WO2022148186A1 - Behavioral sequence data processing method and apparatus - Google Patents

Abstract

The present disclosure relates to the technical field of artificial intelligence, and relates to a behavioral sequence data processing method and apparatus, an electronic device, and a storage medium. The method comprises obtaining a historical behavioral sequence of a target object, the historical behavioral sequence comprising a plurality of historical behavioral records of the target object; determining a time difference between a behavior time in each historical behavioral record and the current time; generating, on the basis of the time difference, location coding information corresponding to each historical behavioral record, the location coding information representing the distinctness between each historical behavioral record and other historical behavioral records in the plurality of historical behavioral records, and the distinctness corresponding to each historical behavioral record being inversely proportional to the time difference corresponding to each historical behavioral record; and performing coding processing on the historical behavioral sequence on the basis of the location coding information to obtain a target behavioral sequence feature.

Description

Behavior sequence data processing method and device

cross reference

This application is based on the Chinese patent application with the application number of 202110034304.X and the filing date of January 11, 2021, and claims the priority of the Chinese patent application. The entire content of the Chinese patent application is incorporated herein by reference.

technical field

The present disclosure relates to the technical field of artificial intelligence, and in particular, to a behavior sequence data processing method, device, electronic device and storage medium.

Background technique

In many scenarios, user behavior sequences need to be analyzed and processed. User behavior sequence is the occurrence process of a series of events such as clicks, visits, and purchases generated by users in daily operations. It contains the characteristics of users' fine-grained interests and preferences, and is one of the important feature sources of user-level machine learning models. .

In the related art, behavior sequences including a large number of historical behavior records of users in a long period of time are often directly used as historical data for learning user interest preferences.

SUMMARY OF THE INVENTION

The present disclosure provides a behavior sequence data processing method, device, electronic device and storage medium. The technical solutions of the present disclosure are as follows:

According to a first aspect of the embodiments of the present disclosure, a method for processing behavior sequence data is provided, including:

Obtain the historical behavior sequence of the target object, and the historical behavior sequence includes a plurality of historical behavior records of the target object;

Determine the time difference between the behavior time in each historical behavior record and the current time;

Based on the time difference, position coding information corresponding to each of the historical behavior records is generated, and the position coding information represents the degree of distinction between each of the historical behavior records and other historical behavior records in the plurality of historical behavior records, The degree of discrimination corresponding to each of the historical behavior records is inversely proportional to the time difference corresponding to each of the historical behavior records;

The historical behavior sequence is encoded based on the position encoding information to obtain the target behavior sequence feature.

In some embodiments, the method further includes:

obtaining the current behavior data of the target object, the current behavior data representing the behavior data of the recommendation information recommended by the target object to the target object at the current time;

The encoding processing of the historical behavior sequence based on the position encoding information, and obtaining the target behavior sequence features include:

The historical behavior sequence is encoded based on the position encoding information and the current behavior data, to obtain the target behavior sequence feature.

The encoding processing of the historical behavior sequence based on the location coding information and the current behavior data, to obtain the target behavior sequence features include:

Replacing the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position coding information to obtain the target behavior sequence;

Perform feature extraction on the target behavior sequence and the current behavior data to obtain initial behavior sequence features corresponding to the target behavior sequence and behavior feature information corresponding to the current behavior data;

Perform attention learning on the initial behavior sequence feature and the behavior feature information to obtain the target behavior sequence feature.

In some embodiments, the generating, based on the time difference, the location coding information corresponding to each of the historical behavior records includes:

Perform logarithmic transformation on the time difference to obtain the target time difference;

Equal interval classification is performed on the target time difference to obtain a first time difference group corresponding to a plurality of categories;

performing one-hot encoding on the first time difference groups corresponding to the multiple categories to obtain the position encoding information;

or,

The time difference is incrementally classified based on the numerical value of the time difference, and a second time difference group corresponding to multiple categories is obtained, wherein the time difference interval range of the category corresponding to the time difference corresponding to each historical behavior record is the same as that of each The time difference corresponding to the historical behavior record is inversely proportional;

One-hot encoding is performed on the second time difference groups corresponding to the multiple categories to obtain the position encoding information.

In some embodiments, the encoding processing of the historical behavior sequence based on the position encoding information to obtain the target behavior sequence feature includes:

Replacing the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position coding information to obtain the target behavior sequence;

Perform feature extraction on the target behavior sequence to obtain initial behavior sequence features corresponding to the target behavior sequence;

Perform attention learning on the initial behavior sequence features to obtain the target behavior sequence features.

In some embodiments, the encoding processing of the historical behavior sequence based on the position encoding information to obtain the target behavior sequence feature includes:

The position coding information is input to the position coding network for encoding the historical behavior sequence, so as to obtain the target behavior sequence feature.

In some embodiments, the method further includes:

Obtain sample behavior sequences of multiple sample objects and multi-task annotation results corresponding to the multiple sample objects, where the sample behavior sequence of each sample object includes multiple sample behavior records of each sample object before a preset historical time;

Determine the sample time difference between the behavior time in each sample behavior record and the preset historical time;

Based on the sample time difference, sample location coding information corresponding to each of the sample behavior records is generated, and the sample location coding information represents the corresponding value of each of the sample behavior records corresponding to each sample object and each of the sample objects. The degree of distinction between other sample behavior records in the plurality of sample behavior records, the degree of distinction corresponding to each of the sample behavior records is inversely proportional to the sample time difference corresponding to each of the sample behavior records;

Inputting the sample behavior sequence and the sample position coding information into the first neural network to be trained for encoding processing to obtain the sample behavior sequence feature;

Inputting the sample sequence features into the second neural network to be trained to perform multi-task processing to obtain multi-task prediction results corresponding to the multiple sample objects;

determining the target loss according to the multi-task prediction result and the multi-task labeling result;

The first neural network to be trained and the second neural network to be trained are trained based on the target loss to obtain the target encoding network and multitasking network.

In some embodiments, the method further includes:

Inputting the target behavior sequence feature into a multi-task processing network for multi-task processing to obtain a multi-task processing result;

Recommend target information to the target object according to the multitasking result.

