CN118446750A - Application recommendation method, electronic equipment and cloud server - Google Patents

Abstract

The present application provides an application recommendation method, an electronic device and a cloud server, the method comprising: the cloud can determine a user prediction profile based on the click behavior data of the terminal device, and divide multiple terminal devices into user profile groups based on the user prediction profile. The terminal devices of the same user profile group have similar application usage habits, so the click behavior data of the terminal devices of the same user group can provide a large amount of data support for the training process of the application recommendation model, thereby improving the training accuracy of the recommendation model, and then sending the recommendation model corresponding to the user portrait group to each terminal device corresponding to the user portrait group can improve the application recommendation accuracy of each terminal device.

Description

Application recommendation method, electronic device and cloud server

Technical Field

The present application relates to the field of terminal devices, and in particular to an application recommendation method, an electronic device and a cloud server.

Background technique

With the development of terminal equipment technology, the number of applications supported by terminal equipment has also increased. In the face of numerous applications, it is possible to provide users with a faster usage experience by predicting the applications that users are about to use and recommending applications on the terminal equipment.

Currently, terminal devices often recommend recently used or most frequently used applications to users, and the recommendation accuracy is low, which affects the user experience.

Summary of the invention

In order to solve the above technical problems, the present application provides an application recommendation method, an electronic device and a cloud server. In this method, the user portrait of the terminal device user can be predicted based on the click behavior data of the terminal device. Since the terminal devices in the same user portrait group have similar application usage habits or application usage behaviors, the click behavior data of each terminal device in the same user portrait group can provide data support for the training of the recommendation model, thereby improving the training accuracy of the recommendation model. Then, the recommendation model corresponding to the user portrait group is sent to each terminal device corresponding to the user portrait group, which can improve the application recommendation accuracy of each terminal device.

In the first aspect, the present application provides an application recommendation method. The method includes: the cloud obtains the first click behavior data of multiple terminal devices. The cloud obtains the user prediction portrait corresponding to each terminal device in the multiple terminal devices based on the first click behavior data of the multiple terminal devices. The cloud divides the multiple terminal devices into multiple user portrait groups based on the user prediction portrait corresponding to each terminal device in the multiple terminal devices. The cloud performs model training based on the first click behavior data of each terminal device in a single user portrait group to obtain an application recommendation model corresponding to the single user portrait group. The cloud sends the application recommendation model corresponding to the single user portrait group to each terminal device in the single user portrait group, so that each terminal device in the single user portrait group makes application recommendations based on the application recommendation model. In this way, the present application provides a method for application recommendation, and the cloud can determine the user prediction portrait based on the click behavior data of the terminal device, and divide the terminal devices with similar application usage habits into the same user portrait group according to the user prediction portrait. Therefore, when the cloud uses the click behavior data of each terminal device of the same user portrait group to train the recommendation model, the trained recommendation model can accurately simulate the application click behavior of users who have the habit of using this type of application, thereby providing sufficient data support for the training of the recommendation model, improving the training accuracy of the recommendation model, and further improving the application recommendation accuracy of each terminal device.

According to the first aspect, the cloud obtains the predicted user portrait corresponding to each of the multiple terminal devices based on the first click behavior data of the multiple terminal devices, including: the cloud inputs the first click behavior data of a single terminal device into the portrait prediction model to obtain the predicted user portrait corresponding to the single terminal device, wherein the portrait prediction model is a large language model. In this way, the cloud can use the reasoning ability of the large language model itself to automatically infer the application click rules that cannot be discovered or summarized manually, thereby improving the accuracy of user portrait prediction.

According to the first aspect, or any implementation of the first aspect above, the cloud obtains the user prediction portrait corresponding to each of the multiple terminal devices based on the first click behavior data of the multiple terminal devices, including: the cloud performs text representation on the first click behavior data of a single terminal device to generate a description text in a fixed text format; the cloud inputs the description text into the portrait prediction model to obtain the user prediction portrait corresponding to the single terminal device. In this way, by converting the format of the click behavior data into a text representation, the cloud can sort out the information of each dimension from the click behavior data and organize it into a fixed format, so that the format of the click behavior data sent by different terminal devices is more consistent, which is convenient for the portrait prediction model to process and improve the model prediction effect.

According to the first aspect, or any implementation of the first aspect above, the first click behavior data of a single terminal device includes user dimension information, and the cloud obtains a user prediction portrait corresponding to each of the multiple terminal devices based on the first click behavior data of multiple terminal devices, including: the cloud inputs the first click behavior data of the single terminal device into the portrait prediction model to obtain an initial user portrait, and the initial user portrait includes user dimension information; the cloud uses the user dimension information in the first click behavior data of the single terminal device to correct the user dimension information in the initial user portrait to obtain a user prediction portrait corresponding to the single terminal device. In this way, by correcting the user dimension in the initial user portrait through the user dimension information in the first behavior data, the prediction accuracy of the user portrait prediction in the cloud can be improved, and then the accuracy of the division of the user portrait group can be improved, so that the accuracy of the application prediction can be improved by improving the accuracy of the application model.

According to the first aspect, or any implementation of the first aspect above, the user dimension information includes subjective user dimension information and objective user dimension information, and the user prediction portrait corresponding to the single terminal device includes the objective user dimension information in the first click behavior data of the single terminal device and the objective user dimension information in the initial user portrait; the cloud uses the user dimension information in the first click behavior data of the single terminal device to correct the user dimension information in the initial user portrait to obtain the user prediction portrait corresponding to the single terminal device, including: the cloud uses the objective user dimension information in the first click behavior data of the single terminal device to replace the objective user dimension information in the initial user portrait; the cloud keeps the objective user dimension information in the initial user portrait unchanged. In this way, since the objective user dimension information in the first behavior click data is relatively accurate, the user dimension information in the user prediction portrait can more accurately reflect the user's recent application usage habits. Therefore, the accuracy of the user prediction portrait can be further improved through this correction method, thereby improving the accuracy of the division of the user portrait group, and the accuracy of the application prediction can be improved by improving the accuracy of the application model.

According to the first aspect, or any implementation of the first aspect above, before the cloud inputs the click behavior data of a single terminal device into the portrait prediction model to obtain the user prediction portrait corresponding to the single terminal device, it also includes: the cloud obtains multiple prediction model training data and a basic large language model, wherein each prediction model training data includes the second click behavior data of a terminal device and the user portrait description text corresponding to a terminal device; the cloud uses the multiple prediction model training data to fine-tune the instructions of the basic large language model to obtain a trained portrait prediction model. In this way, the portrait prediction model can be accurately trained through the user portrait description text and the second click behavior data, thereby improving the prediction accuracy of the user prediction model. In addition, by fine-tuning the instructions of the basic large language model, the portrait prediction model can be accurately trained using a small amount of prediction training model training data, thereby reducing the training difficulty and training cost.

According to the first aspect, or any implementation of the first aspect above, the first click behavior data includes: at least one of time dimension information, space dimension information, device dimension information and application dimension information. In this way, when the recommendation model is trained using the above-mentioned dimension information, the recommendation model can predict the application that the user expects to use based on the change pattern of the application selection behavior within the above-mentioned dimension, which can further improve the application recommendation accuracy of the recommendation model. And, since the above-mentioned dimension information and the user prediction portrait are both associated with the user's application usage habits, accordingly, the cloud can discover the correlation between the above-mentioned dimension information and the user prediction portrait, so that the above-mentioned dimension information can be used to improve the prediction accuracy of the user prediction portrait.

In the second aspect, the present application provides an application recommendation method. The method includes: a terminal device receives an application recommendation model sent by the cloud, specifically, the application recommendation model corresponds to the user portrait group to which the target terminal device belongs, the application recommendation model is trained by the cloud using the first click behavior data of each terminal device in the user portrait group, the user portrait group is based on the user prediction portrait corresponding to each terminal device in multiple terminal devices, and the multiple terminal devices are divided, and the user prediction portrait corresponding to each terminal device in the multiple terminal devices is determined based on the first click behavior data of multiple terminal devices. The terminal device obtains current scene information. The terminal device inputs the current scene information into the application recommendation model to obtain a recommended application. The recommended application is displayed. In this way, the present application provides a method for application recommendation, because the cloud can determine the user prediction portrait based on the click behavior data of the terminal device, and divide the terminal devices with similar application usage habits into the same user portrait group according to the user prediction portrait. Therefore, when the cloud uses the click behavior data of each terminal device in the same user portrait group to train the recommendation model, the trained recommendation model can accurately simulate the application click behavior of users with such application usage habits, thereby providing sufficient data support for the training of the recommendation model and improving the training accuracy of the recommendation model. Furthermore, when the terminal device uses the application recommendation model, it can accurately predict the application that the user expects to use and recommend it, thereby improving the accuracy of application recommendation. In addition, it is convenient for the user to quickly find the application he needs to use in the terminal device, thereby improving the application use experience.

According to the second aspect, after the terminal device displays the recommended application, it also includes: the terminal device detects the user's application selection operation; the terminal device generates third click behavior data in response to the application selection operation; the terminal device uses the third click behavior data in the current time period to update the application recommendation model. In this way, the local recommended applications of the terminal device can be continuously updated following the user's application selection operation, so that the application recommendation model is more in line with the user's personal application usage habits, and the accuracy of application recommendation is further improved.

According to the second aspect, or any implementation of the second aspect above, the terminal device uses the third click behavior data in the current time period to update the application recommendation model, including: the terminal device determines the application recommendation accuracy of the target terminal device in the current time period based on the third click behavior data in the current time period; when the application recommendation accuracy is less than a preset accuracy threshold, the terminal device uses the third click behavior data in the current time period to update the application recommendation model. In this way, when the recommendation accuracy of the application recommendation model is not high, the user's click behavior data can be used to align and update it, so that the application recommendation model is more in line with the user's personal application usage habits, and the application recommendation accuracy is further improved. And, when the recommendation accuracy of the application recommendation model is high, the terminal device does not update the application recommendation model, which ensures the application recommendation accuracy while saving the computing power of the terminal device.

According to the second aspect, or any implementation of the second aspect above, after the terminal device generates the third click behavior data, it also includes: the terminal device sends the third click behavior data to the cloud, so that the cloud trains the recommendation model based on the third click behavior data. In this way, since the user's application click habits are constantly changing, the terminal device can continuously upload the click behavior data, so that the cloud can continuously train the application recommendation model based on the continuously uploaded click behavior data, thereby ensuring the accuracy and timeliness of the application recommendation model on the cloud.

According to the second aspect, or any implementation of the second aspect above, the terminal device inputs the current scenario information into the application recommendation model to obtain the recommended application, including: the terminal device inputs the current scenario information into the application recommendation model to obtain the selection probabilities of multiple candidate applications; the terminal device selects the recommended application from the multiple candidate applications based on the selection probabilities of the multiple candidate applications and the preset recommendation rules. In this way, the terminal device can select applications with a higher probability of being clicked by users for recommendation based on the selection probabilities, thereby improving the recommendation accuracy.

According to the second aspect, or any implementation of the second aspect above, a recommended folder is displayed on the main interface of the terminal device; and the terminal device displays the recommended application, including: the terminal device displays the recommended application in the recommended folder. In this way, the terminal device can place the recommended application in a more intuitive entrance by placing the recommended application in the recommended folder, which makes it easier for the user to find the application they want to use, thereby improving the convenience of using the application.

According to the second aspect, or any implementation of the second aspect above, the user prediction portrait corresponding to each terminal device among the multiple terminal devices is obtained by processing the first click behavior data of each terminal device using a portrait prediction model, and the portrait prediction model is a large language model.

According to the second aspect, or any implementation of the second aspect, the first click behavior data includes:

At least one of time dimension information, space dimension information, device dimension information, and application dimension information.

In a third aspect, the present application provides a cloud server. The cloud server includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and when the computer programs are executed by the one or more processors, the cloud server performs the following steps: obtaining first click behavior data of multiple terminal devices; obtaining a user prediction portrait corresponding to each terminal device in the multiple terminal devices based on the first click behavior data of the multiple terminal devices; dividing the multiple terminal devices into multiple user portrait groups based on the user prediction portrait corresponding to each terminal device in the multiple terminal devices; performing model training based on the first click behavior data of each terminal device in a single user portrait group to obtain an application recommendation model corresponding to the single user portrait group; sending the application recommendation model corresponding to the single user portrait group to each terminal device in the single user portrait group, so that each terminal device in the single user portrait group performs application recommendation based on the application recommendation model.

According to the third aspect, when the computer program is executed by one or more processors, the cloud server performs the following steps: inputting the first click behavior data of a single terminal device into a portrait prediction model to obtain a user prediction portrait corresponding to the single terminal device, wherein the portrait prediction model is a large language model.

According to the third aspect, or any implementation of the third aspect above, when the computer program is executed by one or more processors, the cloud server performs the following steps: textually represent the first click behavior data of a single terminal device, and generate a descriptive text in a fixed text format; input the descriptive text into a portrait prediction model to obtain a user prediction portrait corresponding to the single terminal device.

