CN115774817A - Information processing model training method, information processing method and related equipment - Google Patents

Abstract

This application relates to the field of artificial intelligence, and specifically discloses an information processing model training method, information processing method and related equipment. The training method includes: obtaining the interaction information sequence of the first user and the interaction information sequence of the second user; according to the A user's first user feature generates the first user's local feature and the first user's global feature; and generates the second user's local feature according to the second user's first user feature; according to the first user's local feature and the first Calculating the first score based on the global features of the user; and calculating the second score based on the global features of the first user and the local features of the second user; pairing the first neural network model, the user feature model and the discriminant model according to the first score and the second score Carry out training; determine an information processing model according to the trained first neural network model and the trained user feature model. This scheme can effectively guarantee the matching between the target information determined in the target domain and the user.

Description

Information processing model training method, information processing method and related equipment

technical field

The present application relates to the technical field of artificial intelligence, and more specifically, to an information processing model training method, an information processing method and related equipment.

Background technique

With the development of Internet technology, more and more users share and disseminate information through the sending platform. After the sending platform receives the user's information request, the sending platform screens the information sent to the user from the information database. In related technologies, the information sent by the sending platform to the user is not highly compatible with the user, resulting in a large difference between the sent information and the information actually needed by the user, which in turn requires the user to interact with the sending platform multiple times, and the sending platform multiple times. Only by filtering and sending information can users obtain the information they need. Since the information sending platform needs to perform information screening and information sending many times, the efficiency of information sending is not high.

Contents of the invention

In view of the above problems, the embodiment of the present application proposes an information processing model training method, an information processing method, and related equipment, so as to solve the problem of low information transmission efficiency caused by the low matching between information and users.

According to an aspect of an embodiment of the present application, there is provided a method for training an information processing model, including: acquiring the interaction information sequence of the first user and the interaction information sequence of the second user from the first user behavior data in the source domain; The first neural network model outputs the first user features of the first user according to the interaction information sequence of the first user, and the first neural network model outputs the first user feature according to the interaction information sequence of the second user. The first user feature of the second user; the first neural network model is determined according to the pre-training of the second user historical behavior data in the source domain; the user feature model generates the first user feature according to the first user feature The local features of the first user and the global features of the first user; and the local features of the second user are generated by the user feature model according to the first user features of the second user; calculating a first score based on the local features of the user and the global features of the first user; and calculating a second score based on the global features of the first user and the local features of the second user by the discriminant model; according to the The first score and the second score train the first neural network model, the user feature model and the discriminant model until the training end condition is reached; according to the trained first neural network model and the training The latter user characteristic model determines an information processing model, and the information processing model determines the target information to be sent to the user in the target domain.

According to an aspect of an embodiment of the present application, an information processing method is provided, including: acquiring a set of candidate information in a target domain; predicting a target user's behavior label for each information in the set of candidate information through the information processing model; The information processing model is trained according to the information processing model training method described above; according to the behavior label, the target information to be sent to the target user is determined in the candidate information set.

According to an aspect of an embodiment of the present application, an information processing model training device is provided, including: a first user behavior data acquisition module, configured to acquire the interaction information sequence of the first user from the first user behavior data in the source domain and the interaction information sequence of the second user; the first output module is used to output the first user characteristics of the first user according to the interaction information sequence of the first user by the first neural network model, and the second output module, The first neural network model is used to output the first user characteristics of the second user according to the interaction information sequence of the second user; the first neural network model is based on the historical behavior data of the second user in the source domain Determined by pre-training; the first generation module is used to generate the local characteristics of the first user and the global characteristics of the first user according to the first user characteristics of the first user by the user characteristic model; and the second generation module , used to generate the local features of the second user according to the first user features of the second user by the user feature model; the first score calculation module is used to use the discriminant model according to the local features of the first user calculating a first score with the global features of the first user; and a second score calculation module, configured to calculate a second score by the discriminant model according to the global features of the first user and the local features of the second user ; The first training module is used to train the first neural network model, the user feature model and the discriminant model according to the first score and the second score until the training end condition is reached; information processing A model determination module, configured to determine an information processing model according to the trained first neural network model and the trained user characteristic model, and the information processing model determines the target information to be sent to the user in the target domain.

In some embodiments of the present application, the user feature model includes at least two convolutional layers and a pooling layer, and the length of the convolution window corresponding to each convolutional layer in the at least two convolutional layers is different; In this embodiment, the first generating module includes: a first convolution processing unit, configured to perform each convolutional layer of the at least two convolutional layers on the first performing convolution processing on the first user features of the user to obtain multiple intermediate local features of the first user; the first concatenation unit is configured to concatenate the multiple intermediate local features of the first user to obtain the first user A local feature of a user; a first pooling processing unit, configured to perform pooling processing on the local feature of the first user by the pooling layer to obtain a global feature of the first user; in this embodiment , the second generating module, comprising: a second convolution processing unit, configured to respectively perform the first user of the second user by each convolution layer in the at least two convolution layers according to the corresponding convolution window performing convolution processing on features to obtain a plurality of intermediate local features of the second user; a second splicing unit configured to splice the multiple intermediate local features of the second user to obtain local features of the second user .

In some embodiments of the present application, the first neural network model includes a basic neural network model and an adaptive network inserted into the basic neural network model, and the parameters of the basic neural network model are based on the first neural network model of the source domain. Two user historical behavior data are determined by pre-training; the training device of the information processing model also includes: a parameter fixing module, which is used to fix the parameters of the basic neural network model; in this embodiment, the first training module is further configured It is: according to the first score and the second score, adjust the parameters of the adaptive network, the user feature model and the discriminant model and continue training until the training end condition is reached.

In some embodiments of the present application, the basic neural network model includes an embedding layer and multiple cascaded converter neural networks, the first neural network model is divided into N cascaded sub-neural networks, the The sub-neural network includes a converter neural network and at least one adaptive network inserted into the converter neural network; N is an integer greater than 1; the first output module includes: an input information acquisition unit for obtaining k-level sub-neural networks; The input information corresponding to the neural network; wherein, if k=1, the input information corresponding to the k-level sub-neural network is the embedding vector output by the embedding layer according to the interaction information sequence of the first user; if k>1, The input information corresponding to the k-level sub-neural network is that the k-1-level sub-neural network outputs the k-1 level intermediate user characteristics of the first user; 0<k≤N, k is an integer; the first input unit is used to The input information corresponding to the k-level sub-neural network is input into the k-level sub-neural network; the feature extraction unit is used to perform feature extraction by the k-level sub-neural network according to the corresponding input information, and output the first user's k wherein, if k=N, the k-level intermediate user features of the first user are used as the first user features of the first user; if k<N, the k-level intermediate user features of the first user are used as The input information of the k+1-level sub-neural network.

In some embodiments of the present application, the converter neural network includes a cascaded multi-head attention layer, a first feedforward neural network layer, a first summation and normalization layer, a second feedforward neural network layer, and a second summation and normalization layer; each of said subneural networks includes an adaptive network inserted between said first feedforward neural network layer and said first summation and normalization layer, and inserting An adaptive network between the second feed-forward neural network layer and the second summation and normalization layer; and an identity map of inputs to the sum and normalization layer, and an identity map from the input of the second feedforward neural network layer to the input of the second sum and normalization layer.

In some embodiments of the present application, the adaptive network includes a cascaded first fully connected layer, an intermediate fully connected layer, an activation layer and a final fully connected layer, and the number of neurons in the first fully connected layer Equal to the number of neurons in the fully connected layer of the last layer, the number of neurons in the fully connected layer of the first layer is greater than the number of neurons in the middle fully connected layer; the adaptive network also includes the An identity mapping from the input of the first fully connected layer to the output of the last fully connected layer.

In some embodiments of the present application, the information processing model training device further includes: a target node selection module, configured to select multiple target nodes in the reference interaction information sequence; the reference interaction information sequence refers to the second user The interaction information sequence of each user in the historical behavior data; the replacement module is used to block the multiple target nodes in the reference interaction information sequence to obtain a sample interaction information sequence; the sample interaction information sequence input module is used to convert The sample interaction information sequence is input into the basic neural network model; the prediction information output module is used for the basic neural network model to output the prediction information corresponding to the target node according to the sample interaction information sequence; the prediction error is determined A module for determining a prediction error according to the prediction information and the target node; a second training module for adjusting the parameters of the basic neural network model according to the prediction error and continuing training until the pre-training end condition is reached .

According to an aspect of the embodiment of the present application, an information processing device is provided, including: a candidate information set acquisition module, used to acquire a candidate information set in the target domain; a behavior label prediction module, used to predict The target user's behavior label for each information in the candidate information set; the information processing model is obtained by training according to the above-mentioned information processing model training method; the target information determination module is used to, according to the behavior label, in Target information to be sent to the target user is determined in the candidate information set.

In some embodiments of the present application, the information processing model includes a trained first neural network model, a trained user feature model, an embedding search layer, and a classification layer; a behavior label prediction module includes: global feature search A unit, configured to obtain the global features of the target user in the user global feature set; the global feature in the user global feature set is composed of the trained first neural network model and the trained user feature model Generated according to the first user behavior data; the embedding vector generation unit is used to generate the embedding vector of each information in the candidate information set by the embedding search layer; the behavior label output unit is used to be used by the classification layer according to The global features of the target user and the embedding vectors of each information in the candidate information set are used to output the target user's behavior label for each information in the candidate information set.