According to a second aspect of the embodiments of the present disclosure, an apparatus for processing behavior sequence data is provided, including:

A historical behavior sequence acquisition module, configured to execute the acquisition of a historical behavior sequence of a target object, the historical behavior sequence including a plurality of historical behavior records of the target object;

a time difference determination module, configured to execute and determine the time difference between the behavior time in each historical behavior record and the current time;

A location coding information generation module, configured to generate location coding information corresponding to each of the historical behavior records based on the time difference, the location coding information representing each of the historical behavior records and the plurality of historical behavior records The degree of distinction between other historical behavior records in , the degree of distinction corresponding to each of the historical behavior records is inversely proportional to the time difference corresponding to each of the historical behavior records;

The first encoding processing module is configured to perform encoding processing on the historical behavior sequence based on the position encoding information to obtain a target behavior sequence feature.

In some embodiments, the apparatus further includes:

The current behavior data acquisition module is configured to execute and acquire the current behavior data of the target object, the current behavior data representing the behavior data of the recommendation information recommended by the target object to the target object at the current time;

The first encoding processing module is further configured to perform encoding processing on the historical behavior sequence based on the position encoding information and the current behavior data to obtain the target behavior sequence feature.

In some embodiments, the first encoding processing module includes:

The first position encoding unit is configured to perform the replacement of the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position encoding information to obtain the target behavior sequence;

A first feature extraction processing unit, configured to perform feature extraction on the target behavior sequence and the current behavior data, to obtain initial behavior sequence features corresponding to the target behavior sequence and behavior feature information corresponding to the current behavior data ;

The first attention learning unit is configured to perform attention learning on the initial behavior sequence feature and the behavior feature information to obtain the target behavior sequence feature.

In some embodiments, the position coding information generation module includes:

a first logarithmic transformation unit, configured to perform logarithmic transformation on the time difference to obtain a target time difference;

a first equal interval classification unit, configured to perform equal interval classification on the target time difference to obtain first time difference groups corresponding to multiple categories;

a first one-hot encoding unit, configured to perform one-hot encoding on the first time difference groups corresponding to the multiple categories to obtain the position encoding information;

or,

The first incremental classification unit is configured to perform incremental classification on the time difference based on the numerical value of the time difference, and obtain a second time difference group corresponding to a plurality of categories, wherein the time difference corresponding to each of the historical behavior records is determined. The time difference interval range of the corresponding category is inversely proportional to the time difference corresponding to each of the historical behavior records;

The second one-hot encoding unit is configured to perform one-hot encoding on the second time difference groups corresponding to the multiple categories to obtain the position encoding information.

In some embodiments, the first encoding processing module includes:

The second position encoding unit is configured to perform the replacement of the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position encoding information to obtain the target behavior sequence;

A second feature extraction unit configured to perform feature extraction on the target behavior sequence to obtain initial behavior sequence features corresponding to the target behavior sequence;

The second attention learning unit is configured to perform attention learning on the initial behavior sequence feature to obtain the target behavior sequence feature.

In some embodiments, the first encoding processing module is further configured to perform encoding processing of inputting the position encoding information into the historical behavior sequence into a position encoding network to obtain the target behavior sequence feature.

In some embodiments, the apparatus further includes:

The training data acquisition module is configured to perform acquisition of sample behavior sequences of multiple sample objects and multi-task annotation results corresponding to the multiple sample objects, and the sample behavior sequence of each sample object includes the sample behavior sequence of each sample object at a preset historical time Multiple previous sample behavior records;

a sample time difference determination module, configured to execute and determine the sample time difference between the behavior time in each sample behavior record and the preset historical time;

A sample position coding information generation module, configured to generate, based on the sample time difference, sample position coding information corresponding to each of the sample behavior records, the sample position coding information representing each of the samples corresponding to each sample object The degree of distinction between behavior records and other sample behavior records in the plurality of sample behavior records corresponding to each of the sample objects, the degree of discrimination corresponding to each of the sample behavior records and the sample corresponding to each of the sample behavior records The time difference is inversely proportional;

The second encoding processing module is configured to input the sample behavior sequence and the sample position encoding information into the first neural network to be trained for encoding processing to obtain the sample behavior sequence features;

A second multi-task processing module configured to perform multi-task processing by inputting the sample sequence features into the second neural network to be trained to obtain multi-task prediction results corresponding to the plurality of sample objects;

a target loss determination module configured to determine a target loss according to the multi-task prediction result and the multi-task labeling result;

A network training module configured to perform training of the first neural network to be trained and the second neural network to be trained based on the target loss to obtain the target encoding network and multitasking network.

In some embodiments, the apparatus further includes:

a first multi-task processing module, configured to perform multi-task processing by inputting the target behavior sequence feature into a multi-task processing network to obtain a multi-task processing result;

an information recommendation module, configured to perform recommending target information to the target object according to the multitasking result.

According to a third aspect of embodiments of the present disclosure, there is provided an electronic device, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to achieve The method of any one of the first aspects above.

According to a fourth aspect of the embodiments of the present disclosure, a non-volatile computer-readable storage medium is provided, when instructions in the storage medium are executed by a processor of an electronic device, the electronic device can execute the implementation of the present disclosure The method of any one of the first aspects of the examples.

According to a fifth aspect of the embodiments of the present disclosure, there is provided a computer program product comprising instructions which, when run on a computer, cause the computer to perform any one of the methods described in the first aspect of the embodiments of the present disclosure.

In the process of processing the historical behavior sequence, the time difference between the behavior time and the current time in multiple historical behavior records in the historical behavior sequence is combined to generate the position coding information representing the degree of discrimination between each historical behavior record and other historical behavior records. , the degree of discrimination corresponding to each historical behavior record is inversely proportional to the time difference corresponding to each historical behavior record, and when encoding and processing the historical behavior sequence, the position encoding information is added, so that the encoding process can better focus on the recent The learning of behavior records ensures that the obtained target behavior sequence features can retain more recent behavior records, which can better reflect the current real interests and preferences of the object, thereby improving the accuracy and effect of subsequent information recommendation.

It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure.

Description of drawings

The accompanying drawings, which are incorporated into and constitute a part of this specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure and do not unduly limit the present disclosure.