According to the third aspect, or any implementation method of the third aspect above, the first click behavior data of a single terminal device includes user dimension information. When the computer program is executed by one or more processors, the cloud server executes the following steps: inputting the first click behavior data of the single terminal device into the portrait prediction model to obtain an initial user portrait, and the initial user portrait includes user dimension information; using the user dimension information in the first click behavior data of the single terminal device, correcting the user dimension information in the initial user portrait to obtain a user prediction portrait corresponding to the single terminal device.

According to the third aspect, or any implementation method of the third aspect above, user dimension information includes subjective user dimension information and objective user dimension information, and the user prediction portrait corresponding to a single terminal device includes the objective user dimension information in the first click behavior data of the single terminal device, and the objective user dimension information in the initial user portrait; when the computer program is executed by one or more processors, the cloud server performs the following steps: using the objective user dimension information in the first click behavior data of the single terminal device to replace the objective user dimension information in the initial user portrait; keeping the objective user dimension information in the initial user portrait unchanged.

According to the third aspect, or any implementation of the third aspect above, when the computer program is executed by one or more processors, the cloud server performs the following steps: obtain multiple prediction model training data and a basic large language model, wherein each prediction model training data includes the second click behavior data of a terminal device and a user portrait description text corresponding to the terminal device; use the multiple prediction model training data to fine-tune the basic large language model to obtain a trained portrait prediction model.

According to the third aspect, or any implementation method of the third aspect above, when the computer program is executed by one or more processors, the cloud server executes the following steps: input the first click behavior data of a single terminal device into a portrait prediction model to obtain a user portrait description text of the single terminal device; extract keywords from the user portrait description text to obtain key portrait description words and sentences of the single terminal device; input the key portrait description words and sentences of the single terminal device into the portrait prediction model, and use the intermediate output of the portrait prediction model as the representation vector of the key portrait description words and sentences of the single terminal device; based on the representation vectors of the key portrait description words and sentences of multiple terminal devices, divide the multiple terminal devices into user groups to obtain multiple user portrait groups.

According to the third aspect, or any implementation of the third aspect, the first click behavior data includes: at least one of time dimension information, space dimension information, device dimension information and application dimension information.

The third aspect and any implementation of the third aspect correspond to the first aspect and any implementation of the first aspect, respectively. The technical effects corresponding to the third aspect and any implementation of the third aspect can refer to the technical effects corresponding to the first aspect and any implementation of the first aspect, which will not be repeated here.

In a fourth aspect, the present application provides an electronic device. The electronic device includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and when the computer programs are executed by the one or more processors, the electronic device performs the following steps: receiving an application recommendation model sent by the cloud, the application recommendation model corresponds to the user portrait group to which the target terminal device belongs, the application recommendation model is obtained by training the cloud using the first click behavior data of each terminal device in the user portrait group, the user portrait group is based on the user prediction portrait corresponding to each terminal device in multiple terminal devices, and the multiple terminal devices are divided, and the user prediction portrait corresponding to each terminal device in the multiple terminal devices is determined based on the first click behavior data of the multiple terminal devices; obtaining current scene information; inputting the current scene information into the application recommendation model to obtain recommended applications; and displaying recommended applications.

According to the fourth aspect, when the computer program is executed by one or more processors, the electronic device performs the following steps: detecting the user's application selection operation; generating third click behavior data in response to the application selection operation; and updating the application recommendation model using the third click behavior data within the current time period.

According to the fourth aspect, or any implementation method of the fourth aspect above, when the computer program is executed by one or more processors, the electronic device performs the following steps: based on the third click behavior data in the current time period, determine the application recommendation accuracy of the target terminal device in the current time period; when the application recommendation accuracy is less than the preset accuracy threshold, use the third click behavior data in the current time period to update the application recommendation model.

According to the fourth aspect, or any implementation of the fourth aspect above, when the computer program is executed by one or more processors, the electronic device performs the following steps: sending the third click behavior data to the cloud, so that the cloud performs fine-tuning training of the recommendation model based on the third click behavior data.

According to the fourth aspect, or any implementation of the fourth aspect above, when the computer program is executed by one or more processors, the electronic device performs the following steps: inputting current scene information into an application recommendation model to obtain selection probabilities of multiple candidate applications; and selecting a recommended application from multiple candidate applications based on the selection probabilities of multiple candidate applications and preset recommendation rules.

According to the fourth aspect, or any implementation of the fourth aspect above, a recommended folder is displayed on the main interface of the terminal device, and when the computer program is executed by one or more processors, the electronic device performs the following steps: displaying recommended applications in the recommended folder.

According to the fourth aspect, or any implementation method of the fourth aspect above, the user prediction portrait corresponding to each terminal device among the multiple terminal devices is obtained by processing the first click behavior data of each terminal device using a portrait prediction model, and the portrait prediction model is a large language model.

According to the fourth aspect, or any implementation of the fourth aspect, the first click behavior data includes:

At least one of time dimension information, space dimension information, device dimension information, and application dimension information.

The fourth aspect and any implementation of the fourth aspect correspond to the second aspect and any implementation of the second aspect, respectively. The technical effects corresponding to the fourth aspect and any implementation of the fourth aspect can refer to the technical effects corresponding to the above-mentioned second aspect and any implementation of the second aspect, which will not be repeated here.

In a fifth aspect, the present application provides an application recommendation system, which includes the terminal device and cloud involved in the first and second aspects above.

In a sixth aspect, the present application provides a computer-readable medium for storing a computer program, wherein the computer program includes instructions for executing the method in the first aspect or any possible implementation of the first aspect.

In a seventh aspect, the present application provides a computer-readable medium for storing a computer program, wherein the computer program includes instructions for executing the method in the second aspect or any possible implementation of the second aspect.

In an eighth aspect, the present application provides a computer-readable medium for storing a computer program, wherein the computer program includes instructions for executing the method in the first aspect or any possible implementation of the first aspect.

In a ninth aspect, the present application provides a computer program comprising instructions for executing the method in the second aspect or any possible implementation of the second aspect.

In a tenth aspect, the present application provides a chip, the chip comprising a processing circuit and a transceiver pin, wherein the transceiver pin and the processing circuit communicate with each other through an internal connection path, and the processing circuit executes the method in the first aspect or any possible implementation of the first aspect to control the receiving pin to receive a signal and control the sending pin to send a signal.

In an eleventh aspect, the present application provides a chip, the chip comprising a processing circuit and a transceiver pin, wherein the transceiver pin and the processing circuit communicate with each other through an internal connection path, and the processing circuit executes the method in the second aspect or any possible implementation of the second aspect to control the receiving pin to receive a signal and control the sending pin to send a signal.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of an interface provided in an embodiment of the present application;

FIG2 is a schematic diagram of an exemplary application scenario;

FIG3 is a schematic diagram showing the structure of an electronic device;

FIG4 is a software structure block diagram of an electronic device according to an embodiment of the present application;

FIG5 is a hardware structure block diagram of an example of a cloud device provided in an embodiment of the present application;

FIG6 is a flow chart of an application recommendation method provided in an embodiment of the present application;

FIG. 7 is a schematic diagram showing exemplary click behavior data;

FIG8a is a schematic diagram showing an exemplary end-cloud interaction;

FIG8b is a schematic diagram of an exemplary terminal display interface;

FIG8c is another schematic diagram of end-cloud interaction shown as an example;

FIG9 is a schematic diagram showing an exemplary process of training a recommendation model;

FIG10 is a schematic diagram showing an exemplary training process of a portrait prediction model;

FIG11 is a schematic diagram showing an exemplary process of user portrait prediction;

FIG12 is a schematic diagram showing an exemplary process of dividing user groups;

FIG13 is a schematic diagram showing an exemplary training process of a recommendation model;

FIG14 is a schematic diagram showing an exemplary distribution process of a recommendation model;

FIG15 is a flow chart of another application recommendation method provided in an embodiment of the present application;

FIG16 is a schematic diagram showing an exemplary use process of a recommendation model;

FIG17 is a schematic diagram showing an exemplary process of application recommendation;

FIG18a is a schematic diagram of another terminal display interface;

FIG18b is a schematic diagram showing another exemplary end-cloud interaction;

FIG18c is a schematic diagram showing another exemplary terminal display interface;

FIG19 is a schematic diagram showing an exemplary local update process of a recommendation model;

FIG20 is a schematic diagram showing an exemplary local update process of a recommendation model;

FIG21 is a schematic diagram showing an exemplary update process of a recommendation model;

FIG. 22 is a schematic diagram of the structure of a device provided in an embodiment of the present application.

Detailed ways

The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

The term "and/or" in this article is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and/or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

The terms "first" and "second" in the description and claims of the embodiments of the present application are used to distinguish different objects rather than to describe a specific order of objects. For example, a first target object and a second target object are used to distinguish different target objects rather than to describe a specific order of target objects.

In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

In the description of the embodiments of the present application, unless otherwise specified, the meaning of "multiple" refers to two or more than two. For example, multiple processing units refer to two or more processing units; multiple systems refer to two or more systems.

With the development of terminal devices and the diversification of application programs, more and more applications are downloaded on terminal devices, which makes it difficult for users to find applications. In order to solve this problem, application recommendations can be used on terminal devices to enable users to quickly find the application they want to open among the many applications on the terminal device, thereby improving the user experience.

Figure 1 is a schematic diagram of an interface provided by an embodiment of the present application. Before introducing the embodiment of the present application, the application recommendation interface involved in the embodiment of the present application is first described in conjunction with Figure 1. Referring to Figure 1, taking the terminal device as a mobile phone as an example, in the application recommendation scenario, the display interface of the mobile phone may include one or more controls, and the controls may include but are not limited to: a recommended folder 101 (or a recommended folder, an application recommendation folder, etc.). Specifically, the recommended folder may include an application icon (or referred to as the first application icon) of one or more recommended applications. Among them, the recommended application is the application recommended by the terminal device for the user to use. In an embodiment of the present application, the recommended application may be an application predicted by the terminal device through a recommendation model, which the user has a certain probability of clicking to use.

It should be noted that the number of the suggested folders 101 in the embodiment of the present application can be one or more, and there is no limitation on this. When there are multiple suggested folders 101, the multiple suggested folders 101 can be on the same page or on different pages, and there is no limitation on this.

Furthermore, when there are multiple recommended folders 101, the recommended applications can be placed in different recommended folders 101 according to preset rules. For example, in order to further improve the user experience, the application icons of the same or similar types of recommended applications can be placed in one recommended folder 101, and the application icons of different or dissimilar types of recommended applications can be placed in different recommended folders 101. Optionally, the application icons of recommended applications can be placed in different recommended folders 101 in other ways, such as according to the frequency of use or the probability of use, and there is no specific limitation on this.

And, still referring to Figure 1, the control on the display interface of the mobile phone may also include application icons of other applications (or second application icons). Specifically, other applications are applications that are not recommended, such as other applications that need to be displayed on the display interface except for the recommended applications.

In one embodiment, application icons of other applications may be directly placed on the display interface, such as application icon 103 of the first other application in FIG. 1 .

In another embodiment, the application icons of other applications can be placed in other application folders (i.e., folders for placing other applications), for example, the application icon 1021 of the second other application in FIG. 1 can be placed in the other application folder 102. It should be noted that the number of other application folders 102 can be one or more, and there is no limitation on this. When there are multiple suggestion folders 101, the multiple other application folders 102 can be on the same page or on different pages, and there is no limitation on this.

In the description of the embodiment of the present application, the application icon of the recommended application is located in the recommended folder as an example for explanation. In other embodiments, the present application is also applicable to the application icon of the recommended application being located in a display control such as a recommendation card or a pop-up window, and no specific limitation is made to this.

In the application recommendation scenario, in one example, the most recently used or most frequently used application can be used as a recommended application. However, the user's application usage behavior is often relatively random. For example, in the user's life scenario of shopping and returning home, the user will frequently open the payment application to pay during the shopping process, and after returning home, he will open the short video application for entertainment and leisure. If the user's most frequently used payment application is used as a recommended application after the user returns home, it cannot meet the user's actual usage needs because the user may not use it at home. Therefore, this recommendation method has a low recommendation accuracy.

In another example, a recommendation model can be built using the data in the terminal device, and then the recommendation model can be used to predict the program that the user is about to use. However, due to the limited data in the terminal device, the prediction accuracy is often low, which affects the user's experience.

In order to solve the problem that the application recommendation accuracy is low and affects the user experience, the embodiment of the present application provides an application recommendation method, which can predict the user portrait of the terminal device user based on the click behavior data of the terminal device. Since the user usage habits or usage behaviors of the same user portrait group are similar, the click behavior data of the same user portrait group can provide data support for the training of the recommendation model, improve the training accuracy of the recommendation model, and then send the recommendation model corresponding to the user portrait group to each terminal device corresponding to the user portrait group, which can improve the application recommendation accuracy of each terminal device.

Before describing the technical solution of the embodiment of the present application, the application scenario of the embodiment of the present application is first described with reference to the accompanying drawings.