In some embodiments of the present application, the information processing device further includes: a receiving module, configured to receive an information request, the information request indicating the user identifier of the target user; a sending module, configured to identify, and send the target information to the target user.

In some embodiments of the present application, the information processing device further includes: an interactive behavior information acquisition module, configured to acquire the target user's interactive behavior information for the target information; an update training module, configured to The interactive behavior information of the target information is used to update and train the information processing model.

According to an aspect of an embodiment of the present application, there is provided an electronic device, including: a processor; a memory, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by the processor, the above The training method of the information processing model or the information processing method.

According to an aspect of the embodiments of the present application, there is provided a computer-readable storage medium on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor, the information processing model described above is implemented. training methods or information processing methods.

According to an aspect of the embodiments of the present application, a computer program product is provided, including computer instructions, and when the computer instructions are executed by a processor, the above information processing model training method or information processing method is implemented.

In the solution of this application, the first neural network model is pre-trained through the second user behavior data in the source domain, so that the first neural network model can learn the user's user characteristics according to the user's interaction information sequence in the source domain , on this basis, combined with the discriminant model, using the first user behavior data in the source domain to conduct unsupervised comparative learning on the first neural network model and the user feature model, to widen the difference between the features of different users, and improve the interaction-based The accuracy of user features extracted from information sequences. Therefore, after the information processing model determined according to the trained first neural network model and the user feature model, since the information processing model has accurately learned the user's features in the source domain, the user's features reflect the user's preference for information , therefore, the information processing model determines the target information to be sent to the user in the target domain according to the user characteristics learned in the source domain, because the determination process of the target information combines the user characteristics of the user in the source domain , therefore, the matching between the target information and the user can be improved, thereby effectively reducing the number of interactions with the user, and effectively solving the problem of low information sending efficiency caused by the low matching between the information and the user in the prior art The problem.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and together with the description serve to explain the principles of the application. Apparently, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative efforts.

Fig. 1 shows a schematic diagram of an information sending system provided by an exemplary embodiment of the present application.

Fig. 2 is a flowchart showing a method for training an information processing model according to an embodiment of the present application.

Fig. 3 is a schematic diagram illustrating unsupervised contrastive learning in the fine-tuning training phase according to an embodiment of the present application.

Fig. 4 is a schematic diagram of a basic neural network model according to an embodiment of the present application.

Fig. 5 is a schematic diagram of a sub-neural network according to an embodiment of the present application.

Fig. 6 is a schematic structural diagram of an adaptive network according to an embodiment of the present application.

Fig. 7 is a flowchart showing steps before step 220 according to an embodiment of the present application.

Fig. 8 is a flowchart of an information processing method according to an embodiment of the present application.

Fig. 9 is a flow chart showing step 820 according to an embodiment of the present application.

Fig. 10 is a schematic diagram showing an information processing model predicting behavior labels according to an embodiment of the present application.

Fig. 11 is a schematic diagram illustrating displaying target information in a user interface according to an embodiment of the present application.

Fig. 12 is a block diagram of an information processing model training device according to an embodiment.

Fig. 13 is a block diagram of an information processing device according to an embodiment of the present application.

Fig. 14 shows a schematic structural diagram of a computer system suitable for implementing the electronic device of the embodiment of the present application.

Detailed ways

Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments may, however, be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the concepts of example embodiments to those skilled in the art.

Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided in order to give a thorough understanding of the embodiments of the application. However, those skilled in the art will appreciate that the technical solutions of the present application may be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be employed. In other instances, well-known methods, apparatus, implementations, or operations have not been shown or described in detail to avoid obscuring aspects of the application.

The block diagrams shown in the drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor devices and/or microcontroller devices entity.

The flow charts shown in the drawings are only exemplary illustrations, and do not necessarily include all contents and operations/steps, nor must they be performed in the order described. For example, some operations/steps can be decomposed, and some operations/steps can be combined or partly combined, so the actual order of execution may be changed according to the actual situation.

It should be noted that: the "plurality" mentioned in this article refers to two or more than two. "And/or" describes the association relationship of associated objects, indicating that there may be three types of relationships. For example, A and/or B may indicate: A exists alone, A and B exist simultaneously, and B exists independently. The character "/" generally indicates that the contextual objects are an "or" relationship.

Before the application is described in detail, the terms involved in the application are explained as follows:

Source domain: In transfer learning, the source domain refers to the domain to which the existing knowledge belongs.

Target domain: Also known as the target domain, it refers to the domain where new knowledge is to be learned, which can also be understood as the domain that needs to deal with the current target task.

In an information sending scenario, the domain to which an application with sufficient user interaction data already exists may be referred to as a source domain. And the domain of the application with no user interaction data or less user interaction data is called the target domain. For example, in the first application program, if there are a large number of user interaction data for the information sent in the first application program, the domain to which the first application program belongs can be called the source domain; in the second application program, the lack of The user's interaction data, therefore, the domain to which the second application program belongs can be referred to as the target domain.

In order to solve the problem of low information sending efficiency due to the low matching between information and users in the prior art, the solution of this application is proposed, which builds and trains an information processing model based on artificial intelligence technology, and utilizes the user to target the source The user interaction information sequence of the information in the domain is used to train the information processing model and learn the user characteristics of each user in the source domain. On this basis, the user characteristics learned by the information processing model in the source domain are used to determine the target in a targeted manner. The information sent to the user in the target domain, so as to ensure the matching between the target information determined for the user in the target domain and the user, thereby improving the efficiency of information transmission.

In some application scenarios, such as the lack of user interaction data in new applications (this scenario can be called a cold start scenario, and the relevant users are called cold users), there is also a matching between the information sent to the user and the user Not high, resulting in the need to send information to the user multiple times. Therefore, the solution of the present application can also be applied to determine the target information to be sent to the user in a cold start scenario.

Artificial Intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the nature of intelligence and produce a new kind of intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

Artificial intelligence technology is a comprehensive subject that involves a wide range of fields, including both hardware-level technology and software-level technology. Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning/deep learning. The solution of the present application mainly utilizes natural language processing technology to mine user characteristics according to the user's information triggering user behavior in the source domain, so as to obtain the user characteristics of the user. And the user characteristics of the user in the source domain are applied to the target domain across domains to realize information recommendation across neighborhoods.

In some embodiments of the present application, the target domain may be a domain or product domain similar to the source domain. For example, if the information to be recommended in both the source domain and the target domain are news, the target domain and the source domain can be regarded as similar fields; Aggregating information such as articles, videos, etc., the target domain and source domain can be regarded as similar domains. In some embodiments, if the type of information in the target domain is the same as the type of information in the source domain, then the target domain can be regarded as a similar domain to the source domain. Types of information such as articles (blogs, news, public account articles, etc.), small videos, news, advertisements, product links, etc. Correspondingly, the information to be recommended in the source domain and the target domain may be articles, short videos, news, advertisements, and the like.

Fig. 1 shows a schematic diagram of an information sending system provided by an exemplary embodiment of the present application. The information sending system includes a terminal 110 and a sending platform 120 .

The sending platform 120 can be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, and can also provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication , middleware services, domain name services, security services, CDN (Content Delivery Network, content distribution network), and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms.

The sending platform 120 is used to provide background services for applications supporting information sending. Optionally, the sending platform 120 undertakes the main computing work, and the terminal 110 undertakes the secondary computing work; Use distributed computing architecture for collaborative computing.

Optionally, the sending platform 120 includes a first sending server 121 and a second sending server 122 . The first sending server 121 is used to provide services related to information sending in the source domain, such as determining information in the source domain to be sent to each user, collecting user interaction data on information sent in the source domain, and the like.

The second sending server 122 is configured to provide services related to information sending in the target domain, for example, determine the target information in the target domain to be sent to each user. In some embodiments, the second sending server 122 is also communicatively connected with the first sending server 121 . The information processing model is deployed on the second sending server 122, which can be based on the information in the source domain sent by the first sending server 121 to the user and the user's interaction data with respect to the information sent in the source domain as the training data of the information processing model. An information processing model training is performed, so that the trained information processing model is used to determine the target information in the target domain to be sent to the user.

The terminal 110 communicates with the sending platform 120 through a wireless network or a limited network, and the terminal 110 may be an electronic device such as a smart phone, a notebook computer, a game console, a desktop computer, a tablet computer, an e-book reader, or a smart speaker. The terminal 110 installs and runs an application program supporting information transmission. The application can be any one of a news application, an instant messaging application, a reading application, a social application for crowd aggregation based on topics or channels or circles, a shopping-based social application, a browser program, or a video program. A sort of. Schematically, the terminal 110 is a terminal used by a first user, and an application program running on the terminal 110 is registered with an account of the first user.

The application programs running on the terminal 110 that support information sending may be two types of application programs, one of which is the first application program corresponding to the source domain, and the other application program is the second application program corresponding to the target domain. On this basis, the first information in the source domain can be sent to the user through the first application program; the information in the target domain can be recommended to the user through the second application program. For example, the first application program can be an instant messaging application, and the information in the source domain can be articles, news, blogs, etc. from official accounts; the second application program can be a reading application, and the information to be sent in the target domain can be Blogs, book reviews, news, etc. in the reading app.