FIG. 1 is a schematic diagram of an application environment according to an exemplary embodiment;

2 is a flowchart of a method for processing behavior sequence data according to an exemplary embodiment;

FIG. 3 is a schematic flowchart of generating position coding information corresponding to each historical behavior record based on a time difference according to an exemplary embodiment;

FIG. 4 is a schematic flowchart illustrating a process of encoding a historical behavior sequence based on position coding information to obtain a target behavior sequence feature according to an exemplary embodiment;

FIG. 5 is a flowchart showing another kind of encoding processing of historical behavior sequences based on position encoding information to obtain characteristics of target behavior sequences according to an exemplary embodiment;

6 is a flowchart of a training target encoding network and a multitasking network according to an exemplary embodiment;

7 is a block diagram of an apparatus for processing behavior sequence data according to an exemplary embodiment;

Fig. 8 is a block diagram of an electronic device for processing behavior sequence data according to an exemplary embodiment.

Detailed ways

In order to make those skilled in the art better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

It should be noted that the terms "first", "second" and the like in the description and claims of the present disclosure and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific sequence or sequence. It is to be understood that the data so used may be interchanged under appropriate circumstances such that the embodiments of the disclosure described herein can be practiced in sequences other than those illustrated or described herein. The implementations described in the illustrative examples below are not intended to represent all implementations consistent with this disclosure. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present disclosure as recited in the appended claims. Obtaining user information and user account-related information, including social relationship identity information and the like described in the embodiments of the present disclosure, have obtained the user's permission. , devices, and storage media to obtain user-related information.

Please refer to FIG. 1 , which is a schematic diagram of an application environment according to an exemplary embodiment. As shown in FIG. 1 , the application environment may include a server 01 and a terminal 02 .

In some embodiments, Server 01 may be used to train the target encoding network. In some embodiments, the server 01 may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud service, cloud database, cloud computing, cloud function, cloud storage, Network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms.

In some embodiments, the terminal 02 may perform behavior sequence data processing in combination with the target coding network trained by the server 01 . In some embodiments, the terminal 02 may include, but is not limited to, smartphones, desktop computers, tablet computers, laptop computers, smart speakers, digital assistants, augmented reality (AR)/virtual reality (VR) devices, Electronic devices such as smart wearable devices. In some embodiments, the operating system running on the electronic device may include, but is not limited to, an Android system, an IOS system, linux, windows, and the like.

In addition, it should be noted that what is shown in FIG. 1 is only an application environment provided by the present disclosure, and other application environments may also be included in practical applications, such as training of target coding networks, which may also be implemented on the terminal 02 .

In the embodiment of this specification, the above-mentioned server 01 and the terminal 02 may be directly or indirectly connected through wired or wireless communication, which is not limited in this disclosure.

FIG. 2 is a flowchart of a method for processing behavior sequence data according to an exemplary embodiment. As shown in FIG. 2 , the method for processing behavior sequence data is used in electronic devices such as terminals and edge computing nodes, and includes the following steps:

In step S201, obtain the historical behavior sequence of the target object;

In step S203, determine the time difference between the behavior time in each historical behavior record and the current time;

In step S205, based on the time difference, the location coding information corresponding to each historical behavior record is generated;

In step S207, encoding processing is performed on the historical behavior sequence based on the position encoding information to obtain the target behavior sequence feature.

As described above, in step S201, the historical behavior sequence of the target object is acquired.

In the embodiments of the present specification, the target object may be the recommendation object of the information in the recommendation system. In some embodiments, the target object may be a single user in the recommendation system, or may be a certain group or the like.

In some embodiments, the historical behavior sequence may include multiple historical behavior records of the target object. In some embodiments, the multiple historical behavior records may be historical behavior records of the target object within a period of time before the current time (the period of time may be preset), or may be all historical behavior records of the target object before the current time. In some embodiments, each historical behavior record may represent relevant information during the behavior of the target object.

In practical applications, the target object often has multiple behaviors in different services, and the target object in the same service can also correspond to multiple behaviors, such as click behavior, video playback behavior, and the like. In the embodiments of this specification, records corresponding to one or more behaviors may be selected as historical behavior records of the target object in combination with actual application requirements.

In some embodiments, in some scenarios where users passively receive recommended videos, the user often needs to watch (play) the video for a period of time before giving feedback, resulting in a reduction in the user's active choice. It is a historical behavior sequence, which can better reflect the user's interest and preference.

In some embodiments, when multiple play history records of the target object are used as the historical behavior sequence, each play history record in the historical behavior sequence may include the video id (that is, the video identifier) watched by the target object, and the video author id (that is, author identification), video duration, video tag (topic tag), video viewing duration, video viewing time (action time), etc.

As described above, in step S203, the time difference between the action time in each historical action record and the current time is determined.

In practical applications, the long-term behavior of a user reflects the user's multi-interest distribution, while the short-term behavior often reflects the user's current interests. In some embodiments, the multiple historical behavior records are generated by the target object performing a certain behavior at different times. In the embodiment of this specification, the time difference between the behavior time in each historical behavior record and the current time can be determined by determining To distinguish whether the historical behavior record is a short-term behavior record or a long-term behavior record of the target object.

As described above, in step S205, based on the time difference, the location coding information corresponding to each historical behavior record is generated.

In the embodiment of this specification, the above-mentioned location coding information can represent the degree of distinction between each historical behavior record and other historical behavior records in the plurality of historical behavior records, and the time difference between the degree of distinction corresponding to each historical behavior record and the corresponding historical behavior record inversely proportional. That is, the smaller the time difference, the higher the degree of discrimination corresponding to the historical time record; the larger the time difference, the lower the degree of discrimination corresponding to the historical time record.

In some embodiments, as shown in FIG. 3 , based on the time difference, generating the location coding information corresponding to each historical behavior record may include the following steps:

In step S2051, logarithmically transform the time difference to obtain the target time difference;

In step S2053, the target time difference is classified into equal intervals to obtain first time difference groups corresponding to multiple categories;

In step S2055, one-hot encoding is performed on the first time difference groups corresponding to the multiple categories to obtain position encoding information.