Referring to FIG. 2 , a schematic diagram of an application scenario provided in an embodiment of the present application is shown. The application scenario includes a cloud and multiple terminal devices (e.g., terminal device 1-terminal device 3). As shown in FIG. 2 , each terminal device is connected to the cloud for communication. In the specific implementation of the embodiment of the present application, the cloud may include one or more cloud servers for providing services to the terminal device. The terminal device may be a mobile phone, tablet computer, wearable device, vehicle-mounted device, augmented reality (AR)/virtual reality (VR) device, laptop computer, etc. with an application display interface. The embodiment of the present application does not limit the specific type of the terminal device. It should be noted that in actual applications, the number of clouds can be one or more, and the number of clouds and terminal devices in the application scenario shown in FIG. 2 is only an adaptive example, and the present application does not limit this.

FIG3 shows a schematic diagram of the structure of the electronic device 100. Exemplarily, the structure of the electronic device in FIG3 may be applicable to the terminal device in FIG2. It should be understood that the electronic device 100 shown in FIG3 is only an example of an electronic device, and the electronic device 100 may have more or fewer components than those shown in the figure, may combine two or more components, or may have different component configurations. The various components shown in FIG3 may be implemented in hardware, software, or a combination of hardware and software including one or more signal processing and/or application specific integrated circuits.

The electronic device 100 may include: a processor 110, an external memory interface 120, an internal memory 121, a universal serial bus (USB) interface 130, a charging management module 140, a power management module 141, a battery 142, an antenna 1, an antenna 2, a mobile communication module 150, a wireless communication module 160, an audio module 170, a speaker 170A, a receiver 170B, a microphone 170C, an earphone interface 170D, a sensor module 180, a button 190, a motor 191, an indicator 192, a camera 193, a display screen 194, and a subscriber identification module (SIM) card interface 195, etc. The sensor module 180 may include a pressure sensor 180A, a gyroscope sensor 180B, an air pressure sensor 180C, a magnetic sensor 180D, an acceleration sensor 180E, a distance sensor 180F, a proximity light sensor 180G, a fingerprint sensor 180H, a temperature sensor 180J, a touch sensor 180K, an ambient light sensor 180L, a bone conduction sensor 180M, etc.

The processor 110 may include one or more processing units, for example, the processor 110 may include an application processor (AP), a modem processor, a graphics processor (GPU), an image signal processor (ISP), a controller, a memory, a video codec, a digital signal processor (DSP), a baseband processor, and/or a neural-network processing unit (NPU), etc. Different processing units may be independent devices or integrated into one or more processors.

The controller may be the nerve center and command center of the electronic device 100. The controller may generate an operation control signal according to the instruction operation code and the timing signal to complete the control of fetching and executing instructions.

A memory may also be provided in the processor 110 for storing instructions and data. In some embodiments, the memory in the processor 110 is a high-speed cache memory. The USB interface 130 is an interface that complies with USB standard specifications, and may specifically be a Mini USB interface, a Micro USB interface, a USB Type C interface, etc. The USB interface 130 may be used to connect a charger to charge the electronic device 100, or may be used to transfer data between the electronic device 100 and a peripheral device. It may also be used to connect headphones to play audio through the headphones. The interface may also be used to connect other electronic devices, such as AR devices, etc.

The charging management module 140 is used to receive charging input from a charger. The charger may be a wireless charger or a wired charger. In some wired charging embodiments, the charging management module 140 may receive charging input from a wired charger through the USB interface 130. In some wireless charging embodiments, the charging management module 140 may receive wireless charging input through a wireless charging coil of the electronic device 100. While the charging management module 140 is charging the battery 142, it may also power the electronic device through the power management module 141.

The power management module 141 is used to connect the battery 142, the charging management module 140 and the processor 110. The power management module 141 receives input from the battery 142 and/or the charging management module 140, and supplies power to the processor 110, the internal memory 121, the external memory, the display screen 194, the camera 193, and the wireless communication module 160. The wireless communication function of the electronic device 100 can be implemented through the antenna 1, the antenna 2, the mobile communication module 150, the wireless communication module 160, the modem processor and the baseband processor.

Antenna 1 and antenna 2 are used to transmit and receive electromagnetic wave signals. Each antenna in electronic device 100 can be used to cover a single or multiple communication frequency bands. Different antennas can also be reused to improve the utilization of antennas. For example, antenna 1 can be reused as a diversity antenna for a wireless local area network. In some other embodiments, the antenna can be used in combination with a tuning switch.

The mobile communication module 150 can provide solutions for wireless communications including 2G/3G/4G/5G, etc., applied to the electronic device 100. The mobile communication module 150 can include at least one filter, a switch, a power amplifier, a low noise amplifier (LNA), etc. The wireless communication module 160 can provide solutions for wireless communications including wireless local area networks (WLAN) (such as wireless fidelity (Wi-Fi) networks), Bluetooth (BT), global navigation satellite system (GNSS), frequency modulation (FM), near field communication (NFC), infrared technology (IR), etc., applied to the electronic device 100.

In some embodiments, antenna 1 of electronic device 100 is coupled to mobile communication module 150, and antenna 2 is coupled to wireless communication module 160, so that electronic device 100 can communicate with the network and other devices through wireless communication technology.

The electronic device 100 implements the display function through a GPU, a display screen 194, and an application processor. The GPU is a microprocessor for image processing, which connects the display screen 194 and the application processor. The GPU is used to perform mathematical and geometric calculations for graphics rendering. The processor 110 may include one or more GPUs that execute program instructions to generate or change display information.

The display screen 194 is used to display images, videos, etc. The display screen 194 includes a display panel. The display panel may be a liquid crystal display (LCD), an organic light-emitting diode (OLED), etc. In some embodiments, the electronic device 100 may include 1 or N display screens 194, where N is a positive integer greater than 1.

The electronic device 100 can realize the shooting function through ISP, camera 193, video codec, GPU, display screen 194 and application processor.

The camera 193 is used to capture still images or videos. The object generates an optical image through the lens and projects it onto the photosensitive element. The photosensitive element can be a charge coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The photosensitive element converts the optical signal into an electrical signal, and then passes the electrical signal to the ISP to be converted into a digital image signal. The ISP outputs the digital image signal to the DSP for processing. The DSP converts the digital image signal into an image signal in a standard RGB, YUV or other format. In some embodiments, the electronic device 100 may include 1 or N cameras 193, where N is a positive integer greater than 1.

The external memory interface 120 can be used to connect an external memory card, such as a Micro SD card, to expand the storage capacity of the electronic device 100. The external memory card communicates with the processor 110 through the external memory interface 120 to implement a data storage function, such as storing music, video and other files in the external memory card.

The internal memory 121 can be used to store computer executable program codes, which include instructions. The processor 110 enables the electronic device to execute the application recommendation method in this application by running the instructions stored in the internal memory 121. The internal memory 121 may include a program storage area and a data storage area. Among them, the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc. The data storage area can store data created during the use of the electronic device 100 (such as audio data, a phone book, etc.), etc. In addition, the internal memory 121 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, a universal flash storage (UFS), etc.

The electronic device 100 can implement audio functions such as music playing and recording through the audio module 170, the speaker 170A, the receiver 170B, the microphone 170C, the headphone jack 170D, and the application processor.

The audio module 170 is used to convert digital audio information into analog audio signal output, and is also used to convert analog audio input into digital audio signals. The audio module 170 can also be used to encode and decode audio signals. In some embodiments, the audio module 170 can be arranged in the processor 110, or some functional modules of the audio module 170 can be arranged in the processor 110.

The pressure sensor 180A is used to sense the pressure signal and can convert the pressure signal into an electrical signal. In some embodiments, the pressure sensor 180A can be set on the display screen 194. There are many types of pressure sensors 180A, such as resistive pressure sensors, inductive pressure sensors, capacitive pressure sensors, etc. The capacitive pressure sensor can be a parallel plate including at least two conductive materials. When a force acts on the pressure sensor 180A, the capacitance between the electrodes changes. The electronic device 100 determines the intensity of the pressure according to the change in capacitance. When a touch operation acts on the display screen 194, the electronic device 100 detects the touch operation intensity according to the pressure sensor 180A. The electronic device 100 can also calculate the touch position according to the detection signal of the pressure sensor 180A. In some embodiments, touch operations acting on the same touch position but with different touch operation intensities can correspond to different operation instructions. For example: when a touch operation with a touch operation intensity less than the first pressure threshold acts on the short message application icon, an instruction to view the short message is executed. When a touch operation with a touch operation intensity greater than or equal to the first pressure threshold acts on the short message application icon, an instruction to create a new short message is executed.

The touch sensor 180K is also called a "touch panel". The touch sensor 180K can be set on the display screen 194, and the touch sensor 180K and the display screen 194 form a touch screen, also called a "touch screen". The touch sensor 180K is used to detect touch operations acting on or near it. The touch sensor can pass the detected touch operation to the application processor to determine the type of touch event. Visual output related to the touch operation can be provided through the display screen 194. In other embodiments, the touch sensor 180K can also be set on the surface of the electronic device 100, which is different from the position of the display screen 194.

The key 190 includes a power key, a volume key, etc. The key 190 may be a mechanical key or a touch key. The electronic device 100 may receive key input and generate key signal input related to user settings and function control of the electronic device 100.

Motor 191 can generate vibration prompts. Motor 191 can be used for incoming call vibration prompts, and can also be used for touch vibration feedback. For example, touch operations acting on different applications (such as taking pictures, audio playback, etc.) can correspond to different vibration feedback effects. Indicator 192 can be an indicator light, which can be used to indicate charging status, power changes, messages, missed calls, notifications, etc.

The software system of the electronic device 100 may adopt a layered architecture, an event-driven architecture, a micro-core architecture, a micro-service architecture, or a cloud architecture. The embodiment of the present application takes the Android system of the layered architecture as an example to exemplify the software structure of the electronic device 100.

FIG. 4 is a software structure block diagram of the electronic device 100 according to an embodiment of the present application.

The layered architecture of the electronic device 100 divides the software into several layers, each with clear roles and division of labor. The layers communicate with each other through software interfaces. In some embodiments, the Android system is divided into four layers, from top to bottom, namely, the application layer, the application framework layer, the Android runtime (Android runtime) and the system library, and the kernel layer.

The application layer can include a series of application packages.

As shown in FIG. 4 , the application layer may include applications such as camera, gallery, calendar, call, map, navigation, WLAN, Bluetooth, music, video, short message, etc.

In some embodiments, the application layer may also include a desktop organization module.

In an embodiment of the present application, the desktop organization module can be used to collect statistics and call recommendation models. Among them, the desktop organization module can respond to the user's click operation on the application icon on the terminal device, collect the click behavior data generated by the terminal device, and send the click behavior data to the cloud. And, it can also receive the recommendation model returned by the cloud. And obtain the current scene information of the terminal device, and determine the recommended application based on the current scene information and the recommendation model. And, display the recommended application in the suggestion folder of the display interface of the terminal device.

In the description of the embodiment of the present application, the desktop arrangement module is located in the application layer. In other embodiments, the present application is also applicable to the desktop arrangement module being located in the application framework layer, the system layer, and other software structure layers, without specific limitation.

The application framework layer provides an application programming interface (API) and a programming framework for the applications in the application layer. The application framework layer includes some predefined functions.

As shown in FIG. 4 , the application framework layer may include a window manager, a content provider, a view system, a telephony manager, a resource manager, a notification manager, and the like.

The window manager is used to manage window programs. The window manager can obtain the display screen size, determine whether there is a status bar, lock the screen, capture the screen, etc.

Content providers are used to store and retrieve data and make it accessible to applications. The data may include videos, images, audio, calls made and received, browsing history and bookmarks, phone books, etc.

The view system includes visual controls, such as controls for displaying text, controls for displaying images, etc. The view system can be used to build applications. A display interface can be composed of one or more views. For example, a display interface including a text notification icon can include a view for displaying text and a view for displaying images.

The phone manager is used to provide communication functions of the electronic device 100, such as management of call status (including connecting, hanging up, etc.).

The resource manager provides various resources for applications, such as localized strings, icons, images, layout files, video files, and so on.

The notification manager enables applications to display notification information in the status bar. It can be used to convey notification-type messages and can disappear automatically after a short stay without user interaction. For example, the notification manager is used to notify download completion, message reminders, etc. The notification manager can also be a notification that appears in the system top status bar in the form of a chart or scroll bar text, such as notifications of applications running in the background, or a notification that appears on the screen in the form of a dialog window. For example, a text message is displayed in the status bar, a prompt sound is emitted, an electronic device vibrates, an indicator light flashes, etc.

The kernel layer is the layer between hardware and software. The kernel layer contains at least display driver, camera driver, audio driver, and sensor driver.

It is understandable that the layers in the software structure shown in FIG4 and the components contained in each layer do not constitute a specific limitation on the electronic device 100. In other embodiments of the present application, the electronic device 100 may include more or fewer layers than shown in the figure, and each layer may include more or fewer components, which is not limited in the present application.