When the terminal 110 can run the first application program, when the terminal 110 is running the first application program, the first sending server 121 can send the information in the source domain to the user where the first application program is located, and then the first application program The program detects the interaction behavior triggered by the user on the sent information in the source domain, and reports the interaction data of the user on the sent information in the source domain to the first sending server 121, so that the reported user's interaction with the sent information in the source domain The interaction data is used as the training data of the information processing model.

The first sending server 121 can send the interaction data reported by many terminals 110 running the first application program to the second sending server 122 about the recommended information in the source domain, and the second sending server 122 can follow the requirements of this application. The method trains an information processing model, uses the information processing model to specifically determine the target information in the target domain for users in the second application program, and sends the determined target information in the target domain to the users in the second application program.

When the second application program can also be run on the terminal 110, the second sending server 122 can determine the target information to be sent to the user in the target domain for the user where the second application program is located according to the information processing method of the present application, and The target information in the target field is sent to the terminal 110 .

It can be understood that, only the first application program or the second application program may be run on the terminal 110, or both the first application program and the second application program may be run. As shown in FIG. 1 , the number of terminals in the information sending system may be more, and the information sending system may also include other terminals.

The implementation details of the technical solutions of the embodiments of the present application are described in detail below:

FIG. 2 shows a flowchart of a method for training an information processing model according to an embodiment of the present application. The method can be executed by a computer device with processing capabilities, such as a server, and is not specifically limited here. Referring to Figure 2, the method at least includes steps 210 to 280, which are described in detail as follows:

Step 210, acquiring the interaction information sequence of the first user and the interaction information sequence of the second user from the behavior data of the first user in the source domain.

An information set is provided in the source domain, and based on the information set, the information in the source domain can be sent to the user in a targeted manner. After the information in the information collection is sent to the user, the interaction data of the user on the sent information in the source domain is collected. The user's interaction data on the recommended information may be reflected by the user's interaction behavior indicating interest in the recommended information triggered by the user. For ease of description, an interactive behavior indicating that a user is interested in information is referred to as an interested user behavior. Behaviors of interested users, such as click behaviors, collection behaviors, like behaviors, and comment behaviors triggered by recommended information. In some embodiments, if the user clicks to enter the details page of the recommended information, the user can trigger behaviors such as liking, bookmarking, or commenting on the recommended information. interaction data.

For a large number of users, interaction data of multiple users with respect to the sent information in the source domain is collected. According to the interaction data of many users for the sent information, in chronological order, arrange the multiple sent information that triggers the interaction behavior of the same user, and obtain the interaction information sequence of each user. In other words, the user's interaction information sequence refers to the information sequence obtained by arranging the sent information of the user that triggers the user's behavior of interest in chronological order. Wherein, the chronological order may be determined based on the time when the user triggers an interested user behavior on the sent information. It can be understood that, in the user interaction information sequence, the sent information may be represented by an information identifier of the sent information, such as an information ID. For example, if the information sent to the user in the source domain includes A, B, C, D, E, F; among them, the information clicked by the user in turn includes A, C, E, F, then the user's interactive information sequence can be <A, C, E, F>.

Further, in order to indicate the field to which the interaction information sequence belongs and the user corresponding to the interaction information sequence, the user identifier of the user, the field identifier of the field to which the interaction information sequence belongs, and the interaction information sequence may form a triplet. Therefore, the first user behavior data in the source domain is a collection of multiple triplets.

The first user behavior data in the source domain includes the interaction information sequences of multiple users in the source domain. In this scheme, since the information processing model is fine-tuned and trained through unsupervised comparative learning, each time from The interaction information sequences of two different users in the source domain are obtained from the first user behavior data in the source domain. In order to distinguish the two users, one of the users is called the first user, and the other user is called the second user.

In step 220, the first neural network model outputs the first user characteristics of the first user according to the sequence of interaction information of the first user, and in step 230, the first neural network model outputs the characteristics of the second user according to the sequence of interaction information of the second user. The first user characteristics; the first neural network model is determined according to the pre-training of the second user historical behavior data in the source domain.

The solution of this application is to apply the knowledge learned in the source domain to the information sending task in the target domain through transfer learning. Therefore, the information processing model is trained with the help of user interaction information in the source domain, so that the information processing model Can be used to determine the target information to be sent in the target field.

In the scheme of this application, the information processing model is trained through two training stages, which can be called the pre-training (Pre-training) stage and the fine-tuning (Fine-tuning) training stage, and the pre-training stage is carried out first training, followed by training in the fine-tuning training phase. Therefore, the training data in the source domain is divided into two parts, one of which is used in the pre-training phase and the other is used in the fine-tuning training phase. In the solution of the present application, the training data used in the pre-training phase is called second user behavior data, and the training data used in the fine-tuning training phase is called first user behavior data.

It is understandable that, even if the same information is sent to different users, different users may have different interactive behaviors triggered by the same information due to differences in the information that different users are interested in or preferred information. In other words, the user's interactive behavior triggered by the information sent in the source domain reflects the user's preference characteristics in information to a certain extent. Based on this, in the solution of this application, in the pre-training stage, the first neural network model is trained through the second user behavior data in the source domain, so that the first neural network model is capable of outputting the user interaction information sequence according to the user The feature representation of the user, for the convenience of description, the feature representation of the user output by the first neural network model is called the first user feature.

The first neural network model may be a model composed of a neural network. Neural networks such as CNN (Convolutional Neural Network, convolutional neural network), RNN (Recurrent Neural Network, cyclic neural network), LSTM (Long Short-Term Memory Neural Network, long-short-term memory neural network), fully connected neural network, Transformer model (transformer neural network), etc.

After pre-training the first neural network model, since the first neural network model has the ability to output the user's feature representation according to the user's interaction information sequence, the first neural network model can be based on the first user in the source domain. The interaction information sequence of correspondingly outputs the first user characteristics of the first user, and can output the first user characteristics of the second user according to the interaction information sequence of the second user in the source domain.

Step 240, generate the local features of the first user and the global features of the first user according to the first user features of the first user by the user feature model; and step 250, generate the first user feature of the second user by the user feature model Second, the user's local features.

The user feature model is used for deep feature extraction based on the first user feature of the user. The user feature model may be a neural network model constructed by a neural network such as a convolutional neural network, a pooled neural network, or a fully connected neural network.

The features output by the user feature model that reflect the user's local area are called local features, and the features output by the user feature model that reflect the user's overall situation are called global features. Wherein, the global feature of the user is obtained through further processing based on the local feature of the user.

In some embodiments of the present application, the user feature model includes at least two convolutional layers and a pooling layer, and the length of the convolution window corresponding to each convolutional layer in the at least two convolutional layers is different; in this embodiment Among them, the user feature model can output the local features and global features of the first user through the following process: Each convolution layer in the at least two convolutional layers is respectively analyzed according to the corresponding convolution window A user feature is subjected to convolution processing to obtain multiple intermediate local features of the first user; multiple intermediate local features of the first user are spliced to obtain the local features of the first user; the local features of the first user are obtained by the pooling layer The features are pooled to obtain the global features of the first user.

Similarly, the user feature model can output the local features of the second user through the following process: Each convolution layer in the at least two convolutional layers is respectively analyzed by the first user of the second user according to the corresponding convolution window The features are convoluted to obtain multiple intermediate local features of the second user; the multiple intermediate local features of the second user are spliced to obtain the local features of the second user.

For the convolution kernel, the length of the convolution window is different, and the receptive field of view of the convolution kernel is also correspondingly different. Therefore, the convolutional layers with different convolution windows are used to convolve the first user features of the first user respectively. A user feature is extracted. Each convolutional layer outputs an intermediate local feature of the user.

In this embodiment, after obtaining the multiple intermediate local features of the user (the first user and the second user), the multiple intermediate local features are concatenated, and the concatenated features are used as the user's local features. On this basis, the local features obtained by the user's concatenation are input into the pooling layer, and the pooling layer performs pooling processing on the user's local features and outputs the user's global features. Wherein, the pooling processing performed may be average pooling processing, maximum pooling processing, and the like.

In step 260, the discriminant model calculates a first score according to the first user's local features and the first user's global features; and the discriminant model calculates a second score based on the first user's global features and the second user's local features.

The discriminant model may be a model constructed by a neural network such as a convolutional neural network or a fully connected network. The first score is calculated by the discriminant model according to the local features of the first user and the global features of the first user, the second score is calculated by the discriminant model according to the global features of the first user and the local features of the second user, and based on the The calculated score is used to reversely adjust the parameters of the model (the first neural network model, the user feature model, and the discriminant model), thereby narrowing the gap between the local features and the global features of the same user, and widening the global features of the first user. The gap between the features and the local features of the second user maximizes the amount of information between the user's global features and local features, and widens the difference between the features of different users.

It can be seen from the above that in the solution of this application, the user feature model is not only used to extract features from the first user features, but also used to construct negative samples for unsupervised contrastive learning. Specifically, in this solution, the global features of the first user and the local features of the first user are used as positive samples, and the local features of the second user are used as negative samples. Since the negative samples of comparative learning are automatically constructed through the user feature model, there is no need to label the data during the fine-tuning training stage, which greatly reduces the workload of model training. The situation where the training effect of the model is affected by incorrect labeling.