In some embodiments, when the time difference is logarithmically transformed, an irrational number e may be used as the base, and the time difference is a true number. In practical applications, in the process of logarithmic transformation of the time difference, the smaller the time difference, the more discriminating the logarithmically transformed values; ; By performing equal interval classification on the target time difference obtained after the logarithmic change, the smaller the time difference, the finer the classification, and the larger the time difference, the coarser the classification, and then the first time difference group corresponding to multiple categories is subjected to one-hot encoding. , which can make the location coding information corresponding to the first time difference group in the same category of objects the same, and because the smaller the time difference, the finer the classification, the higher the degree of discrimination between the location coding information corresponding to the recent historical behavior records, thereby effectively guaranteeing the follow-up During encoding, recent behavior records have better discrimination.

In some embodiments, based on the time difference, generating the location coding information corresponding to each historical behavior record may include:

1) The time difference is incrementally classified based on the numerical value of the time difference, and a second time difference group corresponding to a plurality of categories is obtained, wherein the time difference interval range of the category corresponding to the time difference corresponding to each historical behavior record is the same as that of each historical behavior record. The time difference is inversely proportional;

2) One-hot encoding is performed on the second time difference groups corresponding to the multiple categories to obtain position encoding information.

In some embodiments, the time differences corresponding to the above-mentioned multiple historical time records are divided into four categories, and after the incremental classification, the four categories are: the first category, the time difference is within 0-10 minutes (including 10 minutes); The second category, the time difference is within 10-60 minutes (including 60 minutes); the third category, the time difference is within 60-180 minutes (including 180 minutes); the fourth category, the time difference is greater than 180 minutes.

Further, for the time difference corresponding to the above-mentioned multiple historical time records: the time difference within 0-10 minutes can be divided into the time difference group corresponding to the first category; the time difference within 10-60 minutes can be divided into the second category. The corresponding time difference group; the time difference within 60-180 minutes can be divided into the time difference group corresponding to the third category; the time difference greater than 180 minutes can be divided into the time difference group corresponding to the fourth category.

In the above-mentioned embodiment, the time difference corresponding to the multiple historical behavior records is incrementally classified according to the numerical value of the time difference, so that the smaller the time difference, the finer the classification, and the larger the time difference, the coarser the classification, effectively ensuring the location coding information corresponding to the recent historical behavior records. The degree of discrimination between them is higher, thereby effectively ensuring that the recent behavior records have a better degree of discrimination in the subsequent encoding process.

In some embodiments, after performing one-hot encoding on the time difference groups corresponding to the above four categories, the position encoding information corresponding to the time differences in the time difference groups corresponding to the first category may be: 1000; The time difference in the time difference group of , the corresponding position coding information may be: 0100; the time difference in the time difference group corresponding to the third category, the corresponding position coding information may be: 0010; the time difference corresponding to the fourth category For the time difference in the group, the corresponding position coding information may be: 0001.

In the above-mentioned embodiment, the time difference between the behavior time and the current time in a plurality of historical behavior records in the historical behavior sequence is combined to generate position coding information representing the degree of discrimination between each historical behavior record and other historical behavior records, and each The degree of discrimination corresponding to the historical behavior records is inversely proportional to the time difference corresponding to each historical behavior record, which can effectively ensure that the recent behavior records can be better distinguished in the subsequent encoding process.

As described above, in step S207, the historical behavior sequence is encoded based on the position encoding information to obtain the target behavior sequence feature.

In some embodiments, as shown in FIG. 4 , encoding the historical behavior sequence based on the position encoding information, and obtaining the target behavior sequence feature may include the following steps:

In step S401, replace the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position coding information to obtain the target behavior sequence;

In step S403, feature extraction is performed on the target behavior sequence to obtain initial behavior sequence features corresponding to the target behavior sequence;

In step S405, attention learning is performed on the initial behavior sequence feature to obtain the target behavior sequence feature.

In some embodiments, the initial behavior sequence feature corresponding to the target behavior sequence may be a feature vector corresponding to the target behavior sequence. In some embodiments, feature extraction may be performed on the target behavior sequence, including but not limited to combining with one-hot coding network, N-Gram (Chinese language model) and other feature extraction networks.

In some embodiments, the initial behavior sequence features may include behavior features corresponding to multiple historical behavior records. Correspondingly, performing attention learning on the initial behavior sequence features to obtain the target behavior sequence features may include adding each of the initial behavior sequence features to Dot product each behavior feature with three preset matrices, respectively, to obtain three new eigenvectors corresponding to each behavior feature; perform attention learning based on the three new eigenvectors to obtain the target behavior sequence feature.

In some embodiments, the dot product of the initial behavior sequence feature and three preset matrices is performed to obtain three corresponding new feature vectors, which may include combining the following formulas:

Q _i =X _i ·w ₁

K _i =X _i ·w ₂

V _i =X _i ·w ₃

Wherein, X _i represents the i-th behavior feature in the initial behavior sequence features of the target object, and w ₁ , w ₂ , and w ₃ represent three preset matrices, where w ₂ and w ₃ may be the same matrix. Q _i represents the first eigenvector of the three new eigenvectors corresponding to the ith behavioral feature; K _i represents the second eigenvector of the three new eigenvectors corresponding to the ith behavioral feature; V _i represents the th The third eigenvector among the three new eigenvectors corresponding to the i behavioral features.

In the above embodiment, the dot product of each behavior feature in the initial behavior sequence feature and the three preset matrices can add more features, thereby improving the coding effect.

In some embodiments, performing attention learning based on three new feature vectors to obtain target behavior sequence features may include combining the following formulas:

Among them, Z _i represents the target behavior sequence feature corresponding to the ith behavior feature; Q _i represents the first feature vector among the three new feature vectors corresponding to the ith behavior feature; V _i represents the corresponding feature of the ith behavior feature. The third eigenvector in the three new eigenvectors; K ^T represents the second eigenvector corresponding to multiple behavior features in the initial behavior sequence feature; d _k represents the second eigenvector corresponding to multiple behavior features in the initial behavior sequence feature dimension.

In the above-mentioned embodiment, in the process of encoding the historical behavior sequence, the position coding information that can characterize the degree of distinction between the historical behavior record of the target object and other historical behavior records of the target object is combined, and the distinction corresponding to each historical behavior record is combined. The degree is inversely proportional to the time difference corresponding to the historical behavior record, which can effectively ensure that the encoding process can better focus on the learning of recent behavior records, so that the obtained target behavior sequence features retain more recent behavior records, which can better Reflect the object's current real interest preferences, thereby improving the accuracy of subsequent information recommendation.