FIG5 is a block diagram of the hardware structure of an example of a cloud device provided in an embodiment of the present application. The cloud in the embodiment of the present application may be one or more server clusters. The cluster may include one or more cloud devices. The cloud device may be a cloud server or other device, which is not limited in the present application. As shown in FIG5 , the cloud device 200 includes a processor 222, which further includes one or more processors, and a memory resource represented by a memory 232 for storing instructions executable by the processor 222, such as an application. The application stored in the memory 232 may include one or more modules, each of which corresponds to a set of instructions. In addition, the processor 222 is configured to execute instructions to execute the application recommendation method provided in an embodiment of the present application.

The cloud device 200 may further include a power supply component 226 configured to perform power management of the cloud device 200, a wired or wireless network interface 250 configured to connect the cloud device 200 to a network, and an input/output interface 258. The cloud device 200 may operate based on an operating system stored in the memory 232.

The methods in the following embodiments can be implemented in a terminal device and a cloud with the above-mentioned hardware structure. In the following embodiments, taking the application recommendation system including the terminal device and the cloud as shown in Figure 2 as an example, the specific implementation method of the embodiment of the present application is described in detail in combination with the following Figures 6-21. Specifically, the application recommendation process in the embodiment of the present application can be divided into two parts. The first part is the model generation scenario, that is, the cloud side trains the recommendation model based on the click behavior data of the terminal device and sends the trained recommendation model to the terminal device. The second part is the application prediction scenario, that is, the scenario in which the terminal device predicts and recommends applications based on the recommendation model.

The following is a detailed description of the training process of the recommendation model in conjunction with the method flow diagram shown in FIG6 . Referring to FIG6 , the training process specifically includes:

S101, the cloud receives click behavior data sent by the terminal device.

Click behavior data, i.e., relevant data used to describe the user's application click behavior. Exemplarily, click behavior data may include, but is not limited to, one or more of the following information: user dimension information, time dimension information, space dimension information, device dimension information, and application dimension information. In one example, in order to improve the accuracy of application recommendation, FIG7 is a schematic diagram of click behavior data shown as an example. Referring to FIG7, click behavior data may include user dimension information, time dimension information, space dimension information, device dimension information, and application dimension information. In another example, since user dimension information may be difficult to obtain, click behavior information may not include user dimension information. The following is an explanation of each dimension information in conjunction with FIG7.

User dimension information, that is, relevant characteristic information that can reflect application click behavior in user dimension. Optionally, referring to FIG. 7 , user dimension information may include but is not limited to: one or more of user identifier (Identity, ID), gender, age, occupation, usage habits, and hobbies.

Exemplarily, the user ID may be information that can represent the user's identity, such as the user's mobile phone number, terminal device code, user nickname, etc.

Exemplarily, gender can be "male" or "female". It should be noted that considering that users of different genders may prefer different applications, that is, user application preferences may have gender patterns, accordingly, when the cloud uses gender to train the recommendation model, the recommendation model can predict the application that the user expects to use based on the user's gender characteristics, thereby improving the accuracy of application recommendations. In addition, when the cloud uses gender as a user portrait label, the user's gender can reflect the user's application usage habits, so that the predicted portrait predicted by the portrait prediction model can accurately describe the user's application usage habits, thereby accurately classifying users according to their application usage habits, thereby improving classification accuracy.

Exemplarily, age can be information that can describe the age characteristics of the user, for example, it can be the age group corresponding to the user, such as "youth", "middle-aged", etc. It should be noted that since users of different ages may prefer different applications, for example, teenagers tend to use learning applications, and the elderly tend to use short video applications. Accordingly, when the cloud uses age to train the recommendation model, the recommendation model can predict the application that the user expects to use based on the user's age characteristics, thereby improving the accuracy of application recommendations. In addition, when the cloud uses age as a user portrait label, since the user's age can reflect the user's application usage habits, the predicted portrait predicted by the portrait prediction model can accurately describe the user's application usage habits, so that users can be accurately classified according to their application usage habits, thereby improving classification accuracy.

For example, occupation may be information that can describe the occupational characteristics of the user, such as occupational classifications such as "clerk" and "researcher", or specific occupations such as "doctor" and "clerk".

It should be noted that users of different professions may have different preferred applications due to their different professional habits, that is, user application preferences may have professional rules. Accordingly, when the cloud uses profession to train the recommendation model, the recommendation model can predict the applications that the user expects to use based on the user's professional characteristics, thereby improving the application recommendation accuracy of the recommendation model. In addition, when the cloud uses profession as a user portrait label, the user's profession can reflect the user's application usage habits, so that the predicted portrait predicted by the portrait prediction model can accurately describe the user's application usage habits, thereby accurately classifying users according to their application usage habits, thereby improving classification accuracy.

Exemplarily, usage habits may be characteristic information that can represent applications that users are accustomed to using, such as applications that users use more frequently. For example, usage habits may be applications that users use more frequently over a period of time, or applications that users use more frequently in different time periods.

It should be noted that since usage habits represent the applications that users are accustomed to using (or prefer to use), for example, users who frequently use short video applications may prefer to use short video applications, and users who frequently use photo applications may prefer to use photo applications or photo editing applications, that is, user application preferences may have usage habit patterns. Accordingly, when the cloud uses usage habits to train the recommendation model, the recommendation model can predict the applications that users expect to use based on the user's usage habit characteristics, which can further improve the application recommendation accuracy of the recommendation model. In addition, when the cloud uses usage habits as user portrait labels, user usage habits can reflect the user's application usage habits, so that the predicted portraits predicted by the portrait prediction model can accurately describe the user's application usage habits, so that users can be accurately classified according to their application usage habits, thereby improving the classification accuracy.

For example, hobbies may be information describing user hobby characteristics, such as "drawing", "sports", etc. It should be noted that user dimension information may also include other user dimension information such as speed that can reflect user characteristics of application click behavior, and there is no specific limitation on this.

It should be noted that since users may prefer to use applications related to their hobbies, for example, users who frequently use short video applications may prefer to use short video applications, and users who frequently use photo applications may prefer to use photo applications or photo editing applications, that is, user application preferences may have hobby patterns. Accordingly, when the cloud uses hobbies to train the recommendation model, the recommendation model can predict the applications that the user expects to use based on the user's hobby characteristics, which can further improve the application recommendation accuracy of the recommendation model. In addition, when the cloud uses hobbies as user portrait labels, user hobbies can reflect the user's application usage habits, so that the predicted portraits predicted by the portrait prediction model can accurately describe the user's application usage habits, so that users can be accurately classified according to their application usage habits, thereby improving the classification accuracy.

Time dimension information, i.e., relevant characteristic information that can reflect the application click behavior in the time dimension (e.g., relevant time characteristic information of the specific time when the terminal device 1 detects the application click behavior). Optionally, referring to FIG. 7 , the time dimension information may include, but is not limited to, one or more of: hours, days of the week, months, quarters, holidays. It should be noted that the time dimension information may also include other time dimension information such as speed that can reflect the time characteristics of the application click behavior, and there is no specific limitation on this.

It should be noted that since users may have different application usage preferences in different time periods, such as watching videos at lunch and using office software at work, there are often obvious time patterns. Accordingly, when the cloud uses time dimension information to train the recommendation model, the recommendation model can predict the application that the user expects to use based on the time pattern of application clicks, which can further improve the application recommendation accuracy of the recommendation model. In addition, since both time dimension information and user prediction portraits are associated with the user's application usage habits, the cloud can use the portrait prediction model to discover the correlation between time dimension information and user prediction portraits, so that the time dimension information can be used to improve the prediction accuracy of user prediction portraits.

Spatial dimension information, that is, the relevant characteristic information that can reflect the application click behavior in the spatial dimension (for example, the relevant location data of the user's location when performing the application click behavior). Optionally, referring to Figure 7, the spatial dimension information may include, but is not limited to, one or more of the global positioning system (GPS) position, base station ID, and point of interest (POI) distribution. Exemplarily, the GPS position may be the GPS positioning information of the user's location when performing the application click behavior. Exemplarily, the base station ID may be the base station ID of the base station to which the terminal device is connected when the user performs the application click behavior. Exemplarily, the POI distribution may be the POI points distributed around the user when the user performs the application click behavior. For example, the POI distribution may be a POI point centered on the user and within a preset range. Exemplarily, the POI point may refer to a place or landmark with special significance on the map that can attract people's attention, such as a shop, a tourist attraction, a hospital, a transportation hub, etc.

Optionally, the spatial dimension information may also include other device dimension information such as speed that can reflect the location characteristics of the application click behavior, and there is no specific limitation on this.

It should be noted that users may have specific usage preferences when they are in specific locations. For example, users may prefer to use travel applications when they are in subway stations or bus stations. For example, users may prefer to use railway ticket purchasing applications when they are in train stations. Users may prefer to use payment applications, ordering applications, or consumer review applications when they are in shopping malls. The above application usage behaviors often have obvious spatial usage patterns, and location changes affect user click behaviors. Accordingly, when the cloud uses spatial dimension information to train the recommendation model, the recommendation model can predict the application that the user expects to use based on the location pattern of application clicks, which can further improve the application recommendation accuracy of the recommendation model. In addition, since both spatial dimension information and user prediction portraits are associated with the user's application usage habits, the cloud can use the portrait prediction model to explore the correlation between spatial dimension information and user prediction portraits, so that the spatial dimension information can be used to improve the prediction accuracy of user prediction portraits.

Device dimension information, i.e., relevant characteristic information that can reflect the application click behavior in the device dimension (e.g., the device information of the terminal device 1 when the application click behavior occurs). Optionally, referring to FIG. 7 , the device dimension information may include, but is not limited to, one or more of acceleration, power, charging and discharging, network connection, and Bluetooth connection.

For example, acceleration may refer to the acceleration of the terminal device when the application click behavior occurs, for example, it may be an acceleration interval or a specific acceleration value.

It should be noted that users may have different application usage preferences at different accelerations. For example, when a user is lying down and browsing their phone (acceleration is 0), they tend to open short video applications, information sharing platforms, and video applications. When a user is moving at high speed on a high-speed rail or other means of transportation (at this time, the acceleration is large), due to the poor signal of the terminal device, they may prefer to open instant messaging applications, novel applications, and other applications with low network signal requirements. The above application usage behaviors often have obvious acceleration patterns. Accordingly, when the cloud uses acceleration to train the recommendation model, the recommendation model can predict the application that the user expects to use based on the acceleration characteristics of the terminal device, which can further improve the application recommendation accuracy of the recommendation model. In addition, since both acceleration and user prediction portraits are associated with the user's application usage habits, the cloud can use the portrait prediction model to explore the correlation between acceleration and user prediction portraits, so that acceleration can be used to improve the prediction accuracy of user prediction portraits.

For example, the power level may refer to the remaining power level of the terminal device when the application click behavior occurs, for example, the power level interval.

It should be noted that users may have different application usage preferences at different power levels. For example, users may not open games, short video applications, and other power-intensive applications when the battery is low. The above application usage behaviors often have obvious power patterns. Accordingly, when the cloud uses power to train the recommendation model, the recommendation model can predict the application that the user expects to use based on the power pattern of the terminal device when the application is clicked, which can further improve the application recommendation accuracy of the recommendation model. In addition, since both the power level and the user prediction portrait are related to the user's application usage habits, the cloud can use the portrait prediction model to discover the correlation between the power level and the user prediction portrait, so that the power level can be used to improve the prediction accuracy of the user prediction portrait.

Exemplarily, charging and discharging can be the charging and discharging state of the terminal device when the application click behavior occurs. For example, charging and discharging can include "charging" or "discharging". For another example, when the terminal device is in a charging state, since the user may have different application usage preferences when connecting to different charging devices, charging and discharging can also include relevant information of the charging device.

It should be noted that since users may have different application usage preferences in different charging and discharging states, for example, users may be around a charger when charging with a charger, and accordingly, users may not use sports applications. The above application click behaviors often have charging and discharging patterns. Accordingly, when the cloud uses charging and discharging to train the recommendation model, the recommendation model can predict the application that the user expects to use based on the charging and discharging pattern of the terminal device when the application is clicked, which can improve the application recommendation accuracy of the recommendation model. In addition, since both the charging and discharging information and the user prediction portrait are associated with the user's application usage habits, accordingly, the cloud can use the portrait prediction model to explore the correlation between the charging and discharging information and the user prediction portrait, so that the charging and discharging information can be used to improve the prediction accuracy of the user prediction portrait.

For example, the network connection may refer to the network connection status of the terminal device when the application click behavior occurs, such as wireless network connection, cellular network connection, etc. For another example, when the cellular network is connected, the network connection may also include network mode information such as 2G, 3G, 4G, 5G, etc.