In some embodiments, the first score is used to reflect the similarity between the first user's local features and the first user's global features, and the second score is used to reflect the first user's global features and the second user's local features similarity between. Therefore, in the training process, it is expected that the higher the first score, the better, and the lower the second score, the better, so as to widen the difference between the characteristics of different users.

Step 270, train the first neural network model, the user feature model and the discriminant model according to the first score and the second score until the training end condition is met.

In some embodiments of the present application, loss functions can be set for the first neural network model, user feature model, and discriminant model, and based on the loss function of the first neural network model, the loss function of the user feature model, and the discriminant model loss function sets the target loss function. On this basis, the loss functions of the first neural network model, the user feature model, and the discriminant model are calculated respectively according to the first score and the second score, and then the function value of the target loss function is determined. According to the calculated function of the target loss function value to adjust the parameters of at least one model in the first neural network model, user feature model and discriminant model, and perform the process of the above steps 220-260 again by adjusting the parameters of the first neural network model, user feature model and discriminant model , until the calculated function value of the target loss function makes the target function converge, and then obtain the interaction information sequence of the other two users from the first user behavior data to repeat the process of the above steps 210-270 until the training end condition is reached.

In some embodiments, in the process of adjusting the parameters of the first neural network model, the user feature model, and the discriminant model according to the function value of the target loss function, the parameters of each model can be adjusted in stages, and only the parameters of one model can be adjusted in one stage , and in the next stage, adjust the parameters of other models. For example, if in a certain stage, the parameters of the fixed user feature model and the parameters of the fixed discriminant model are adjusted, only the parameters of the first neural network model are adjusted; in the next stage, the parameters of the fixed first neural network model and the fixed discriminant model are adjusted. Parameters, adjust the parameters of the user feature model; in the next stage, fix the parameters of the first neural network model and the parameters of the fixed user feature model, adjust the parameters of the discriminant model; thus, in the fine-tuning training stage, alternate according to the above process Adjust the parameters of each model. Of course, in other embodiments, it can also be set that only the parameters of the two models are adjusted at one stage, which is not specifically limited here. Since only one or two model parameters are adjusted in a parameter adjustment process, the parameter adjustment amount is reduced, and the training speed and training efficiency of the model are improved.

In some embodiments, in order to maximize the amount of information between the user's global features and local features, and to separate the differences between the features of different users, the loss function set for the discriminant model can be:

Loss1 = log(1+exp ^-t1 )+log(1+exp ^t2 );

Wherein, Loss1 is the function value of the loss function of the discriminant model, t1 denotes the first score calculated by the discriminant model, and t2 denotes the second score calculated by the discriminant model.

Step 280: Determine an information processing model according to the trained first neural network model and the trained user characteristic model, and the information processing model determines target information to be sent to the user in the target domain.

In this solution, the discriminant model may be a model for auxiliary fine-tuning training, therefore, the information processing model may be determined by combining the trained first neural network model and the trained user feature model. In some embodiments, due to the difference between the task in the fine-tuning training phase and the recommendation task in the target domain, that is, the task in the fine-tuning training phase is to output user features according to the user’s interaction information sequence, while the recommendation task in the target domain is to predict user For the behavior label of the information in the target domain, the behavior label user indicates the predicted user behavior triggered by the user for the information. Therefore, on the basis of the trained first neural network model and the trained user feature model, A classification layer is added, and the predicted user behavior labels for information in the target domain are output through the classification layer.

The target user generally refers to a user in the target domain to whom information is to be sent. In some embodiments, the target user may be a new user in the target domain, or a user whose interaction data in the target domain is less than a preset threshold.

In the target domain, due to the lack of user interaction data for information in the target domain, or the lack of user interaction data for information in the target domain, in this case, due to the first user behavior data and The information processing model trained by the second user behavior data has accurately learned the characteristics of the user. Therefore, the user characteristics learned by the information processing model can be transferred to the information sending task of the target domain, and the information processing model can be applied in The user characteristics learned in the source domain are used to predict the user's behavior label for the information in the target domain, and then the target information to be sent to the user is determined according to the behavior label.

In the solution of this application, the first neural network model is pre-trained through the second user behavior data in the source domain, so that the first neural network model learns the user's user characteristics according to the user's interaction information sequence in the source domain , on this basis, combined with the discriminant model, using the first user behavior data in the source domain to conduct unsupervised comparative learning on the first neural network model and the user feature model, to widen the difference between the features of different users, and improve the interaction-based The accuracy of user features extracted from information sequences. Therefore, after the information processing model determined according to the trained first neural network model and the user feature model, since the information processing model has accurately learned the user's features in the source domain, the user's features reflect the user's preference for information , therefore, the information processing model determines the target information to be sent to the user in the target domain according to the user characteristics learned in the source domain, since the determination of the target information combines the user characteristics of the user in the source domain, Therefore, the matching between the target information and the user can be improved, thereby effectively reducing the number of interactions with the user, and effectively solving the problem of low information sending efficiency caused by the low matching between the information and the user in the prior art. question.

Moreover, the solution of the present application can also be applied to determine the target information to be sent for the user in the cold start scenario, which can effectively improve the matching between the target information determined for the user and the user in the cold start scenario of the target domain, and There is no need to send information to users multiple times, which improves the efficiency of information sending. For users, the efficiency of user acquisition is improved, and the user experience is improved.

Fig. 3 is a schematic diagram showing unsupervised contrastive learning in fine-tuning training according to an embodiment of the present application. As shown in FIG. 3 , the first user feature 341 of the first user and the first user feature 342 of the second user are respectively output by the pre-trained first neural network model 310 . Then, local features and global features are generated for the first user and local features are generated for the second user by the user feature model 320 . Specifically, as shown in Figure 3, the convolutional layers of different window lengths in the user feature model perform N-gram local feature extraction on the first user feature 341 of the first user to obtain multiple intermediate local features of the first user features 351, and then concatenate multiple intermediate local features 351 of the first user to obtain the local features 361 of the first user, and then perform average pooling processing on the local features 361 of the first user to obtain the global Feature 362. Following a similar process, the user feature model 320 outputs the intermediate local features 352 of the second user, the local features 363 of the second user.

Afterwards, the discriminant model 330 calculates the first score according to the local features 361 of the first user and the global features 362 of the first user, and the discriminant model 330 calculates the second score based on the local features 363 of the second user and the global features 362 of the first user. Score. Then, the parameters of the first neural network model 310, the user feature model 320, and the discrimination model 330 are reversely adjusted according to the calculated first score and second score.

In some embodiments of the present application, the first neural network model includes a basic neural network model and an adaptive network inserted into the basic neural network model, and the parameters of the basic neural network model are determined by pre-training according to the second training data; step 220 Before, the method also includes: fixing the parameters of the basic neural network model; in this embodiment, step 270 includes: adjusting the parameters of the adaptive network, the user feature model and the discriminant model according to the first score and the second score and continuing Train until the training end condition is reached.

In this scheme, during the pre-training process, the parameters of the basic neural network model are adjusted; after pre-training, the parameters of the basic neural network model are fixed, and an adaptive network is inserted into the basic neural network model, so that in the fine-tuning training stage, for The first neural network model only adjusts the parameters of the adaptive network while keeping the parameters of the basic neural network model unchanged. It can retain the learning results of the pre-training stage in the fine-tuning training stage, and at the same time greatly reduces the amount of parameter adjustment, effectively improving the Model training speed and training effect.

In some embodiments of the present application, the basic neural network model includes an embedding layer and multiple cascaded converter neural networks. FIG. 4 is a schematic diagram of a basic neural network model according to an embodiment of the present application. The embedding layer of the basic neural network model 400 is not shown in FIG. 4 , and only multiple cascades in the basic neural network model are shown. Transformer Neural Networks 410 . The embedding layer is used to convert the information to obtain the embedding vector of each information. For example, if the input is a user's interaction information sequence in the source domain, the embedding layer can generate a corresponding embedding vector for the information ID corresponding to each information in the user interaction information sequence, as shown in Figure 4, for each information The embedding vector corresponding to the ID, according to the position of the information ID in the user interaction information sequence, encode the position of the information ID, and then superimpose the position encoding corresponding to the information ID and the embedding vector corresponding to the information ID as the information ID The embedding vector of . Then, the embedding vectors of all information IDs in the user interaction information sequence are input into the primary converter neural network 410 for feature extraction, and the output features are input into the next-level converter neural network 410 .

The converter neural network is also called the Transformer model. The Transformer model is a model structure used to encode words and sentences in the research field of NLP (Natural Language Processing, Natural Language Processing). It is a model based on the attention mechanism and can solve long-distance dependencies of sequences. The problem is that there is no serious gradient dispersion and fixed structural limitations like the cyclic neural network, and there is no need to consider the range of the receptive field like the convolutional neural network. In general, the Transformer model not only has good implementation effects, but also can be parallelized.

In some embodiments of the present application, the adaptive network may be inserted into the converter network, of course, the adaptive network may be inserted into each level of the converter neural network, or it may be in cascaded multiple Adaptive networks are inserted at intervals in the transformer neural network. In other embodiments, the adaptive network can also be inserted between adjacent two-stage converter networks.