In some embodiments, the above method may further include: acquiring current behavior data of the target object, where the current behavior data represents the behavior data of the recommendation information recommended by the target object to the target object at the current time.

In practical applications, other modules responsible for information recommendation in the recommender system can recommend information to the target object at the current time, and accordingly, can obtain the current behavior data of the target object. The current behavior data represents the target object's recommendation to the target object at the current time. Behavioral data for referral information;

Correspondingly, as shown in Figure 5, the historical behavior sequence is encoded based on the location coding information and the current behavior data, and the obtained target behavior sequence features may include:

In step S501, replace the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position coding information to obtain the target behavior sequence;

In step S503, feature extraction is performed on the target behavior sequence and the current behavior data to obtain initial behavior sequence features corresponding to the target behavior sequence and behavior feature information corresponding to the current behavior data;

In step S505, perform attention learning on the initial behavior sequence feature and behavior feature information to obtain the target behavior sequence feature.

In some embodiments, the initial behavior sequence feature corresponding to the target behavior sequence may be a feature vector corresponding to the target behavior sequence, and the behavior feature information corresponding to the current behavior data may be a feature vector corresponding to the current behavior data. In some embodiments, it may include, but is not limited to, feature extraction networks such as one-hot coding network, N-Gram (Chinese language model), etc., to perform feature extraction on the target behavior sequence and current behavior data.

In some embodiments, performing attention learning on the initial behavior sequence feature and behavior feature information to obtain the target behavior sequence feature may include: performing a dot product on the behavior feature information and a first preset matrix to obtain a fourth feature vector; The behavior sequence features are dot-producted with the second preset matrix and the third preset matrix, respectively, to obtain the fifth feature vector and the sixth feature vector; based on the fourth feature vector, the fifth feature vector and the sixth feature vector, attention learning is performed, Get the target behavior sequence features.

In some embodiments, obtaining the fourth eigenvector, the fifth eigenvector and the sixth eigenvector above may include combining the following formulas:

Q=Y·w ₁

K=X·w ₂

V=X·w ₃

Wherein, X represents the initial behavior sequence feature of the target object, Y represents the behavior feature information of the target object; w ₁ , w ₂ , w ₃ represent the first preset matrix, the second preset matrix and the third preset matrix in sequence, Wherein, w ₂ and w ₃ may be the same matrix. Q represents the fourth feature vector corresponding to the behavior feature information; K represents the fifth sample feature vector corresponding to the initial behavior sequence feature; V represents the sixth sample feature vector corresponding to the initial behavior sequence feature.

In some embodiments, the attention learning is performed based on the fourth feature vector, the fifth feature vector and the sixth feature vector to obtain the target behavior sequence feature, which may include combining the following formulas:

Among them, Z represents the feature of the target behavior sequence; Q represents the fourth feature vector corresponding to the behavior feature information; K represents the fifth sample feature vector corresponding to the initial behavior sequence feature; V represents the sixth sample feature vector corresponding to the initial behavior sequence feature; K ^T represents the fifth sample feature vector corresponding to the initial behavior sequence feature (because in the self-attention learning process, except for the fifth sample feature vector corresponding to the initial behavior sequence feature, there is no fifth sample feature vector corresponding to its feature); d _k represents the dimension of the fifth sample feature vector.

In the above embodiment, in the process of encoding the historical behavior sequence, the current behavior data of the target object at the current time is added, and more object interest information can be learned. The amount of data in the encoding is small, which can effectively reduce the complexity of the encoding process, thereby improving the processing efficiency.

In some embodiments, encoding the historical behavior sequence based on the position encoding information to obtain the target behavior sequence feature may include inputting the historical behavior sequence and the position encoding information into the target encoding network for encoding processing to obtain the target behavior sequence feature. Correspondingly, the target coding network can be pre-trained. In practical applications, during the training process of the target coding network, the training can be carried out in combination with the task requirements corresponding to the practical application. In some embodiments, in the case where multi-task processing needs to be performed in combination with the target behavior sequence features output by the target encoding network, correspondingly, the above method may further include: pre-training the target encoding network and the multi-tasking network step, in some cases In an embodiment, as shown in Figure 6, it may include:

In step S601, the sample behavior sequences of the multiple sample objects and the multi-task annotation results corresponding to the multiple sample objects are obtained;

In step S603, determine the sample time difference between the behavior time in each sample behavior record and the preset historical time;

In step S605, based on the sample time difference, generate sample location coding information corresponding to each sample behavior record;

In step S607, the sample behavior sequence and the sample position coding information are input into the first neural network to be trained for encoding processing to obtain the sample behavior sequence feature;

In step S609, the sample sequence features are input into the second neural network to be trained for multi-task processing to obtain multi-task prediction results corresponding to multiple sample objects;

In step S611, the target loss is determined according to the multi-task prediction result and the multi-task labeling result;

In step S613, the first neural network to be trained and the second neural network to be trained are trained based on the target loss to obtain a target encoding network and a multitasking network.

As described above, in step S601, the sample behavior sequences of the multiple sample objects and the multi-task annotation results corresponding to the multiple sample objects are obtained.

In some embodiments, the multiple sample objects may be any number of objects in the recommendation system, and the sample behavior sequence of each sample object may include multiple sample behavior records of each sample object before a preset historical time; in some implementations In an example, the preset historical time may be a preset historical moment at which information is recommended to the sample object. In some embodiments, the historical behavior data of the information recommended by the sample object for the preset historical time may be obtained, and the multi-task annotation result corresponding to the sample object may be determined in combination with the historical behavior data. In some embodiments, the subtask annotation results of each task determined in combination with historical behavior data may be used.

In some embodiments, the historical behavior data of a sample object includes: the object identifier of the sample object, the information identifier of the historical recommendation information, the click information of clicking the historical recommendation information, the historical recommendation information is not forwarded, and it belongs to a long play. In the case where a certain task in the multi-task is to predict whether the sample object will click on the historical recommendation information in the historical behavior data; correspondingly, the subtask corresponding to the task is marked as a click. In some embodiments, 1 means click, 0 means no click; in some embodiments, in combination with the above historical behavior data, when a certain task in the multi-task is to predict whether the sample object will forward the historical recommendation information in the historical behavior data; correspondingly, The subtask corresponding to the task is marked as unforwarded. In some embodiments, 1 may be used to indicate forwarding, and 0 may be used to indicate unforwarded.