It should be noted that users may have different application usage preferences in different network connection states. In one example, when the user is connected to a wireless network, he may prefer to use applications such as video applications that consume more traffic. When connected to a cellular network, he may prefer to use applications such as instant messaging applications that consume less traffic or have higher communication real-time requirements. In another example, in the 2G network mode, users may use applications that do not require high network communication quality, such as applications that can be used offline. The probability of using applications such as short video applications that require high network communication quality is often low. And, in the 5G network mode, users may use applications that require high network communication quality, such as short video applications. The above application click behavior often has a network connection rule. Accordingly, when the cloud uses the network connection to train the recommendation model, the recommendation model can predict the application that the user expects to use based on the network connection characteristics of the terminal device when the application is clicked, which can improve the application recommendation accuracy of the recommendation model. And, since both the network connection and the user prediction portrait are associated with the user's application usage habits, accordingly, the cloud can use the portrait prediction model to discover the correlation between the network connection and the user prediction portrait, so that the network connection can be used to improve the prediction accuracy of the user prediction portrait.

Exemplarily, the Bluetooth connection may refer to the Bluetooth connection status of the terminal device when the application click behavior occurs, such as "sports watch connected", "headphones connected", "stylus pen connected", etc.

It should be noted that users may have different application usage preferences when using different Bluetooth connections. In one example, when the terminal device is connected to a sports watch via Bluetooth, sports applications may be preferred. When connected to headphones via Bluetooth, video applications or music playback applications may be preferred. When connected to a stylus, drawing applications or other applications that use a stylus may be preferred. The above-mentioned application click behaviors often have Bluetooth connection rules. Accordingly, when the cloud uses Bluetooth connection to train the recommendation model, the recommendation model can predict the application that the user expects to use based on the Bluetooth connection characteristics of the terminal device when the application is clicked, thereby improving the application recommendation accuracy of the recommendation model. In addition, since both Bluetooth connection and user prediction portrait are associated with the user's application usage habits, the cloud can use the portrait prediction model to explore the correlation between Bluetooth connection and user prediction portrait, thereby using Bluetooth connection to improve the prediction accuracy of user prediction portrait.

Optionally, the device dimension information may also include other device dimension information such as speed, which can reflect device characteristics of application click behavior, and there is no specific limitation on this.

Application dimension information, that is, the relevant characteristic information that can reflect the application click behavior in the application dimension. Optionally, the application dimension information may refer to the relevant application information of the application clicked by the application click behavior. Exemplarily, referring to Figure 7, the application dimension information may include but is not limited to one or more of application ID, category, scoring, ranking, and charging. Another example is that the application dimension information may also include application list information of the terminal device. Among them, the application list information is used to record the information of the downloaded applications in the terminal device. It should be noted that if the user likes to take pictures, he will often download multiple photo-taking apps, so the application list information can reflect the user's application usage preferences.

It should be noted that users' multiple application click behaviors often have certain click patterns, and information such as application lists can reflect users' application usage preferences. Accordingly, when the cloud uses application dimension information to train the recommendation model, the recommendation model can predict the applications that users expect to use based on the application click patterns, which can further improve the application recommendation accuracy of the recommendation model. In addition, since both application dimension information and user prediction portraits are associated with users' application usage habits, the cloud can use the portrait prediction model to discover the correlation between application dimension information and user prediction portraits, so that the application dimension information can be used to improve the prediction accuracy of user prediction portraits.

After introducing the click behavior data, the specific implementation of S101 will be described next.

In some embodiments, in response to a received user operation (user operations include but are not limited to: single click, double click, slide, etc.), the terminal device collects click behavior data related to this user operation. After collecting the click behavior data, the terminal device sends the click behavior data to the cloud. Exemplarily, if the collection conditions of different information in the click behavior data are different, the time when the terminal device sends different information to the cloud may be different. The sending process of click behavior data is described below in conjunction with Figures 8a-8c.

In one example, after the user clicks on the application, the terminal device collects the click behavior data (such as time dimension data, space dimension data, device dimension data, application dimension data) generated by this click operation. Figure 8a is an exemplary diagram of a terminal-cloud interaction, and the display interface of the terminal device 1 may include one or more application icons. When the user is at location B1, the application icon 401 of the short video application B is displayed on the display interface of the terminal device 1 (the specific click time is time A1), and the terminal device 1 can respond to the user's click operation and collect the click behavior data generated by this click operation. That is, the data generated in the user's click operation that records the important information of the click behavior (such as "short video application B was clicked on the terminal device 1 at time A1 and location B1)", and the generated click behavior data is sent to the cloud. Optionally, the click behavior data may also include other information that can be used for user portrait prediction, such as user dimension information, and there is no specific restriction on this.

In another example, after the user fills in the click behavior data (such as user dimension data), the terminal device collects the click behavior data filled in by the user. Figure 8b is an exemplary schematic diagram of a terminal display interface. Referring to Figure 8b, when the terminal device 1 displays the selection window "Whether to provide user information", if it is detected that the user has triggered the "Yes" control, the terminal device 1 can display a personal information filling interface for the user to fill in the user dimension information. And, Figure 8c is an exemplary schematic diagram of another end-cloud interaction. After the user has completed the filling, if it is detected that the user has triggered the "Submit" control, the click behavior data containing the user dimension information can be sent to the cloud.

In another example, if the click behavior data includes application list information, the terminal device can send the application list information to the cloud at every preset time period or after detecting that the user has updated the application of the terminal device (such as downloading a new application or deleting an existing application).

In other embodiments, the terminal device may periodically send the collected click behavior data to the cloud. For example, the terminal device may store the collected click behavior data, and after a certain period of time, send the collected click behavior data within the period of time to the cloud.

After introducing the transmission process of click behavior data through S101, the prediction process of user portrait is explained in conjunction with S102.

In the embodiment of the present application, the cloud can receive click behavior data sent by some or all terminal devices in the system (such as shown in FIG. 2 ). That is, the cloud can receive click behavior data sent by multiple terminal devices.

S102, the cloud performs user portrait prediction based on click behavior data and portrait prediction model.

The portrait prediction model, in the embodiment of the present application, is a large language model, that is, a deep learning model trained with text data, which can generate natural language text or understand the meaning of language text. For example, llama-2, GPT, Palm and other models.

Regarding the use of the portrait prediction model, in S102, the cloud can input the click behavior data of each terminal device into the portrait prediction model to obtain a user prediction portrait. Among them, the user prediction portrait is a user portrait predicted by the user prediction model. Exemplarily, the user prediction portrait can be descriptive information that can represent the user's application usage habits. For example, it can include but is not limited to one or more of the following user dimension information: gender, age, occupation, hobbies, usage habits, etc.

Optionally, the cloud can input the click behavior data of each terminal device into the portrait prediction model one by one, or input the click behavior data of multiple terminal devices into the portrait prediction model in batches, without limitation. For example, as shown in FIG9, a flowchart of the exemplary recommendation model training is shown, after user 1 performs an application selection operation on terminal device 1 (i.e., an operation for selecting a target application from multiple applications to be selected, such as click, tap, drawing gestures, etc.), terminal device 1 sends the collected click behavior data 1 to the cloud, and the cloud can input multiple click behavior data 1 within a preset period into the portrait prediction model to obtain a predicted portrait of user 1 (i.e., a predicted portrait of user 1). Among them, the multiple click behavior data 1 are obtained by terminal device 1 based on different click operations within the period, and the multiple click behavior data 1 are different. For example, after user 1 performs a click operation on terminal device 1, terminal device 1 can obtain a click behavior data 1 related to this click operation.

Also, after user 2 performs an application selection operation on terminal device 2, terminal device 2 sends the collected click behavior data 2 to the cloud, and the cloud can input multiple click behavior data 2 within a preset period into the portrait prediction model to obtain a predicted portrait of user 2 (i.e., a predicted portrait of user 2). Also, after user 3 performs an application selection operation on terminal device 3, terminal device 3 sends the collected click behavior data 3 to the cloud, and the cloud can input multiple click behavior data 3 within a preset period into the portrait prediction model to obtain a predicted portrait of user 3 (i.e., a predicted portrait of user 3). It should be noted that the embodiment of the present application takes the interaction between the cloud and terminal devices 1-terminal devices 3 as an example to illustrate the interaction process between the cloud and multiple terminal devices. It should be noted that the number of terminal devices interacting with the cloud can be more, and this application does not limit it.

The training process of the portrait prediction model is explained in conjunction with FIG10. For example, FIG10 is a schematic diagram of the training process of the portrait prediction model shown as an example. Specifically, the cloud can use multiple prediction model training data (i.e., the training data of the portrait prediction model, such as prediction model training data 1, prediction model training data 2, and prediction model training data 3. One prediction model training data includes a click behavior data and a portrait description text of the terminal device corresponding to the click behavior data. Among them, the portrait description text is used to describe the user portrait of the terminal device corresponding to the click behavior data, such as gender, age, occupation, hobbies, usage habits, etc.) to fine-tune the basic large language model (i.e., the pre-trained language model obtained by training a large amount of text), and obtain the trained portrait prediction model when the training requirements (such as convergence conditions, etc.) are met. Exemplarily, in the process of fine-tuning the instructions, the cloud can fine-tune the parameters (such as model weights) of all layers of the basic large language model, or fine-tune the parameters of the top layer of the basic large language model, etc. to fine-tune the instructions of the basic large language model. Exemplarily, in the process of fine-tuning the instructions, the prediction model training data can be used to construct an example of the instruction format. For example, a prediction model training data can be constructed as an example of an instruction format. Specifically, the example of the instruction format may include a task description and a pair of input and output. In one example, an example of an instruction format may be "Please describe the user portrait based on the following user click behavior data (i.e., task description), the time is xxxx, the location is xxxx, the device status is xxxx, and the xxxx application is clicked (i.e., the input of the question)", and the answer is "User A is a 28-year-old female user, a programmer by profession, and has the habit of using food delivery applications at noon and video playback applications late at night. She likes to chase dramas and watch movies (i.e., the output of the answer)".

It should be noted that by fine-tuning the instructions of the basic large language model, the portrait prediction model can be accurately trained using a small amount of prediction training model training data, reducing the difficulty and cost of training. Optionally, other fine-tuning methods other than instruction fine-tuning can be used for model fine-tuning, and model training can also be performed according to other model training methods other than model fine-tuning, without specific restrictions. In an embodiment of the present application, the cloud can receive click behavior data sent by some or all terminal devices in the system (such as shown in Figure 2) for user prediction model training. That is, the cloud can receive click behavior data sent by multiple terminal devices for user prediction model training.

Wherein, referring to Figure 10, each prediction model training data may include multiple groups of corresponding click behavior data and portrait description texts (such as click behavior data 1 and portrait description text 1, click behavior data 2 and portrait description text 2, click behavior data 3 and portrait description text 3). Wherein, the click behavior data can be used as a sample, and the portrait description text can be used as a label. Correspondingly, the cloud can adjust the model parameters of the basic large language model according to the output results of the basic large language model on the click behavior data and the portrait description text. Wherein, the portrait description text is used to describe the user portrait of the terminal device corresponding to the click behavior data, for example, it may include gender, age, occupation, hobbies, usage habits, etc. Exemplarily, the portrait description text may be manually annotated. For example, it may be filled out by the user of the terminal device, for example, the portrait description text may be obtained in the form of a questionnaire. Optionally, in order to ensure the accuracy of the portrait prediction model, the portrait prediction model may be fine-tuned periodically. For example, the prediction model training data in each training cycle can be obtained, and the prediction model training data of the current training cycle can be used to fine-tune the portrait prediction model of the previous training cycle (or the basic large language model) to obtain the portrait prediction model of the current training cycle.

It should be noted that, since the large language model itself has reasoning capabilities, the portrait prediction model is used to predict user portraits in the embodiments of the present application. It is possible to extract application click behavior features from click behavior data through the reasoning capabilities of the large language model itself, and construct a logical chain of application click behavior. Compared with traditional machine learning, there is no need to manually extract features. The large language model can automatically analyze which features are more meaningful for application recommendations, so that the automatic reasoning capabilities of the large voice model can be used to automatically reason out application click rules that cannot be discovered or summarized manually, thereby improving the accuracy of user portrait prediction. In addition, compared with using expert experience rules to construct user prediction portraits, there is no need to manually construct experience rules, which avoids the limitations of manual experience and improves the accuracy of user portrait prediction.

In addition, since the large language model can process natural language, when using the large language model as a user prediction model, the user only needs to describe the relevant click behavior in text and input the click behavior data that is closer to human natural language to accurately predict the user portrait, which reduces the difficulty of using the user prediction model and improves the user experience.

After introducing the portrait prediction model, we will continue to explain the user prediction portrait.

User prediction portrait, that is, the user portrait predicted by the user prediction model. Exemplarily, the user prediction portrait can be descriptive information that can represent the user's application usage habits. For example, it may include but is not limited to one or more of the following user dimension information: gender, age, occupation, hobbies, usage habits, etc. Among them, the above-mentioned user dimension information can refer to the relevant description of the above part of the embodiment of the present application in combination with Figure 7, which will not be repeated here. In an embodiment of the present application, if the click behavior data of the terminal device is input into the user portrait prediction model, the cloud can obtain the user prediction portrait of the user to which the terminal device belongs. At this time, it can be considered that the user prediction portrait corresponds to the user, or the user prediction portrait corresponds to the terminal device.