In some embodiments of the present application, the first neural network model is divided into N cascaded sub-neural networks, and the sub-neural networks include a converter neural network and at least one adaptive network inserted into the converter neural network; N is an integer greater than 1; in this embodiment, step 220 includes: obtaining the input information corresponding to the k-level sub-neural network; wherein, if k=1, the input information corresponding to the k-level sub-neural network is the embedding layer according to the first The embedding vector output by the user's interactive information sequence; if k>1, the input information corresponding to the k-level sub-neural network is the k-1-level intermediate user feature of the first user output by the k-1-level sub-neural network; 0<k≤N , k is an integer; the input information corresponding to the k-level sub-neural network is input into the k-level sub-neural network; the k-level sub-neural network performs feature extraction according to the corresponding input information, and outputs the k-level intermediate user features of the first user; wherein , if k=N, use the k-level intermediate user feature of the first user as the first user feature of the first user; if k<N, use the k-level intermediate user feature of the first user as the k+1-level sub-neural network Enter information.

The multi-level sub-neural network is used to perform feature extraction level by level according to the interaction information sequence of the user in the source domain, so as to ensure the accuracy of the extracted first user feature. It can be understood that the process of extracting the first user features of the second user is similar to the process of extracting the first user features of the first user, and the process of extracting the first user features of the second user will not be repeated here.

FIG. 5 is a schematic diagram of a sub-neural network according to an embodiment of the present application. As shown in FIG. 5 , the sub-neural network 500 includes a converter neural network and two adaptive networks inserted into the converter neural network. Specifically, as shown in Figure 5, the converter neural network includes a cascaded multi-head attention layer 510, a first feedforward neural network layer 520, a first summation and normalization layer 530, a second feedforward neural network layer 540, and a second summation and normalization layer 550; each sub-neural network 500 includes an adaptive network 560 inserted between the first feedforward neural network layer 520 and the first summation and normalization layer 530, and an adaptive network 560 inserted between the second feed-forward neural network layer 540 and the second summation and normalization layer 550; The identity map of the input of the normalization layer 530 , and the identity map from the input of the second feedforward neural network layer 540 to the input of the second summation and normalization layer 550 .

Generally speaking, the deeper the network layer, the more features of different levels can be extracted, and the more combinations of different levels of features will be. However, as the number of network layers continues to deepen, problems such as gradient disappearance and gradient explosion are extremely prone to occur, which will lead to degradation of network performance. In order to solve this problem, a residual connection is introduced in this scheme. For example, in Figure 5, the identity mapping from the input of the multi-head attention layer to the input of the first summation and normalization layer, and from the second feedforward neural The input of the network layer points to the identity mapping of the input of the second summation and normalization layer, so that the information of the shallow layer is applied to the calculation of the deep layer, and the problems of gradient disappearance and gradient explosion are avoided.

FIG. 6 is a schematic structural diagram of an adaptive network shown according to an embodiment of the present application. As shown in FIG. 6 , an adaptive network (Adapter) 560 includes a cascaded first fully connected layer 610, an intermediate fully connected layer 620, and an activation layer. 630 and the last fully connected layer 640, the number of neurons in the first fully connected layer 610 is equal to the number of neurons in the last fully connected layer 640, and the number of neurons in the first fully connected layer 610 is greater than that of the middle fully connected layer The number of neurons in 620; the adaptive network 560 also includes an identity mapping from the input of the first fully connected layer 610 to the output of the last fully connected layer 640.

It is worth mentioning that in this embodiment, according to the transmission direction of data in the adaptive network, each fully connected layer in the adaptive network (Adapter) is divided into the first fully connected layer, the middle fully connected layer and the last layer fully connected layer.

For a fully connected layer, the number of neurons in the fully connected layer determines the dimension of the output vector of the fully connected layer, that is, if the number of neurons in the fully connected layer is M, then the output vector of the fully connected layer It is M dimension. Therefore, in this scheme, the number of neurons in the first fully connected layer is the same as the number of neurons in the last fully connected layer, while the number of neurons in the middle fully connected layer is smaller than the number of neurons in the first fully connected layer. The number, so that the adaptive network is a bottleneck-style MLP (multilayer perceptron, multi-layer perceptron) structure.

An activation function is configured in the activation layer 630 , and the output of the previous layer is nonlinearly transformed through the activation function. The activation function may be a ReLU function, a nonlinear activation function based on Gaussian Error Linear Units (Gaussian Error Linear Units, GELU), a hyperbolic tangent function (Tanh function), etc., which are not specifically limited here. Since the fully connected layer only performs linear transformation (affine transformation) on the data, and the superposition of multiple linear transformations is still a linear transformation, the nonlinearity of the adaptive network is increased by the activation layer. It is worth mentioning that FIG. 6 only exemplarily shows one intermediate fully-connected layer 620 in the adaptive network, and in other embodiments, the adaptive network may include multiple intermediate fully-connected layers.

Similarly, in the adaptive network, the identity mapping from the input of the first fully connected layer 610 to the output of the last fully connected layer 640 is used to avoid gradient disappearance and gradient explosion due to the increase in the number of network layers in the adaptive network and other problems, the input of the shallow first fully connected layer is introduced into the deep last fully connected layer for calculation.

In some embodiments of the present application, before step 220, as shown in FIG. 7 , the method further includes:

Step 710, selecting a plurality of target nodes in the reference interaction information sequence; the reference interaction information sequence refers to the interaction information sequence of each user in the second user historical behavior data.

As described above, the mutual information sequence is formed by sequentially sorting a plurality of information, and an information in the mutual information sequence can be regarded as a node. Therefore, the reference mutual information sequence includes multiple nodes, and then the reference mutual information Select one or more nodes in the sequence as the target node.

Of course, in other embodiments, it is also possible to preset a specified number or a specified ratio, and select target nodes from the reference interaction information sequence according to the specified number or specified ratio. Wherein, the specified number is used to indicate the number of target nodes, and the specified ratio indicates the proportion of the number of target nodes in the reference interaction information sequence. In a specific embodiment, a node may be randomly selected as the target node.

Step 720, occlude multiple target nodes in the reference interaction information sequence to obtain a sample interaction information sequence.

As described above, each node in the reference interaction information sequence indicates the information of the interaction behavior triggered by the corresponding user, such as the information of the click behavior triggered by the user. In this embodiment, the target node is covered, so that the basic neural network model can learn the information indicated by the blocked target node. The sample mutual information sequence refers to the information sequence after occluding the target node in the reference mutual information sequence.

Step 730, input the sample interaction information sequence into the basic neural network model.

In step 740, the basic neural network model outputs prediction information corresponding to the target node according to the sample interaction information sequence.

As described above, the basic neural network model needs to output the information indicated by the occluded target node in the sample interaction information sequence according to the sample interaction information sequence. Specifically, the basic neural network model will make predictions according to the information indicated by the adjacent nodes before and after the occluded node in the sample interaction information sequence. In other words, the basic neural network model learns the context of the nodes in the sample interaction information sequence. The unoccluded nodes predict the occluded information of the target nodes, and obtain the predicted information corresponding to each target node. The prediction information corresponding to each target node is used to indicate the information indicated by the occluded target node predicted by the basic neural network model.

Step 750: Determine the prediction error according to the prediction information and the target node.

In the reference mutual information sequence, the information indicated by the target node is clear. Therefore, the prediction error can be determined according to the information indicated by the target node in the reference mutual information sequence and the prediction information corresponding to the target node. In some embodiments, the loss function of the basic neural network model can be preset, and the loss function can be a cross-entropy loss function, a mean square error loss function, an exponential loss function, etc., which are not specifically limited here.

In some embodiments, the output of the basic neural network model for the target node is the prediction information embedding vector. On this basis, it can correspond to the embedding vector obtained according to the information indicated by the target node in the reference interaction information sequence. Therefore, it can be obtained according to The function value of the loss function is calculated corresponding to the embedding vector of the prediction information of the target node and the embedding vector of the information indicated by the target node in the reference mutual information sequence, and the calculated function value of the loss function is the prediction error.

Step 760, adjust the parameters of the basic neural network model according to the prediction error and continue the training until the end condition of the pre-training is reached.

According to the calculated prediction error, reversely adjust the parameters of the basic neural network model, and then repeat the process of 730-750 through the adjusted basic neural network model, so that the recalculated prediction error makes the basic neural network model The loss function converges, and then use the next reference interaction information sequence to repeat the training of the basic neural network model according to the above steps 710-750 until the pre-training end condition is reached.

In the embodiment, a certain proportion of target nodes in the reference interaction information sequence is occluded to obtain a sample interaction information sequence, and the basic neural network model is used to predict the target node according to the context node of the blocked node in the sample interaction information sequence. With reference to the information indicated in the interaction information sequence, self-supervised training of the underlying neural network model is achieved.

FIG. 8 is a flow chart of an information processing method according to an embodiment of the present application. As shown in FIG. 8 , the method includes: Step 810, acquiring a candidate information set in the target domain. Step 820: Predict the target user's behavior label for each information in the candidate information set through the information processing model; the information processing model is trained according to any embodiment of the above information processing model training method.