In addition, it should be noted that in the embodiment of this specification, multitasking is not limited to the two tasks listed above. In practical applications, it may also include more services in combination with actual business requirements, for example, it may also include duration-related predictions Tasks (such as whether it is an effective play, whether it is a long play, whether it is a short play, and viewing duration prediction), subdivided business prediction tasks (such as whether to download the recommended information, enter the introduction page of whether to enter the recommended information, and stay on the introduction page. duration forecast) and so on.

As described above, in step S603, the sample time difference between the behavior time in each sample behavior record and the preset historical time is determined.

In some embodiments, it can be determined whether the sample behavior record is a short-term behavior record or a long-term behavior record of the sample object at the preset historical time by determining the sample time difference between the behavior time in each sample behavior record and the preset historical time. .

As described above, in step S605, based on the sample time difference, the sample position coding information corresponding to each sample behavior record is generated.

In the embodiment of this specification, the sample location coding information represents the degree of distinction between each sample behavior record corresponding to each sample object and other sample behavior records in the multiple sample behavior records corresponding to the sample object, and the corresponding sample behavior record of each sample behavior record. The degree of discrimination is inversely proportional to the sample time difference corresponding to the sample behavior record. In some embodiments, generating the sample location coding information corresponding to each sample behavior record based on the sample time difference may include: performing logarithmic transformation on the sample time difference to obtain the target sample time difference; performing equal interval classification on the target sample time difference to obtain multiple The first sample time difference group corresponding to the category; one-hot encoding is performed on the first sample time difference group corresponding to the multiple categories to obtain the sample position encoding information.

In the above embodiment, by performing equal interval classification on the target sample time difference obtained after logarithmic change, the smaller the sample time difference is, the finer the classification is, and the larger the sample time difference is, the coarser the classification is. The one-hot encoding of the sample time difference group can make the position encoding information corresponding to the first sample time difference group in the same category of objects the same, and because the smaller the time difference, the finer the classification, and the sample position encoding corresponding to the recent sample behavior records. The discrimination between the information is higher, thereby effectively ensuring that the recent behavior records have a better discrimination in the subsequent encoding process.

In some embodiments, generating the sample location coding information corresponding to each sample behavior record based on the sample time difference may include: incrementally classifying the sample time difference based on the numerical value of the sample time difference, and obtaining a second sample time difference group corresponding to multiple categories ; Perform one-hot encoding on the second sample time difference groups corresponding to multiple categories to obtain sample position encoding information.

In the above embodiment, the sample time differences corresponding to multiple sample behavior records are incrementally classified according to the numerical value of the sample time difference, so that the smaller the sample time difference, the finer the classification, and the larger the sample time difference, the coarser the classification, which effectively ensures that the recent sample behavior records correspond to each other. The discriminative degree between the sample position coding information is higher, which effectively ensures that the recent behavior records have a better discrimination degree in the subsequent coding process.

In the embodiment of this specification, for the specific refinement of the relevant steps for generating the sample location coding information corresponding to each sample behavior record based on the sample time difference, please refer to the above-mentioned relevant steps for generating the location coding information corresponding to each historical behavior record based on the time difference. The refinement will not be repeated here.

In the above-mentioned embodiment, the sample time difference between the behavior time in the multiple sample behavior records in the sample behavior sequence and the preset historical time is combined to generate a sample behavior record that can characterize each sample behavior record corresponding to the sample object. The sample location coding information of the discrimination degree between other sample behavior records in each sample behavior record, and the discrimination degree corresponding to each sample behavior record is inversely proportional to the sample time difference corresponding to each sample behavior record, which can effectively ensure that in the subsequent encoding process, Better differentiation of recent behavior records.

As described above, in step S607, the sample behavior sequence and the sample position encoding information are input into the first neural network to be trained for encoding processing to obtain the sample behavior sequence feature.

In some embodiments, the first neural network to be trained may be an encoding network to be trained. In some embodiments, the first neural network to be trained includes: a position encoding layer to be trained, a feature extraction layer to be trained, and an attention learning layer to be trained; correspondingly, the above-mentioned sample behavior sequence and sample position encoding information are input into the first The neural network to be trained performs encoding processing to obtain the sample behavior sequence features, which may include: inputting the sample behavior sequence and sample position encoding information into the position encoding layer to be trained for position encoding to obtain the target sample behavior sequence; inputting the target sample behavior sequence into the to-be-trained feature The extraction layer performs feature extraction to obtain the initial sample behavior sequence features corresponding to the target sample behavior sequence; the initial sample behavior sequence features are input into the attention learning layer to be trained for attention learning, and the sample behavior sequence features are obtained.

In some embodiments, inputting the sample behavior sequence and the sample position encoding information into the position encoding layer to be trained for position encoding, and obtaining the target sample behavior sequence may include replacing the behavior time in each sample behavior record of the sample behavior sequence with the corresponding The sample position encoding information is used to obtain the target sample behavior sequence.

In some embodiments, the initial sample behavior sequence feature corresponding to the target sample behavior sequence may be a feature vector corresponding to the target sample behavior sequence. In some embodiments, the feature extraction layer to be trained may include, but is not limited to, one-hot (one-hot) encoding network, N-Gram (Chinese language model), and the like.

In some embodiments, the attention learning layer to be trained may be a self-attention layer in an encoding network in the Transformer. In some embodiments, inputting the initial sample behavior sequence features into the attention learning layer to be trained for attention learning, and obtaining the sample behavior sequence features may include: performing a dot product on the initial sample behavior sequence features and three preset matrices to obtain corresponding The three new feature vectors of ; based on the three new feature vectors, the attention learning is performed to obtain the sample behavior sequence features.

In the embodiment of this specification, the initial sample behavior sequence features are input into the attention learning layer to be trained for attention learning, and the specific refinement steps to obtain the sample behavior sequence features can refer to the above-mentioned attention learning of the initial behavior sequence features to obtain the target behavior The specific refinement of the sequence feature will not be repeated here. The three preset matrices may be network parameters.