Optionally, in the case where the click behavior data includes user dimension information, the cloud can use the user dimension information in the click behavior data to modify the user prediction portrait. For example, if the click behavior data includes objective user dimension information (user information that is more objective or more fixed, such as gender, age, and occupation), the objective user dimension information in the user prediction portrait can be replaced with the objective user dimension information in the click behavior data. For another example, if the click behavior data includes subjective user dimension information (user information that is more subjective or more easily changed, such as usage habits and hobbies), the subjective user dimension information in the user prediction portrait may not be modified.

Optionally, when the click behavior data does not include user dimension information, the cloud may not modify the user prediction portrait.

After introducing concepts such as the portrait prediction model and user prediction portrait, the implementation method of S102 will be described next.

In some embodiments, the cloud can periodically perform user portrait prediction. Specifically, each terminal device can send click behavior data to the cloud, and the cloud can perform user portrait prediction based on the click behavior data obtained in the current time period. Exemplarily, for each terminal device, the cloud inputs the click behavior data of the terminal device obtained in the current time period into the portrait prediction model to obtain the predicted user portrait corresponding to the terminal device.

In some embodiments, after acquiring the click behavior data, the cloud can perform a text representation on it (for example, convert it into a text format), generate a description text in a fixed text format (i.e., a data form shown in a text format for describing the click behavior data), and use the description text and the portrait prediction model to generate a user prediction portrait. Exemplarily, the cloud can perform a text description from multiple information dimensions of the click behavior data (such as the multiple information dimensions shown in FIG. 7). For example, in the case where the click behavior data does not include user dimension information, the text format of the description text may include "time is xxxx (time dimension information), location is xxxx (space dimension information), device status is xxxx (device dimension information), and xxxx application is clicked (application dimension information)". For another example, in the case where the click behavior data includes user dimension information, the text format of the description text may also include "user is xxxx (user dimension information)". And, it should be noted that in order to ensure the prediction reliability of the user portrait, the cloud can first perform a text representation on the click behavior data as a sample and then input it into the basic large language model during the training process of the portrait recommendation model to train the portrait recommendation model.

Optionally, in order to improve the accuracy of portrait prediction, when one or more dimensional information of a certain click behavior data is missing, the cloud can use the dimensional information in the adjacent click behavior data sent by the same terminal device to complete the missing information. For example, after terminal device 1 sends user dimensional information to the cloud, it sends 5 click behavior data in succession. If there is no user dimensional data in the 5 click behavior data sent in succession, the cloud can use the received user dimensional information to complete the 5 click behavior data.

It should be noted that through the text representation of the click behavior data, the cloud can sort out the information of each dimension from the click behavior data and organize it into a fixed format, so that the format of the click behavior data sent by different terminal devices is more consistent, which is convenient for portrait prediction model processing and improves the model prediction effect.

In some embodiments, the portrait prediction model of the large language model can output a portrait description text (i.e., a text sentence used to describe the predicted user portrait). The cloud can perform keyword extraction on the portrait description text to obtain key portrait description words and sentences (i.e., the predicted user portrait. Specifically, the key portrait description words and sentences are key words and/or key sentences extracted from the portrait description text to describe the user portrait, such as "programmer" (key words), "like to browse short video applications late at night" (key sentences)). Among them, the predicted user portrait can include multiple portrait tags.

In one example, S102 is specifically described by taking the portrait prediction process of terminal device 1 as an example. FIG. 11 is a schematic diagram of the process of user portrait prediction shown as an example. After the cloud obtains the click behavior data of terminal device 1 (including time dimension information, space dimension information, device dimension information and application dimension information, and optionally, user dimension information), the click dimension information can be formatted and the description text 1 can be obtained. Then the description text 1 is input into the portrait prediction model to obtain the portrait description text of the user prediction portrait "User A is a 28-year-old female user, a programmer by profession, and has the habit of using food delivery applications at noon and video playback applications at night, and likes to chase dramas and watch movies". After the cloud extracts keywords from the portrait description text, the predicted portrait of user 1 can be obtained. Among them, the predicted portrait of user 1 can include multiple user portrait tags, such as "28 years old (age)", "female (gender)", "programmer (occupation)", "use video playback applications at night (usage habit 1)", "use food delivery applications at noon (usage habit 2)", "chase dramas (hobby 1)" and "take pictures (hobby 2)".

After introducing the prediction process of user portraits in S102, the user group segmentation process will be described in conjunction with S103.

S103, the cloud divides the user portrait groups based on the user prediction portraits.

In S103, the cloud can divide the terminal devices (or user prediction portraits) to obtain multiple user portrait groups. Each user portrait group may include at least one terminal device (or user prediction portrait). Exemplarily, referring to FIG. 9 , the cloud can divide the terminal devices according to the user 1 prediction portrait, the user 2 prediction portrait, and the user 3 prediction portrait through the user portrait group division unit, and divide the terminal device 1 and the terminal device 3 into the same user portrait group, and divide the terminal device 2 into another user portrait group.

In some embodiments, in order to ensure the accuracy of the portrait, the cloud can periodically divide the user portrait groups. For example, the cloud can divide the user portrait groups once a month. It should be noted that the specific period can be set according to the actual scenario and specific needs, and there is no specific restriction on this.

In some embodiments, the cloud can divide user groups based on some typical users of known user groups (such as user portrait groups 1 and 2 in the attached figure). Exemplarily, the cloud can vectorize all user prediction portraits to represent the user prediction portraits as multidimensional vectors. Then, the similarity measurement (such as cosine similarity) is performed on the user prediction portraits in the form of multidimensional vectors of each user and some typical users of the known user group (i.e., the user portraits in the form of multidimensional vectors of typical users). Then, among the user groups to which the K typical users with the highest similarity ranking belong, the user group with the largest number of users is used as the group division result of the user (for example, for each known user group, the number of users belonging to the user group can be counted among the K typical users, and then among the number of users in each known group, the known user group with the largest number of users is determined as the group division result of the user). The specific value of K should be greater than the number of known user groups, and can be set according to specific scenarios and actual needs. For example, K=3 is set when dividing two user groups, and there is no specific restriction on this. Exemplarily, 2 or more typical users can be selected from each known user portrait group to divide the user group to ensure the accuracy of the division of the user group.

For example, FIG12 is a flowchart of an exemplary user group division process. The cloud can measure the similarity between each user's predicted portrait and some typical users of the known user group. For example, the similarity between the predicted portrait of user 1 and the typical user P1 of user portrait group 1 is 0.5, the similarity between the predicted portrait of user 1 and the typical user P2 of user portrait group 1 is 0.7, the similarity between the predicted portrait of user 1 and the typical user P3 of user portrait group 2 is 0.3, and the similarity between the predicted portrait of user 1 and the typical user P4 of user portrait group 2 is 0.4. When K=3, the cloud can take the user portrait group 1 with the largest number of user groups among the three users with the highest similarity (typical user P2 of user portrait group 1, typical user P1 of user portrait group 1, and typical user P4 of user portrait group 2) as the user portrait group of user 1.

Similarly, the similarity between the predicted portrait of user 2 and the typical user P1 of user portrait group 1 is 0.3, the similarity between the predicted portrait of user 2 and the typical user P2 of user portrait group 1 is 0.5, the similarity between the predicted portrait of user 2 and the typical user P3 of user portrait group 2 is 0.6, and the similarity between the predicted portrait of user 2 and the typical user P4 of user portrait group 2 is 0.7. The similarity between the predicted portrait of user 3 and the typical user P1 of user portrait group 1 is 0.8, the similarity between the predicted portrait of user 3 and the typical user P2 of user portrait group 1 is 0.7, the similarity between the predicted portrait of user 3 and the typical user P3 of user portrait group 2 is 0.6, and the similarity between the predicted portrait of user 3 and the typical user P4 of user portrait group 2 is 0.4. Therefore, user 2 belongs to user portrait group 2 and user 3 belongs to user portrait group 1.

In one embodiment, in order to reduce the training cost, the portrait prediction model may include a vector representation layer. Taking the user predicted portrait as the key portrait description phrase as an example, the cloud can input the key portrait description phrase of a single terminal device into the portrait prediction model, and use the intermediate output of the portrait prediction model (such as the output of the vector representation layer of the user predicted portrait) as the representation vector of the key portrait description phrase of the single terminal device. Then, based on the representation vectors of the key portrait description phrases of the multiple terminal devices, the multiple terminal devices can be divided into user groups to obtain multiple user portrait groups.

In one example, the method for obtaining some typical users of a known user group may include: clustering the representation vectors of multiple user portraits, and sampling some users from each cluster (each cluster corresponds to a known user group). Alternatively, based on manual settings of user analysis, the implementation example of this application does not limit the specific acquisition setting method.

In other embodiments, the cloud can use a clustering algorithm to divide the user prediction portrait into user groups. For example, the cloud can use a clustering algorithm such as Kmeans and FCM. Exemplarily, the cloud can vectorize the user prediction portrait as a multi-dimensional vector, and then use a clustering algorithm to cluster the user prediction portrait in the form of a multi-dimensional vector, so as to obtain the user portrait group to which each user belongs.

Since each user prediction portrait corresponds to a terminal device (or a user), through the embodiment of the present application, based on the user prediction portrait, the cloud can accurately classify the terminal device according to the user's application usage habits, thereby improving the accuracy of user group division. Furthermore, when the cloud trains the recommendation model based on the click behavior data of the same user portrait group, since the users of the same user portrait group have similar application usage habits, it can provide more accurate and more data support for the training of the recommendation model.

After introducing the user group division process in S103, the training process of the recommendation model is described in conjunction with S104.

S104, the cloud performs recommendation model training for each user portrait group based on the click behavior data of each user portrait group.

In S104, in each user portrait group, the cloud can use the click behavior data corresponding to each user prediction portrait (the click behavior data corresponding to the user prediction portrait of each terminal device, that is, the click behavior data of the terminal device) to train the basic recommendation model to obtain the recommendation model of the user portrait group. Exemplarily, with continued reference to Figures 9 and 12, when terminal device 1 (corresponding to the predicted portrait of user 1) and terminal device 3 (corresponding to the predicted portrait of user 3) belong to the same user portrait group (such as user portrait group 1), (terminal device 2 corresponding to the predicted portrait of user 2) belongs to another user portrait group (such as user portrait group 2), the recommendation model A training unit in the cloud can use training data 1 (the click behavior data of terminal device 1) and training data 3 (the click behavior data of terminal device 3) to train recommendation model A (that is, the recommendation model of user portrait group 1), and the recommendation model B training unit in the cloud can use training data 2 (the click behavior data of terminal device 2) to train recommendation model B (that is, the recommendation model of user portrait group 2). It should be noted that in S104, the cloud can use the click behavior data of some or all terminal devices in each user portrait group to train the recommendation model of the user portrait group. Also, the cloud can use some or all click behavior data of a single terminal device for training, and there is no specific restriction on this.

The recommendation model is used to predict the applications that the user is about to use and recommend the predicted applications to the user. Exemplarily, the recommendation model can be a deep network model or a large language model, which is not specifically limited. In an embodiment of the present application, a user portrait group can correspond to a recommendation model, and different user portrait groups correspond to different recommendation models.

For the training process of the recommendation model, exemplarily, as shown in FIG13 , there is an exemplary training flow diagram of the recommendation model. Among them, the click behavior data of each user may include application dimension information, and may also include other click behavior data (such as time dimension information, space dimension information, device dimension information, etc.) in addition to the application dimension information. Among them, other click behavior data can be used as samples, and application dimension information (in this click behavior, the application information of the application clicked by the user) can be used as a label.

Specifically, when terminal device 1 and terminal device 3 belong to user portrait group 1, and terminal device 2 belongs to user portrait group 2, the cloud can use the click behavior data of user 1 (training data of recommendation model A) and the click behavior data of user 3 (training data of recommendation model A) to train the initial model (model to be trained), and when the training conditions (such as loss function convergence) are met, the recommendation model A (trained application recommendation model) is obtained, wherein the recommendation model A corresponds to user portrait group 1. In addition, the cloud can use the click behavior data of user 2 (training data of recommendation model B) to train the initial model (model to be trained), and when the training conditions are met, the recommendation model B (trained application recommendation model) is obtained, wherein the recommendation model B corresponds to user portrait group 2.

It should be noted that since users corresponding to the same user portrait group often have similar application usage habits, by using the click behavior data of the same user portrait group to train the recommendation model, the trained recommendation model can accurately simulate the click habits of this type of users, and thus can accurately predict the application that the user is about to click for recommendation, thereby improving the accuracy of application recommendations.

For example, for each click behavior data, the cloud can extract features from other click behavior data therein to extract the scene features of the click behavior (the application click behavior that generates the click behavior data). The cloud then inputs the scene features into the recommendation model to be trained for model training.

After introducing the training process of the recommendation model through S104, the distribution process of the recommendation model will be described in conjunction with S105.

S105, the cloud sends the recommendation model of each user portrait group to the terminal device corresponding to each user portrait group.