In some embodiments, the target user may be a new user in the target domain, and the target user is also a user in the source domain, since the information processing model has learned the user characteristics of the users in the source domain during the training process, therefore , the user characteristics of the user in the source domain can be directly applied to the target domain as the user characteristics of the target user in the target domain. On this basis, the information processing model uses the user characteristics of the target user in the source domain to predict the target user's behavior labels for each information in the candidate information set.

The behavior tag is used to indicate the user behavior triggered by the target user for the corresponding information. In some embodiments, the behavior tag is used to indicate whether the user will click on the corresponding information. Correspondingly, the behavior tag includes Labels and labels that indicate that no click behavior will be triggered.

Step 830, according to the behavior tag, determine the target information to be sent to the target user in the candidate information set.

According to the predicted behavior tags corresponding to each information in the candidate information set, the information that the behavior tag indicates that the target user will click can be filtered out from the candidate information set, and the filtered information can be used as the target information to be sent to the target user .

In this solution, the information processing model is trained by using the first user behavior data and the second user behavior data in the source domain, so as to ensure that the information processing model can accurately learn the characteristics of users in the source domain. On this basis, the user-specific features learned by the information processing model in the source domain are applied to the target domain, and the user's features learned in the source domain reflect the user's preference for information. The characteristics in the target domain are used to determine the target information to be sent in the target domain, thereby ensuring the matching between the target information determined for the user and the target user. Therefore, it is possible to avoid multiple times of information screening and information sending, which is effective It solves the problem of low information sending efficiency in the prior art due to low matching between information and users. Moreover, the solution of the present application can also be applied to determine target information to the user in a cold start scenario of the target domain.

In some embodiments of the present application, the information processing model includes a trained first neural network model, a trained user feature model, an embedded search layer and a classification layer; as shown in FIG. 9, step 820 includes:

Step 910, acquire the global features of the target user in the user global feature set; the global features in the user global feature set are generated by the trained first neural network model and the trained user feature model according to the first user behavior data.

As described above, through the trained first neural network model and the trained user feature model, the global features of each user in the source domain can be generated according to the first user behavior data in the source domain, and the global features of each user in the source domain The global features in are added to the user global feature set. After applying the information processing model to the target domain, when it is necessary to recommend information to a target user, if the target user is both a new user in the target domain and a user in the source domain, you can use the user global feature Find the global features of the target user in the collection.

In some embodiments of the present application, when the source domain and the target domain are divided according to applications, it is assumed that the source domain corresponds to the first application program, and the target domain corresponds to the second application program, wherein the first application program and the second application program The two applications can be applications with login accounts open to each other. For example, the login account in the first application can authorize login in the second application. In this case, when the account authorized to log in in the second application is In the case of a login account in the first application program, the user can be regarded as both a user in the source domain and a user in the target domain. Therefore, the user's login account is used as the user identifier, and the user's global feature is searched in the user's global feature set to obtain the user's global feature.

In step 920, an embedding vector of each information in the candidate information set is generated by the embedding search layer. In some embodiments, the embedding lookup layer can use the embedding matrix corresponding to the embedding layer in the first neural network model to generate an embedding vector of each information in the candidate information set.

In step 930, the classification layer outputs the target user's behavior label for each information in the candidate information set according to the global features of the target user and the embedding vectors of each information in the candidate information set.

In some embodiments of the present application, the classification layer may use a softmax function to output the target user's behavior labels for each information in the candidate information set according to the target user's global features and the embedding vectors of each information in the candidate information set.

Fig. 10 is a schematic diagram showing an information processing model predicting behavior labels according to an embodiment of the present application. As shown in Figure 10, the first neural network model 310 outputs the first user features of each user according to the interaction information sequence of each user in the source domain, and then the user feature model 320 performs deep feature extraction from the first user features of each user, Get the global features of each user on the source domain, and add the global features of each user on the source domain to the user global feature set.

On this basis, when information in the target domain needs to be determined from the target user, the global features of the target user are obtained from the user global feature set according to the user identifier of the target user. The embedded search layer 1010 in the information processing model generates the embedding vector of the information according to the target domain ID corresponding to each information in the target domain (the target domain ID is used to identify the information in the target domain), and finally the classification in the information processing model Layer 1020 performs classification according to the global features of the target user and the embedding vectors corresponding to the information in the target domain, and outputs the behavior labels of the target user for the information in the target domain.

In some embodiments of the present application, before step 820, the method further includes: receiving an information request, and the information request indicates the user identity of the target user; in this embodiment, after step 820, the information processing method further includes: according to User ID of the target user, to send target information to the target user.

Fig. 11 is a schematic diagram illustrating displaying target information in a user interface according to an embodiment of the present application. In the embodiment corresponding to FIG. 11 , the target domain is the field to which the reading application belongs, and the subpage of "Kankan" in the reading application displays the information sent by the reading application server to the user. Specifically, the The information sent is an article. Concrete, can display the article title of sent article (article title I, article title II and article title III among Fig. 11), article author title (as in Fig. 11) in the sub-page of this " take a look " author I, author II and author III), further, the cover image of the sent article can also be displayed.

In a specific embodiment, the information processing model of this scheme is rigorously tested on the first data set ColdRec-2 and the second data set. In the experiment, the training method of the present application is used to obtain the data shown in Figure 10. The presented information processing model is compared with other recommendation models. Among them, the recommended models for comparison and reference include NeuFM, DeepFM (recommended model proposed by Huawei Noah's Ark Laboratory in 2017), MTL, PeterRec (Parameter-Efficient Transfer from Sequential Behaviors for User Modeling and Recommendation, used for user modeling and recommended sequential behavior for efficient transfer of parameters). In the experiment, HR@5 is used as the parameter index to test the performance of each recommendation model. Among them, the full name of HR@5 is Hit Ratio Top5, which refers to the accuracy rate of the first 5 items recalled. The experimental results are shown in Table 1 below.

Table 1

In Table 1 above, the ETM model is trained according to the training method of the present application to obtain the information processing model shown in FIG. 10 . It can be seen from Table 1 that, compared with other models, the information processing model trained by this application has achieved an accuracy rate improvement of more than 15% on both the first data set ColdRec-2 and the second data set. It shows that the information processing model obtained according to the training method of the present application can effectively improve the accuracy of information determination in the target domain.

In some embodiments of the present application, after the step of pushing the target information to the target user according to the user identification of the target user, the information recommendation method further includes: acquiring the target user's interaction behavior information for the target information; The interactive behavior information of the information processing model is updated and trained.

The target user's interaction behavior information on the target information is used to indicate the target user's interaction behavior triggered by the target information, such as clicking behavior. The server can send information in the target domain to multiple target users. When the collected data volume of the target user’s interactive behavior information on the target information is sufficient, the target user’s interactive behavior information on the target information can be used to The information processing model is updated and trained, so that the information processing model can generate user features of each user in the target domain. Furthermore, in the subsequent process, the information processing model can send targeted information directly according to the global characteristics of the user in the target domain, without using the global characteristics of the user in the source domain to process information in the target domain. sent.

The following introduces device embodiments of the present application, which may be used to implement the methods in the foregoing embodiments of the present application. For details not disclosed in the device embodiments of the present application, please refer to the foregoing method embodiments of the present application.

FIG. 12 is a block diagram of an information processing model training device according to an embodiment. As shown in FIG. Obtain the interaction information sequence of the first user and the interaction information sequence of the second user from the user behavior data; the first output module 1220 is configured to output the first user's first User features, and a second output module 1230, configured to use the first neural network model to output the first user features of the second user according to the interaction information sequence of the second user; the first neural network model is based on the second user history of the source domain The behavior data is determined by pre-training; the first generation module 1240 is used to generate the local characteristics of the first user and the global characteristics of the first user by the user characteristic model according to the first user characteristics of the first user; and the second generation module 1250 , used to generate the local features of the second user according to the first user features of the second user by the user feature model; the first score calculation module 1260 is used to use the discriminant model to generate the local features of the first user and the global features of the first user Calculate the first score; and the second score calculation module 1270 is used to calculate the second score according to the global features of the first user and the local features of the second user by the discriminant model; the first training module 1280 is used to calculate the second score according to the first score and the local features of the second user. The second score trains the first neural network model, the user feature model and the discriminant model until the training end condition is reached; the information processing model determination module 1290 is used for according to the trained first neural network model and the trained user feature model An information processing model is determined, and the information processing model determines the target information to be sent to the user in the target domain.

In some embodiments of the present application, the user feature model includes at least two convolutional layers and a pooling layer, and the length of the convolution window corresponding to each convolutional layer in the at least two convolutional layers is different; in this embodiment Among them, the first generation module 1240 includes: a first convolution processing unit, which is used to perform the first user feature of the first user by each convolution layer in the at least two convolution layers according to the corresponding convolution window Convolution processing to obtain multiple intermediate local features of the first user; a first splicing unit for splicing multiple intermediate local features of the first user to obtain local features of the first user; a first pooling processing unit, It is used to perform pooling processing on the local features of the first user by the pooling layer to obtain the global features of the first user; in this embodiment, the second generation module 1250 includes: a second convolution processing unit for Each of the at least two convolutional layers performs convolution processing on the first user features of the second user according to the corresponding convolution window to obtain multiple intermediate local features of the second user; the second splicing unit, It is used to concatenate multiple intermediate local features of the second user to obtain the local features of the second user.