In some embodiments, the attention learning layer to be trained may be a multi-head attention learning layer (ie, multiple attention learning layers), and each sample object can obtain a sample after performing attention learning in each attention learning layer Behavior sequence features, correspondingly, by splicing the sample behavior sequence features output by multiple attention learning layers, the sample behavior sequence features of each sample object after being learned by the multi-head attention learning layer can be obtained.

In the above-mentioned embodiment, in the process of encoding the sample behavior sequence, a method that can characterize the discrimination between each sample behavior record corresponding to each sample object and other sample behavior records among the multiple sample behavior records corresponding to the sample object is combined. The sample location encodes information, and the degree of discrimination corresponding to each sample behavior record is inversely proportional to the sample time difference corresponding to each sample behavior record, which can effectively ensure that the encoding process can better focus on the learning of recent behavior records, so that the obtained The sample behavior sequence feature retains more recent behavior records, which can better reflect the current real interest and preference of the object, thereby improving the accuracy of subsequent information recommendation.

In some embodiments, the above method may further include:

Obtain sample behavior data of recommendation information of multiple sample objects for a preset historical time;

Correspondingly, inputting the sample behavior sequence and the sample position encoding information into the first neural network to be trained for encoding processing, and obtaining the sample behavior sequence features may include:

The sample behavior sequence, the sample location coding information and the sample behavior data are input into the first neural network to be trained for encoding processing to obtain the sample behavior sequence feature.

In some embodiments, inputting the sample behavior sequence, the sample location coding information and the sample behavior data into the first neural network to be trained for encoding processing, and obtaining the sample behavior sequence features may include: inputting the sample behavior sequence and the sample location coding information into the to-be-trained neural network The position encoding layer performs position encoding to obtain the target sample behavior sequence; input the target sample behavior sequence and sample behavior data into the feature extraction layer to be trained for feature extraction, and obtain the initial sample behavior sequence features corresponding to the target sample behavior sequence and the sample behavior data corresponding to Sample behavior feature information; input the initial sample behavior sequence feature and sample behavior feature information into the attention learning layer to be trained for attention learning, and obtain the sample behavior sequence feature.

In some embodiments, in the above step, the initial sample behavior sequence features and sample behavior feature information are input into the attention learning layer to be trained for attention learning, and the specific refinement of the sample behavior sequence features can be found in the initial behavior sequence features and behavior features. The specific refinement steps of performing attention learning on the information to obtain the feature of the target behavior sequence will not be repeated here, wherein the first preset matrix, the second preset matrix and the third preset matrix may be network parameters.

In some embodiments, when the attention learning layer to be trained is a multi-head attention learning layer (ie, multiple attention learning layers), each sample object can obtain a sample after performing attention learning in each attention learning layer Behavior sequence features, correspondingly, by splicing the sample behavior sequence features output by multiple attention learning layers, the sample behavior sequence features of each sample object after being learned by the multi-head attention learning layer can be obtained.

In the above embodiment, in the process of encoding the sample behavior sequence, the historical behavior data of the sample object is added, and more object interest information can be learned, and the amount of the historical behavior data is often larger than that in the sample behavior sequence. It can effectively reduce the complexity of the encoding process and improve the processing efficiency.

As described above, in step S609, the sample sequence features are input into the second neural network to be trained for multi-task processing, and multi-task prediction results corresponding to multiple sample objects are obtained.

In some embodiments, the second neural network to be trained may be a multitasking network to be trained. In some embodiments, the multi-task processing network to be trained may be mmoe (Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts, multi-task learning model). In some embodiments, the second neural network to be trained includes a plurality of sub-feature extraction layers to be trained, a plurality of sub-task weighting layers to be trained, and a plurality of sub-task processing layers to be trained. Inputting the second neural network to be trained for multi-task processing, and obtaining multi-task prediction results corresponding to multiple sample objects may include:

1) Inputting the sample sequence features into multiple sub-feature extraction layers to be trained for feature extraction to obtain multiple sample sequence sub-feature information;

2) Inputting multiple sample sequence sub-feature information into multiple sub-task weighting layers to be trained to perform task weighting processing to obtain sample weighted feature information corresponding to each task;

3) The sample weighted feature information corresponding to each task is input into the corresponding subtask processing layer for subtask processing, and multitask prediction results corresponding to multiple sample objects are obtained.

In the embodiment of this specification, multiple sub-feature extraction layers to be trained can extract features from sample sequence features from different angles, but since multiple tasks share multiple sub-feature extraction layers to be trained, in order to highlight the differences between different tasks, The above-mentioned multiple weighted layers of subtasks to be trained may respectively correspond to one task; each weighted layer of subtasks to be trained may be used to weight the subfeature information of multiple sample sequences in combination with the requirements of the corresponding task. Correspondingly, the sub-feature information of the sample sequence corresponding to each task can better reflect the feature information concerned by the task. In some embodiments, the weights corresponding to the sub-feature information of a plurality of sample sequences may be determined in combination with the degree of attention of different tasks to the sub-feature information of the corresponding sample sequences.

Further, the sample weighted feature information corresponding to each task can be input into the subtask processing layer corresponding to the task for subtask processing, and the subtask prediction result corresponding to each task can be obtained, and the subtask prediction results corresponding to multiple tasks can be used as. The multi-task prediction results corresponding to the above multiple sample objects.

As described above, in step S611, the target loss is determined according to the multi-task prediction result and the multi-task labeling result.

In some embodiments, according to the multi-task prediction result and the multi-task labeling result, determining the target loss may include calculating the loss between the sub-task prediction result and the sub-task labeling result corresponding to each sample behavior data based on a preset loss function; The losses corresponding to the sample behavior data are added to obtain the above target loss.

In the embodiment of this specification, the preset loss function may include, but is not limited to, a cross-entropy loss function, a logistic loss function, a Hinge (hinge) loss function, an exponential loss function, and the like, and the embodiment of this specification is not limited to the above.

As described above, in step S613, the first neural network to be trained and the second neural network to be trained are trained based on the target loss to obtain the target encoding network and the multitasking network.

In some embodiments, training the first neural network to be trained and the second neural network to be trained based on the target loss to obtain the target encoding network and the multitasking network may include

In the case that the target loss does not meet the preset condition, update the network parameters in the first neural network to be trained and the second neural network to be trained;

The target loss is updated based on the updated first neural network to be trained and the second neural network to be trained. When the target loss meets the preset condition, the current first neural network to be trained is used as the target coding network, and the current second neural network to be trained is used as the target encoding network. The network acts as a multitasking network.