In S105, for each user portrait group, the cloud can send the recommendation model of the user portrait group to the terminal device corresponding to the user portrait group (i.e., the terminal device corresponding to each user prediction portrait in the user portrait group). Exemplarily, referring to FIG9 , the recommendation model A of user portrait group 1 can be sent to terminal device 1 and terminal device 3, and the recommendation model B of user portrait group 2 can be sent to terminal device 2. It should be noted that the cloud can send the recommendation model to all terminal devices in a single user portrait group, or send the recommendation model to the terminal device with the application recommendation function enabled in a single user portrait group, without limitation.

In some embodiments, the cloud can send the model parameters of the recommendation model to each terminal device, and each terminal device updates the model by updating the model parameters of the local recommendation model. For example, FIG14 is a schematic diagram of the distribution process of the recommendation model. After the cloud obtains the recommendation model A (the application recommendation model corresponding to the user portrait group 1) through training, the model parameters of the recommendation model A can be sent to the terminal device 1, so that the terminal device 1 can make application recommendations based on the recommendation model A.

After introducing the training scenario of the recommendation model, the following describes the usage scenario of the recommendation model in conjunction with the method flow diagram shown in FIG15 , referring to FIG15 , which specifically includes:

S201, the terminal device recommends applications, collects and uploads data based on the recommendation model.

In S201, for application recommendation, the terminal device can obtain current scene information and input the current scene information into the recommendation model to obtain recommended applications when the preset recommendation conditions are met. For example, FIG16 is a schematic diagram of the use process of the recommendation model shown exemplarily. Taking terminal device 1 as an example, terminal device 1 can use recommendation model A to recommend applications and display the recommended applications in the recommendation folder.

In some embodiments, the recommended application can be displayed in a suggestion folder of the terminal device. In the embodiment of the present application, by placing the recommended application in the suggestion folder, the recommended application can be placed in a more intuitive entrance, which is convenient for users to find the application they want to use, thereby improving the convenience of application use.

In order to facilitate understanding of the application recommendation process, the preset recommendation conditions and current scenario information involved in the application recommendation process are first explained.

The preset recommendation condition is the condition that needs to be met when an application recommendation is required. In one example, the preset recommendation condition may include detecting that the display interface of the terminal device jumps to the main interface. For example, when the user switches from the application interface to the main interface, the recommended application can be displayed on the main interface of the terminal device. In another example, the preset recommendation condition may include detecting that the usage scenario of the terminal device has changed. Among them, the usage scenario can be a scenario that can affect the user's application usage preference, such as time, location, and device status. For example, the terminal device can re-recommend applications to the user when it detects the noon period. At this time, the terminal device can recommend catering applications to the user. For another example, the terminal device can re-recommend applications to the user after detecting that the user moves from home to the subway station, and the terminal device can recommend travel applications to the user. For another example, the terminal device can re-recommend applications to the user after detecting that the user is connected to a Bluetooth headset, and the terminal device can recommend music applications to the user. In another example, the preset recommendation condition may include detecting that the interval between the last recommendation time and the preset time is long. That is to say, the terminal device can periodically recommend applications. It should be noted that the terminal device can also set other preset recommendation conditions according to specific recommendation scenarios and recommendation requirements, and there is no specific restriction on this.

Current scenario information, which may be a usage scenario of a terminal device related to application usage preferences. Exemplarily, it may include one or more of user dimension information, time dimension information, space dimension information, and device dimension information. For its specific content, please refer to the relevant description of click behavior data in the above part of the embodiment of this application, which will not be repeated here. As another example, it may also include application list information in the application dimension information.

After introducing the preset recommendation conditions and current scenario information, the application recommendation process is described in detail.

FIG17 is a flowchart of an exemplary application recommendation process. As shown in FIG17 , continuing to take the terminal device 1 as an example, the terminal device 1 can obtain the current scene information, wherein the current scene information may include time dimension information, space dimension information, device dimension information, application dimension information, etc. And, the terminal device 1 extracts the scene features of the obtained current scene information, inputs the extracted scene features into the recommendation model A, and obtains the selection probability of multiple applications. Then, from multiple applications, select N (N is any positive integer, such as 4) applications that meet the preset recommendation rules, such as application A, application B, application D and application E. Then, application A, application B, application D and application E are displayed in the display interface of the terminal device 1 (such as the recommended folder on the main page of the terminal device 1). Optionally, in order to improve the user experience, according to the selection probability of each application, the M (M is a positive integer less than or equal to N, such as 2) applications with the highest selection probability can also be preloaded, such as preloading application A and application B.

Among them, the preset recommendation rules are the rules or conditions that the recommended applications need to meet. Exemplarily, the preset recommendation rules may include: selecting N downloaded applications with the highest probability of being selected. Among them, the downloaded applications are applications that have been downloaded on the terminal device 1. For example, referring to Figure 17, if the order of selection probability from high to low is application A, application B, application C, application D and application E, and application C is not downloaded, then when 4 applications need to be selected, application A, application B, application D and application E can be selected. In another exemplary embodiment, the preset recommendation rules may include: selecting in order of selection probability from high to low, and the number of selected applications in each application category is less than or equal to K. Among them, K is a positive integer less than or equal to N, for example, 2. For example, referring to Figure 17, if the order of selection probability from high to low is Application A, Application B, Application C, Application D, Application E, and Application F, if Application A, Application D, and Application E are all short video applications, and Application C has not been downloaded, then when you need to select 4 applications, you can choose Application A, Application B, Application D, and Application F.

Optionally, in order to improve the user's downloading experience, the terminal device may display applications that have a high probability of being selected and have not been downloaded in the recommended download folder. For example, application C may be displayed in the recommended download folder of the terminal device 1.

After introducing the application recommendation process, the process of data collection and uploading is explained next.

In S201, after the user performs an application selection operation on the application interface of the terminal device displaying the recommended application, the terminal device can collect the click behavior data generated by the local click operation. And upload the local click behavior data to the cloud for the cloud to train the recommendation model. Among them, the local click behavior data can refer to the relevant description of the above part of the embodiment of this application, which will not be repeated here.

18a is a schematic diagram of another terminal display interface, and taking terminal device 1 as an example, the home page of terminal device 1 can display suggested folders and other folders. When the user clicks on the other application folder, the display screen of terminal device 1 displays multiple other applications in the other application folder.

And, as shown in Figure 18b, which is another exemplary end-cloud interaction diagram, when the user clicks on "Communication Application B" in other application folders, the terminal device 1 detects the local application click behavior, and can generate the click behavior data and send the click behavior data to the cloud.

In the embodiment of the present application, the terminal devices used for application recommendation, data collection and uploading in S201 may be part or all of the terminal devices in the system (such as shown in FIG. 2 ), and no specific limitation is imposed on this.

S202, the terminal device periodically updates the model based on the collected data.

Exemplarily, the terminal device can periodically (such as daily) update the local recommendation model based on the click behavior data collected over a period of time. For example, FIG. 18c is another schematic diagram of a terminal display interface shown as an example. As shown in FIG. 18c, the terminal device 1 can update the local recommendation model A. And, with the update of the recommendation model, under the same scenario information, the recommended applications of the terminal device 1 also have certain differences. Continuing to refer to FIG. 18c, before and after the update of the recommendation model A, the recommended folder of the terminal device 1 also changed, that is, the application icon 181 of the browser A changed to the application icon 182 of the communication application B.

Exemplarily, each terminal device can update the recommendation model based on its own click behavior data. For example, as shown in FIG19, an exemplary schematic diagram of the local update process of the recommendation model is shown. After the cloud sends the same recommendation model A to the terminal device 1 and the terminal device 3, the terminal device 1 can locally update the recommendation model A based on its own click behavior data to obtain the recommendation model A'. In an example, as shown in FIG20, an exemplary schematic diagram of the local update process of the recommendation model is shown. After the user performs an application selection operation on the application display interface, the terminal device obtains the current click behavior data, wherein the current click behavior data may include one or more of time dimension information, space dimension information, device dimension information, application dimension information, etc. The terminal device uses the feature embedding representation module to extract features from the current click behavior data, and the scene features can be obtained in advance. Then the terminal device uses the scene features (for example, the clicked application can be used as a label, and the scene features extracted from other click behavior data can be used as samples) to train the recommendation model A, and obtain the recommendation model A' when the training requirements are met. Exemplarily, in order to ensure the accuracy of the recommendation model, the feature embedding representation module of the terminal device is the same as the feature embedding representation module of the cloud, so that the same scene features can be extracted from the same click behavior data, thereby ensuring the consistency of the local recommendation model training of the terminal device and the recommendation model training of the cloud.

And, continuing to refer to Figure 19, the terminal device 3 can locally update the recommendation model A based on its own click behavior data to obtain the recommendation model A". That is to say, different terminal devices corresponding to the same user portrait group receive the same recommendation model. Under the same scenario information, if the downloaded applications are the same, the applications recommended by different terminal devices are the same. However, as the user's click behavior changes, the local recommendation models of different terminal devices also change accordingly. Accordingly, the applications recommended by different terminal devices in the same scenario also have personalized differences.

Optionally, FIG21 is a schematic diagram of an exemplary update process of a recommendation model. The update process of the recommendation model may include:

S2021, the terminal device obtains the click behavior data within the current time period.

The current time period may be an update period of the recommendation model, for example, every day.

It should be noted that the click behavior data can be found in the relevant description of the above part of the embodiment of the present application, which will not be repeated here.

S2022, the terminal device determines whether the application recommendation accuracy rate in the current time period is greater than a preset accuracy rate threshold. If the determination result is no, step S2023 is executed. If the determination result is yes, step S2024 is executed.

Among them, the application recommendation accuracy is used to characterize the accuracy of the application recommendation of the recommendation model locally stored in the terminal device within the current time period, and the preset accuracy threshold is used to measure whether the recommendation model locally stored in the terminal device meets the accuracy requirements of the application recommendation. Among them, the preset accuracy threshold can be set to any value according to the actual scenario and specific needs, and there is no specific restriction on this.

In some embodiments, the terminal device may determine the exact number of clicks in the current time period (i.e., the number of times the recommended application is clicked in the current time period), and the total number of application clicks determined by the terminal device in the current time period (i.e., the total number of times the user clicks on the application in the current time period). Then, the terminal device determines the ratio of the exact number of clicks to the total number of clicks as the application recommendation accuracy rate for the current time period. Exemplarily, for each click behavior in the current time period, if the application clicked by the user is not an application icon of a recommended application in the recommended folder, the terminal device determines that this application recommendation is wrong, and keeps the exact number of recommendations in the current time period unchanged. Similarly, if the application clicked by the user is an application icon of any recommended application in the recommended folder, the terminal device determines that this click behavior is accurate, and adds 1 to the exact number of clicks in the current time period. For example, if the user clicks on the application 18 times in total in the current time period, of which 15 times are recommended applications, the application recommendation accuracy rate for the current time period may be 0.833 (i.e., 15/18).

S2023, the terminal device updates the recommendation model A using the click behavior data in the current time period to obtain the recommendation model A'.

Exemplarily, in the next time period, the terminal device can continue to recommend applications using recommendation model A', and when the application recommendation accuracy in the next time period is less than a preset accuracy threshold, the terminal device can continue to update recommendation model A' using the click behavior data generated in the next time period.

It should be noted that S2023 can refer to the relevant description of S104 in the embodiment of the present application, and will not be repeated here.

S2024: The terminal device does not update the recommended model A.

It should be noted that, in the embodiment of the present application, the terminal device that performs model updating in S202 may be part or all of the terminal devices in the system (eg, as shown in FIG. 2 ), and no specific limitation is imposed on this.

It is understandable that, in order to realize the above functions, the electronic device includes hardware and/or software modules corresponding to the execution of each function. In combination with the algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application in combination with the embodiments, but such implementation should not be considered to be beyond the scope of the present application.

In an example, FIG22 shows a schematic block diagram of a device 300 according to an embodiment of the present application. The device 300 may include: a processor 301 and a transceiver/transceiver pin 302 , and optionally, a memory 303 .

The components of the device 300 are coupled together via a bus 304, wherein the bus 304 includes a power bus, a control bus, and a status signal bus in addition to a data bus. However, for the sake of clarity, all buses are referred to as bus 304 in the figure.

Optionally, the memory 303 may be used for the instructions in the aforementioned method embodiment. The processor 301 may be used to execute the instructions in the memory 303, and control the receiving pin to receive a signal, and control the sending pin to send a signal.

The apparatus 300 may be the electronic device or a chip of the electronic device in the above method embodiment.

Among them, all relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding functional module, and will not be repeated here.

The steps performed by the terminal device in the application recommendation method provided in the above embodiment of the present application can also be performed by a chip system included in the terminal device, wherein the chip system may include a processor and a Bluetooth chip. The chip system can be coupled to a memory so that when the chip system is running, it calls a computer program stored in the memory to implement the steps performed by the above terminal device. The processor in the chip system can be an application processor or a processor other than an application processor.