In some embodiments of the present application, the first neural network model includes a basic neural network model and an adaptive network inserted into the basic neural network model, and the parameters of the basic neural network model are pre-trained according to the second user historical behavior data in the source domain determined; the training device of the information processing model also includes: a parameter fixing module, used to fix the parameters of the basic neural network model; in this embodiment, the first training module 1280 is further configured to: according to the first score and the second Score, adjust the parameters of the adaptive network, user feature model and discriminant model and continue training until the training end condition is reached.

In some embodiments of the present application, the basic neural network model includes an embedding layer and a plurality of cascaded converter neural networks, the first neural network model is divided into N cascaded sub-neural networks, and the sub-neural network includes a conversion The neural network of the converter and at least one adaptive network inserted into the neural network of the converter; N is an integer greater than 1; the first output module 1220 includes: an input information acquisition unit, which is used to obtain input information corresponding to the k-level sub-neural network; Among them, if k=1, the input information corresponding to the k-level sub-neural network is the embedding vector output by the embedding layer according to the interaction information sequence of the first user; if k>1, the corresponding input information of the k-level sub-neural network is k-1 The level sub-neural network outputs the k-1 level intermediate user features of the first user; 0<k≤N, k is an integer; the first input unit is used to input the input information corresponding to the k-level sub-neural network into the k-level sub-neural network The feature extraction unit is used to perform feature extraction by the k-level sub-neural network according to the corresponding input information, and output the k-level intermediate user features of the first user; wherein, if k=N, the k-level intermediate user of the first user The feature is used as the first user feature of the first user; if k<N, the k-level intermediate user features of the first user are used as the input information of the k+1-level sub-neural network.

In some embodiments of the present application, the converter neural network includes cascaded multi-head attention layers, a first feedforward neural network layer, a first summation and normalization layer, a second feedforward neural network layer, and a second feedforward neural network layer. Two summation and normalization layers; each sub-neural network includes inserting an adaptive network between the first feedforward neural network layer and the first summation and normalization layer, and inserting the second feedforward neural network layer and An adaptive network between the second summation and normalization layer; the sub-neural network also includes an identity mapping from the input of the multi-head attention layer to the input of the first summation and normalization layer, and from the second previous An identity map that feeds the input of the neural network layer to the input of the second summation and normalization layer.

In some embodiments of the present application, the adaptive network includes a cascaded first fully connected layer, an intermediate fully connected layer, an activation layer, and a final fully connected layer, and the number of neurons in the first fully connected layer is the same as that of the last fully connected layer. The number of neurons in the connection layer is equal, and the number of neurons in the first fully connected layer is greater than the number of neurons in the middle fully connected layer; the adaptive network also includes the input from the first fully connected layer to the last fully connected layer. The identity map of the output.

In some embodiments of the present application, the information processing model training device further includes: a target node selection module, configured to select multiple target nodes in the reference interaction information sequence; the reference interaction information sequence refers to the second user historical behavior data The interaction information sequence of each user; the replacement module is used to block multiple target nodes in the reference interaction information sequence to obtain the sample interaction information sequence; the sample interaction information sequence input module is used to input the sample interaction information sequence into the basic neural network model; the prediction information output module is used to output the prediction information corresponding to the target node by the basic neural network model according to the sample interaction information sequence; the prediction error determination module is used to determine the prediction error according to the prediction information and the target node; the second training The module is used to adjust the parameters of the basic neural network model according to the prediction error and continue training until the pre-training end condition is reached.

Fig. 13 is a block diagram of an information processing device according to an embodiment of the present application. As shown in Fig. 13 , the information processing device includes: a candidate information set acquisition module 1310 for acquiring candidate information sets in the target domain; behavior label prediction The module 1320 is used to predict the target user's behavior label for each information in the candidate information set through the information processing model; the information processing model is obtained according to the training method of the information processing model in any of the above-mentioned embodiments; the target information determination module 1330, It is used to determine the target information to be sent to the target user in the candidate information set according to the behavior tag.

In some embodiments of the present application, the information processing model includes a trained first neural network model, a trained user feature model, an embedding search layer, and a classification layer; the behavior label prediction module 1320 includes: a global feature search unit, using To obtain the global features of the target user in the user global feature set; the global features in the user global feature set are generated by the trained first neural network model and the trained user feature model according to the first user behavior data; the embedding vector The generation unit is used to generate the embedding vector of each information in the candidate information set by the embedding search layer; the behavior label output unit is used to output the target user according to the global characteristics of the target user and the embedding vector of each information in the candidate information set by the classification layer Behavior labels for each information in the candidate information set.

In some embodiments of the present application, the information processing device further includes: a receiving module, configured to receive an information request, and the information request indicates the user ID of the target user; a sending module, configured to send the target user the target user ID according to the target user ID target information.

In some embodiments of the present application, the information recommending device further includes: an interactive behavior information acquisition module, configured to acquire the target user's interactive behavior information on the target information; an update training module, used to update the target information according to the target user's interactive behavior information on the target information , update and train the information processing model.

Fig. 14 shows a schematic structural diagram of a computer system suitable for implementing the electronic device of the embodiment of the present application. It should be noted that the computer system 1400 of the electronic device shown in FIG. 14 is only an example, and should not limit the functions and scope of use of the embodiments of the present application.

As shown in FIG. 14 , a computer system 1400 includes a central processing unit (Central Processing Unit, CPU) 1401, which can be stored in a program in a read-only memory (Read-Only Memory, ROM) 1402 or loaded from a storage part 1408 to a random It accesses programs in the memory (Random Access Memory, RAM) 1403 to execute various appropriate actions and processes, for example, execute the methods in the above-mentioned embodiments. In the RAM 1403, various programs and data necessary for system operation are also stored. The CPU 1401 , ROM 1402 , and RAM 1403 are connected to each other via a bus 1404 . An input/output (Input/Output, I/O) interface 1405 is also connected to the bus 1404 .

The following components are connected to the I/O interface 1405: an input section 1406 including a keyboard, a mouse, etc.; an output section 1407 including a cathode ray tube (Cathode Ray Tube, CRT), a liquid crystal display (Liquid Crystal Display, LCD), etc., and a speaker ; a storage section 1408 including a hard disk or the like; and a communication section 1409 including a network interface card such as a LAN (Local Area Network) card, a modem, or the like. The communication section 1409 performs communication processing via a network such as the Internet. A drive 1410 is also connected to the I/O interface 1405 as needed. A removable medium 1411 such as a magnetic disk, optical disk, magneto-optical disk, semiconductor memory, etc. is mounted on the drive 1410 as necessary so that a computer program read therefrom is installed into the storage section 1408 as necessary.

In particular, according to the embodiments of the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments of the present application include a computer program product, which includes a computer program carried on a computer-readable medium, where the computer program includes program codes for executing the methods shown in the flowcharts. In such an embodiment, the computer program may be downloaded and installed from a network via communication portion 1409 and/or installed from removable media 1411 . When this computer program is executed by a central processing unit (CPU) 1401, various functions defined in the system of the present application are performed.

It should be noted that the computer-readable medium shown in the embodiment of the present application may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable Programmable Read-Only Memory (Erasable Programmable Read Only Memory, EPROM), flash memory, optical fiber, portable compact disk read-only memory (Compact Disc Read-Only Memory, CD-ROM), optical storage device, magnetic storage device, or any suitable The combination. In the present application, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, however, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, in which computer-readable program codes are carried. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. . Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. Wherein, each block in the flowchart or block diagram may represent a module, a program segment, or a part of the code, and the above-mentioned module, program segment, or part of the code includes one or more executable instruction. It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved. It should also be noted that each block in the block diagrams or flowchart illustrations, and combinations of blocks in the block diagrams or flowchart illustrations, can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a A combination of dedicated hardware and computer instructions.

The units described in the embodiments of the present application may be implemented by software or by hardware, and the described units may also be set in a processor. Wherein, the names of these units do not constitute a limitation of the unit itself under certain circumstances.

As another aspect, the present application also provides a computer-readable storage medium. The computer-readable medium may be contained in the electronic device described in the above-mentioned embodiments; in the device. The above-mentioned computer-readable storage medium carries computer-readable instructions, and when the computer-readable storage instructions are executed by a processor, the method in any of the above-mentioned embodiments is implemented.

According to one aspect of the present application, there is also provided an electronic device, which includes: a processor; a memory, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by the processor, the implementation of any of the above-mentioned embodiments is realized. method.

According to an aspect of the embodiments of the present application, a computer program product or computer program is provided, the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method in any one of the above embodiments.

It should be noted that although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. Actually, according to the embodiment of the present application, the features and functions of two or more modules or units described above may be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided to be embodied by a plurality of modules or units.

Through the description of the above implementations, those skilled in the art can easily understand that the example implementations described here can be implemented by software, or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of the present application can be embodied in the form of software products, which can be stored in a non-volatile storage medium (which can be CD-ROM, U disk, mobile hard disk, etc.) or on the network , including several instructions to make a computing device (which may be a personal computer, server, touch terminal, or network device, etc.) execute the method according to the embodiment of the present application.

Other embodiments of the present application will be readily apparent to those skilled in the art from consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any modification, use or adaptation of the application, these modifications, uses or adaptations follow the general principles of the application and include common knowledge or conventional technical means in the technical field not disclosed in the application .

It should be understood that the present application is not limited to the precise constructions which have been described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims (16)

1. A method for training an information processing model, comprising:

acquiring an interaction information sequence of a first user and an interaction information sequence of a second user from first user behavior data of a source domain;

outputting, by a first neural network model, a first user characteristic of the first user according to the interaction information sequence of the first user, and outputting, by the first neural network model, a first user characteristic of the second user according to the interaction information sequence of the second user; the first neural network model is pre-trained and determined according to second user historical behavior data of a source domain;

generating, by a user feature model, local features of the first user and global features of the first user according to first user features of the first user; and generating, by the user feature model, local features of the second user from first user features of the second user;

calculating, by a discriminant model, a first score based on the local features of the first user and the global features of the first user; and calculating, by the discriminative model, a second score based on the global features of the first user and the local features of the second user;

training the first neural network model, the user feature model and the discrimination model according to the first score and the second score until a training end condition is reached;

and determining an information processing model according to the trained first neural network model and the trained user characteristic model, wherein the information processing model is used for determining target information to be sent to a user in a target domain.

2. The method of claim 1, wherein the user feature model comprises at least two convolutional layers and a pooling layer, wherein the length of a convolution window corresponding to each convolutional layer in the at least two convolutional layers is different;

the generating, by the user feature model, the local feature of the first user and the global feature of the first user according to the first user feature of the first user includes:

performing convolution processing on the first user characteristic of the first user by each convolution layer of the at least two convolution layers according to the corresponding convolution window to obtain a plurality of intermediate local characteristics of the first user;

splicing the plurality of intermediate local features of the first user to obtain the local feature of the first user;

pooling the local features of the first user by the pooling layer to obtain global features of the first user;

the generating, by the user feature model, local features of the second user according to the first user features of the second user includes:

performing convolution processing on the first user characteristic of the second user by each convolution layer of the at least two convolution layers according to the corresponding convolution window to obtain a plurality of intermediate local characteristics of the second user;

and splicing the plurality of intermediate local features of the second user to obtain the local feature of the second user.

3. The method of claim 1, wherein the first neural network model comprises a base neural network model and an adaptive network inserted into the base neural network model, and parameters of the base neural network model are pre-trained and determined according to second user historical behavior data of the source domain;

before the outputting, by the first neural network model, the first user characteristic of the first user according to the interaction information sequence of the first user, the method further includes:

fixing parameters of the basic neural network model;

the training the first neural network model, the user feature model and the discriminant model according to the first score and the second score until a training end condition is reached includes:

and adjusting parameters of the adaptive network, the user characteristic model and the discriminant model according to the first score and the second score, and continuing training until a training end condition is reached.

4. The method of claim 3, wherein the base neural network model comprises an embedded layer and a plurality of cascaded transducer neural networks, the first neural network model is divided into N cascaded sub-neural networks, the sub-neural networks comprise a transducer neural network and at least one adaptive network inserted into the transducer neural network; n is an integer greater than 1;

the outputting, by a first neural network model, a first user characteristic of the first user according to the interaction information sequence of the first user includes:

acquiring input information corresponding to a k-level sub-neural network; if k =1, the input information corresponding to the k-level sub-neural network is an embedded vector output by the embedding layer according to the interaction information sequence of the first user; if k is larger than 1, the input information corresponding to the k-level sub-neural network is k-1 level sub-neural network to output k-1 level intermediate user characteristics of the first user; k is more than 0 and less than or equal to N, and k is an integer;

inputting input information corresponding to the k-level sub-neural network into the k-level sub-neural network;

performing feature extraction by the k-level sub-neural network according to the corresponding input information, and outputting k-level intermediate user features of the first user; if k = N, taking the k-level intermediate user feature of the first user as the first user feature of the first user; and if k is less than N, taking the k-level intermediate user characteristics of the first user as input information of the k + 1-level sub-neural network.

5. The method of claim 4, wherein the converter neural network comprises a cascade of a multi-headed attention layer, a first feed-forward neural network layer, a first summing and normalizing layer, a second feed-forward neural network layer, and a second summing and normalizing layer;

each said sub-neural network comprises an adaptive network interposed between said first feedforward neural network layer and said first summing and normalizing layer, and an adaptive network interposed between said second feedforward neural network layer and said second summing and normalizing layer;

the sub-neural network further includes an identity map pointing from the input of the multi-head attention layer to the input of the first summing and normalization layer, and an identity map pointing from the input of the second feed-forward neural network layer to the input of the second summing and normalization layer.

6. The method of claim 5, wherein the adaptive network comprises a first fully-connected layer, a middle fully-connected layer, an active layer and a last fully-connected layer which are cascaded, the number of neurons in the first fully-connected layer is equal to the number of neurons in the last fully-connected layer, and the number of neurons in the first fully-connected layer is greater than the number of neurons in the middle fully-connected layer; the adaptive network further includes an identity map from the input of the first fully-connected layer to the output of the last fully-connected layer.

7. The method of any of claims 3-6, wherein prior to outputting, by the first neural network model, the first user characteristic of the first user in accordance with the sequence of interaction information of the first user, the method further comprises:

selecting a plurality of target nodes in the reference interactive information sequence; the reference interaction information sequence refers to an interaction information sequence of each user in the second user historical behavior data;

shielding the target nodes in the reference interaction information sequence to obtain a sample interaction information sequence;

inputting the sample interaction information sequence into the basic neural network model;

outputting, by the basic neural network model, prediction information corresponding to the target node according to the sample interaction information sequence;

determining a prediction error according to the prediction information and the target node;

and adjusting parameters of the basic neural network model according to the prediction error and continuing training until a pre-training ending condition is reached.

8. An information processing method, characterized by comprising:

acquiring a candidate information set in a target domain;

predicting behavior labels of target users for all information in the candidate information set through the information processing model; the information processing model is trained according to the method of any one of claims 1 to 7;

and determining target information to be sent to the target user in the candidate information set according to the behavior tag.

9. The method of claim 8, wherein the information processing model comprises a trained first neural network model, a trained user feature model, an embedded lookup layer, and a classification layer; the predicting, by the information processing model, behavior labels of target users for each information in the candidate information set includes:

acquiring global features of the target user from a user global feature set; global features in the user global feature set are generated by the trained first neural network model and the trained user feature model according to the first user behavior data;

generating an embedded vector of each information in the candidate information set by the embedded search layer;

and the classification layer outputs behavior labels of the target user for each information in the candidate information set according to the global characteristics of the target user and the embedded vectors of each information in the candidate information set.

10. The method of claim 8, wherein prior to predicting, by the information handling model, a behavior label of a target user for each information in the set of candidate information, the method further comprises:

receiving an information request, wherein the information request indicates the user identification of the target user;

after determining, according to the behavior tag, target information to be sent to the target user in the candidate information set, the method further includes:

and sending the target information to the target user according to the user identification of the target user.

11. The method of claim 10, wherein after sending the target information to the target user according to the user identifier of the target user, the method further comprises:

acquiring interaction behavior information of the target user aiming at the target information;

and updating and training the information processing model according to the interactive behavior information of the target user aiming at the target information.

12. An apparatus for training an information processing model, comprising:

the first user behavior data acquisition module is used for acquiring an interaction information sequence of a first user and an interaction information sequence of a second user from first user behavior data of a source domain;

a first output module, configured to output, by a first neural network model, a first user characteristic of the first user according to the interaction information sequence of the first user, and a second output module, configured to output, by the first neural network model, a first user characteristic of the second user according to the interaction information sequence of the second user; the first neural network model is pre-trained and determined according to second user historical behavior data of a source domain;

the first generation module is used for generating the local features of the first user and the global features of the first user according to the first user features of the first user by the user feature model; the second generation module is used for generating local characteristics of the second user according to the first user characteristics of the second user by the user characteristic model;

a first score calculation module for calculating a first score by a discriminant model according to the local features of the first user and the global features of the first user; the second score calculation module is used for calculating a second score according to the global features of the first user and the local features of the second user by the discriminant model;

the first training module is used for training the first neural network model, the user feature model and the discrimination model according to the first score and the second score until a training end condition is reached;

and the information processing model determining module is used for determining an information processing model according to the trained first neural network model and the trained user characteristic model, and the information processing model determines target information to be sent to a user in a target domain.

13. An information processing apparatus characterized by comprising:

the candidate information set acquisition module is used for acquiring a candidate information set in a target domain;

the behavior tag prediction module is used for predicting the behavior tags of the target user aiming at all information in the candidate information set through the information processing model; the information processing model is trained according to the method of any one of claims 1 to 7;

and the target information determining module is used for determining target information to be sent to the target user in the candidate information set according to the behavior tag.

14. An electronic device, comprising:

a processor;

a memory having computer-readable instructions stored thereon which, when executed by the processor, implement the method of any of claims 1-11.

15. A computer-readable storage medium having computer-readable instructions stored thereon, which when executed by a processor, implement the method of any one of claims 1-11.

16. A computer program product comprising computer instructions, characterized in that the computer instructions, when executed by a processor, implement the method of any of claims 1-11.

Images