Similarly, in the above embodiment, the steps executed in the cloud can also be executed by a chip system included in the cloud, wherein the chip system can include a processor and a Bluetooth chip. The chip system can be coupled to a memory, so that when the chip system is running, it calls a computer program stored in the memory to implement the above steps executed in the cloud. The processor in the chip system can be an application processor or a processor other than an application processor.

As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims (34)

1. An application recommendation method, characterized in that the method is applied in the cloud, and the method comprises: Obtain first click behavior data of multiple terminal devices; Based on the first click behavior data of the multiple terminal devices, obtaining a user prediction portrait corresponding to each terminal device in the multiple terminal devices; Based on the predicted user portrait corresponding to each terminal device in the plurality of terminal devices, dividing the plurality of terminal devices into a plurality of user portrait groups; Performing model training based on the first click behavior data of each terminal device in a single user portrait group to obtain an application recommendation model corresponding to the single user portrait group; The application recommendation model corresponding to the single user portrait group is sent to each terminal device in the single user portrait group, so that each terminal device in the single user portrait group makes application recommendations based on the application recommendation model. 2. The method according to claim 1, characterized in that the step of obtaining a user prediction portrait corresponding to each of the plurality of terminal devices based on the first click behavior data of the plurality of terminal devices comprises: Input the first click behavior data of a single terminal device into the portrait prediction model to obtain the user prediction portrait corresponding to the single terminal device. Among them, the portrait prediction model is a large language model. 3. The method according to claim 2, characterized in that the step of obtaining a user prediction portrait corresponding to each of the plurality of terminal devices based on the first click behavior data of the plurality of terminal devices comprises: Performing text representation on the first click behavior data of a single terminal device to generate a description text in a fixed text format; The description text is input into the portrait prediction model to obtain a user prediction portrait corresponding to the single terminal device. 4. The method according to claim 2, wherein the first click behavior data of the single terminal device includes user dimension information, and the obtaining of a user prediction portrait corresponding to each of the multiple terminal devices based on the first click behavior data of the multiple terminal devices comprises: Inputting the first click behavior data of the single terminal device into the portrait prediction model to obtain an initial user portrait, wherein the initial user portrait includes user dimension information; The user dimension information in the first click behavior data of the single terminal device is used to correct the user dimension information in the initial user portrait to obtain a user prediction portrait corresponding to the single terminal device. 5. The method according to claim 4, characterized in that the user dimension information includes subjective user dimension information and objective user dimension information, and the user prediction portrait corresponding to the single terminal device includes the objective user dimension information in the first click behavior data of the single terminal device and the objective user dimension information in the initial user portrait; the user dimension information in the first click behavior data of the single terminal device is used to correct the user dimension information in the initial user portrait to obtain the user prediction portrait corresponding to the single terminal device, comprising: Using the objective user dimension information in the first click behavior data of the single terminal device to replace the objective user dimension information in the initial user portrait; The objective user dimension information in the initial user portrait is kept unchanged. 6. The method according to claim 2, characterized in that before inputting the click behavior data of a single terminal device into the portrait prediction model to obtain the user prediction portrait corresponding to the single terminal device, the method further comprises: Acquire multiple prediction model training data and a basic large language model, wherein each prediction model training data includes second click behavior data of a terminal device and a user portrait description text corresponding to the terminal device; The basic large language model is fine-tuned using the multiple prediction model training data to obtain a trained portrait prediction model. 7. The method according to claim 1, characterized in that The acquiring, based on the first click behavior data of the multiple terminal devices, a user prediction portrait corresponding to each terminal device in the multiple terminal devices includes: Inputting the first click behavior data of a single terminal device into a portrait prediction model to obtain a user portrait description text of the single terminal device; Perform keyword extraction on the user portrait description text to obtain key portrait description words and sentences of the single terminal device; Inputting the key portrait description words and sentences of the single terminal device into the portrait prediction model, and using the intermediate output of the portrait prediction model as the representation vector of the key portrait description words and sentences of the single terminal device; The dividing the plurality of terminal devices into a plurality of user portrait groups based on the predicted user portrait corresponding to each of the plurality of terminal devices comprises: Based on the representation vectors of the key portrait description words and sentences of the multiple terminal devices, the multiple terminal devices are divided into user groups to obtain multiple user portrait groups. 8. The method according to claim 1, wherein the first click behavior data comprises: At least one of time dimension information, space dimension information, device dimension information, and application dimension information. 9. An application recommendation method, characterized in that the method is applied to a target terminal device, and the method comprises: Receive an application recommendation model sent by the cloud, the application recommendation model corresponds to the user profile group to which the target terminal device belongs, the application recommendation model is trained by the cloud using the first click behavior data of each terminal device in the user profile group, the user profile group is obtained by dividing the multiple terminal devices based on the user prediction profile corresponding to each terminal device in the multiple terminal devices, and the user prediction profile corresponding to each terminal device in the multiple terminal devices is determined based on the first click behavior data of the multiple terminal devices; Get current scene information; Inputting the current scenario information into the application recommendation model to obtain a recommended application; The recommended applications are displayed. 10. The method according to claim 9, characterized in that after displaying the recommended application, the method further comprises: Detect the user's application selection operation; In response to the application selection operation, generating third click behavior data; The application recommendation model is updated using the third click behavior data in the current time period. 11. The method according to claim 10, characterized in that the updating of the application recommendation model by using the third click behavior data in the current time period comprises: Determining the application recommendation accuracy of the target terminal device in the current time period based on the third click behavior data in the current time period; When the application recommendation accuracy is less than a preset accuracy threshold, the application recommendation model is updated using the third click behavior data in the current time period. 12. The method according to claim 10, characterized in that after generating the third click behavior data, the method further comprises: The third click behavior data is sent to the cloud, so that the cloud performs fine-tuning training of the recommendation model based on the third click behavior data. 13. The method according to claim 9, wherein the step of inputting the current scenario information into the application recommendation model to obtain a recommended application comprises: Inputting the current scenario information into the application recommendation model to obtain selection probabilities of multiple candidate applications; The recommended application is selected from the multiple candidate applications based on the selection probabilities of the multiple candidate applications and a preset recommendation rule. 14. The method according to claim 9, characterized in that a suggested folder is displayed on the main interface of the terminal device; The displaying of the recommended application includes: The recommended application is displayed in the suggestion folder. 15. The method according to claim 9, characterized in that The user prediction portrait corresponding to each terminal device among the multiple terminal devices is obtained by processing the first click behavior data of each terminal device using the portrait prediction model. The portrait prediction model is a large language model. 16. The method according to claim 9, wherein the first click behavior data comprises: At least one of time dimension information, space dimension information, device dimension information, and application dimension information. 17. A cloud server, comprising: one or more processors; Memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and when the computer programs are executed by the one or more processors, the cloud server performs the following steps: Obtain first click behavior data of multiple terminal devices; Based on the first click behavior data of the multiple terminal devices, obtaining a user prediction portrait corresponding to each terminal device in the multiple terminal devices; Based on the predicted user portrait corresponding to each terminal device among the multiple terminal devices, dividing the multiple terminal devices into multiple user portrait groups; Performing model training based on the first click behavior data of each terminal device in a single user portrait group to obtain an application recommendation model corresponding to the single user portrait group; The application recommendation model corresponding to the single user portrait group is sent to each terminal device in the single user portrait group, so that each terminal device in the single user portrait group makes application recommendations based on the application recommendation model. 18. The cloud server according to claim 17, wherein when the computer program is executed by the one or more processors, the cloud server performs the following steps: Input the first click behavior data of a single terminal device into the portrait prediction model to obtain the user prediction portrait corresponding to the single terminal device. Among them, the portrait prediction model is a large language model. 19. The cloud server according to claim 18, wherein when the computer program is executed by the one or more processors, the cloud server performs the following steps: Performing text representation on the first click behavior data of a single terminal device to generate a description text in a fixed text format; The description text is input into the portrait prediction model to obtain a user prediction portrait corresponding to the single terminal device. 20. The cloud server according to claim 18, wherein the first click behavior data of the single terminal device includes user dimension information, and when the computer program is executed by the one or more processors, the cloud server performs the following steps: Inputting the first click behavior data of the single terminal device into the portrait prediction model to obtain an initial user portrait, wherein the initial user portrait includes user dimension information; The user dimension information in the first click behavior data of the single terminal device is used to correct the user dimension information in the initial user portrait to obtain a user prediction portrait corresponding to the single terminal device. 21. The cloud server according to claim 20, characterized in that the user dimension information includes subjective user dimension information and objective user dimension information, and the user prediction portrait corresponding to the single terminal device includes the objective user dimension information in the first click behavior data of the single terminal device and the objective user dimension information in the initial user portrait; When the computer program is executed by the one or more processors, the cloud server performs the following steps: Using the objective user dimension information in the first click behavior data of the single terminal device to replace the objective user dimension information in the initial user portrait; The objective user dimension information in the initial user portrait is kept unchanged. 22. The cloud server according to claim 18, wherein when the computer program is executed by the one or more processors, the cloud server performs the following steps: Acquire multiple prediction model training data and a basic large language model, wherein each prediction model training data includes second click behavior data of a terminal device and a user portrait description text corresponding to the terminal device; The basic large language model is fine-tuned using the multiple prediction model training data to obtain a trained portrait prediction model. 23. The cloud server according to claim 17, wherein when the computer program is executed by the one or more processors, the cloud server performs the following steps: Inputting the first click behavior data of a single terminal device into a portrait prediction model to obtain a user portrait description text of the single terminal device; Perform keyword extraction on the user portrait description text to obtain key portrait description words and sentences of the single terminal device; Inputting the key portrait description words and sentences of the single terminal device into the portrait prediction model, and using the intermediate output of the portrait prediction model as the representation vector of the key portrait description words and sentences of the single terminal device; Based on the representation vectors of the key portrait description words and sentences of the multiple terminal devices, the multiple terminal devices are divided into user groups to obtain multiple user portrait groups. 24. The cloud server according to claim 17, wherein the first click behavior data comprises: At least one of time dimension information, space dimension information, device dimension information, and application dimension information. 25. An electronic device, comprising: one or more processors; Memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and when the computer programs are executed by the one or more processors, the electronic device performs the following steps: Receive an application recommendation model sent by a cloud, the application recommendation model corresponding to a user profile group to which a target terminal device belongs, the application recommendation model being trained by the cloud using first click behavior data of each terminal device in the user profile group, the user profile group being obtained by dividing the multiple terminal devices based on a predicted user profile corresponding to each terminal device in a plurality of terminal devices, the predicted user profile corresponding to each terminal device in the plurality of terminal devices being determined based on the first click behavior data of the plurality of terminal devices; Get current scene information; Inputting the current scenario information into the application recommendation model to obtain a recommended application; The recommended applications are displayed. 26. The electronic device according to claim 25, characterized in that when the computer program is executed by the one or more processors, the electronic device performs the following steps: Detect the user's application selection operation; In response to the application selection operation, generating third click behavior data; The application recommendation model is updated using the third click behavior data in the current time period. 27. The electronic device according to claim 26, wherein when the computer program is executed by the one or more processors, the electronic device performs the following steps: Determining the application recommendation accuracy of the target terminal device in the current time period based on the third click behavior data in the current time period; When the application recommendation accuracy is less than a preset accuracy threshold, the application recommendation model is updated using the third click behavior data in the current time period. 28. The electronic device according to claim 26, wherein when the computer program is executed by the one or more processors, the electronic device performs the following steps: The third click behavior data is sent to the cloud, so that the cloud performs fine-tuning training of the recommendation model based on the third click behavior data. 29. The electronic device according to claim 25, characterized in that when the computer program is executed by the one or more processors, the electronic device performs the following steps: Inputting the current scenario information into the application recommendation model to obtain selection probabilities of multiple candidate applications; The recommended application is selected from the multiple candidate applications based on the selection probabilities of the multiple candidate applications and a preset recommendation rule. 30. The electronic device according to claim 25, characterized in that a suggested folder is displayed on the main interface of the terminal device, and when the computer program is executed by the one or more processors, the electronic device performs the following steps: The recommended application is displayed in the suggestion folder. 31. The electronic device according to claim 25, characterized in that: The user prediction portrait corresponding to each terminal device among the multiple terminal devices is obtained by processing the first click behavior data of each terminal device using the portrait prediction model. The portrait prediction model is a large language model. 32. The electronic device according to claim 25, wherein the first click behavior data comprises: At least one of time dimension information, space dimension information, device dimension information, and application dimension information. 33. A chip, characterized in that it comprises one or more interface circuits and one or more processors; The interface circuit is used to receive a signal from a memory of the cloud server and send the signal to the processor, wherein the signal includes computer instructions stored in the memory; and when the processor executes the computer instructions, the cloud server executes the application recommendation method described in any one of claims 1 to 8. 34. A chip, characterized in that it comprises one or more interface circuits and one or more processors; The interface circuit is used to receive a signal from a memory of an electronic device and send the signal to the processor, wherein the signal includes computer instructions stored in the memory; and when the processor executes the computer instructions, the terminal device executes the application recommendation method described in any one of claims 9 to 16.