In some embodiments, the target loss meeting the preset condition may be that the target loss is less than or equal to a specified threshold, or the difference between the corresponding target losses in the two training processes before and after is less than a certain threshold. In the embodiment of this specification, the specified threshold and a certain threshold may be set in combination with actual training requirements.

In addition, it should be noted that, in practical applications, single-task processing can also be performed in combination with the target behavior sequence features output by the target encoding network.

In the above embodiment, during the training of the target coding network, the sample time difference between the behavior time in the multiple sample behavior records in the sample behavior sequence and the preset historical time is first combined to generate each sample that can characterize each sample object corresponding to each sample. The sample location coding information of the discrimination degree between the behavior record and other sample behavior records in the multiple sample behavior records corresponding to the sample object, and the discrimination degree corresponding to each sample behavior record is inversely proportional to the sample time difference corresponding to each sample behavior record , and then, in the process of encoding the sample behavior sequence, the target encoding network is added, which can effectively ensure that the encoding process can better focus on the learning of recent behavior records, so that the obtained sample behavior sequence features retain more The recent behavior records can better reflect the current real interests and preferences of the object, improve the prediction accuracy of multi-task processing results, and greatly improve the learning recommendation accuracy and recommendation effect in the recommendation system.

In some embodiments, combined with the network structure of the encoding network to be trained, the trained target encoding network may include: a feature extraction layer, a position encoding layer and an attention learning layer;

Correspondingly, the above-mentioned historical behavior sequence and position encoding information are input into the target encoding network for encoding processing, and the characteristics of the target behavior sequence obtained may include:

Input the historical behavior sequence and position coding information into the position coding layer for position coding to obtain the target behavior sequence;

Input the target behavior sequence into the feature extraction layer for feature extraction, and obtain the initial behavior sequence feature corresponding to the target behavior sequence;

The initial behavior sequence features are input into the attention learning layer for attention learning, and the target behavior sequence features are obtained.

In some embodiments, inputting the initial behavior sequence feature into the attention learning layer for attention learning, and obtaining the target behavior sequence feature may include: performing a dot product on the initial behavior sequence feature and three preset matrices to obtain corresponding three new The eigenvectors of ; based on three new eigenvectors for attention learning, the target behavior sequence features are obtained.

In the embodiment of this specification, the historical behavior sequence and the position coding information are input into the target coding network for coding processing, and the specific refinement of the relevant steps to obtain the characteristics of the target behavior sequence can refer to the above-mentioned coding processing of the historical behavior sequence based on the position coding information to obtain The specific refinement of the relevant steps of the target behavior sequence feature will not be repeated here.

In the above embodiment, in the process of encoding the historical behavior sequence, combined with the location coding information that can characterize the degree of distinction between each historical behavior record and other historical behavior records, the degree of distinction corresponding to each historical behavior record is combined with each historical behavior record. The time difference corresponding to the behavior records is inversely proportional, which can effectively ensure that in the coding process, the learning of recent behavior records can be better focused, so that the obtained target behavior sequence features retain more recent behavior records, which can better reflect the current behavior of the object. Real interest preferences, thereby improving the accuracy of subsequent information recommendation.

In some embodiments, the above-mentioned inputting the historical behavior sequence and position encoding information into the target encoding network for encoding processing, and obtaining the target behavior sequence feature may include:

The historical behavior sequence, position coding information and current behavior data are input into the target coding network for coding processing, and the target behavior sequence features are obtained.

In some embodiments, the historical behavior sequence, position encoding information and current behavior data are input into the target encoding network for encoding processing, and obtaining the target behavior sequence feature may include:

Input the historical behavior sequence and position coding information into the position coding layer for position coding to obtain the target behavior sequence;

Input the target behavior sequence and the current behavior data into the feature extraction layer for feature extraction, and obtain the initial behavior sequence feature corresponding to the target behavior sequence and the behavior feature information corresponding to the current behavior data;

The initial behavior sequence features and behavior feature information are input into the attention learning layer for attention learning, and the target behavior sequence features are obtained.

In the embodiment of this specification, the historical behavior sequence, location coding information and current behavior data are input into the target coding network for coding processing, and the specific refinement of the relevant steps to obtain the characteristics of the target behavior sequence can refer to the above-mentioned pairing based on the location coding information and the current behavior data. The historical behavior sequence is encoded and processed to obtain the specific refinement of the relevant steps of the feature of the target behavior sequence, which will not be repeated here.

In the above embodiment, in the process of encoding the historical behavior sequence, the current behavior data of the target object is added, and more object interest information can be learned, and the quantity of the current behavior data is often larger than that in the historical behavior sequence. It can effectively reduce the complexity of the encoding process and improve the processing efficiency.

In some embodiments, further optimization screening may be performed for subsequent information recommendation based on the above-mentioned target behavior sequence features. Correspondingly, the above-mentioned method further includes:

Input the target behavior sequence features into the multi-task processing network for multi-task processing, and obtain the multi-task processing result;

Recommend target information to target objects according to multi-tasking results.

In some embodiments, when the multi-task includes a task of predicting whether the target object clicks on certain recommendation information, correspondingly, the recommendation information whose corresponding task processing result is clicked may be used as the target information, and recommended to the target. object.

In the above embodiment, the information recommendation is performed by the target behavior sequence feature that can effectively reflect the current real interest preference of the target object, which can effectively ensure that the recommended information better meets the user's needs, and improves the recommendation accuracy and recommendation effect.

It can be seen from the technical solutions provided by the above embodiments of this specification that, in the process of processing the historical behavior sequence, this specification combines the time difference between the behavior time in multiple historical behavior records in the historical behavior sequence and the current time to generate a representation of each historical behavior. Record the location coding information of the degree of discrimination with other historical behavior records. The degree of discrimination corresponding to each historical behavior record is inversely proportional to the time difference corresponding to each historical behavior record, and the location code is added when encoding the historical behavior sequence. information, so that in the coding process, the learning of recent behavior records can be better focused, ensuring that the obtained target behavior sequence features can retain more recent behavior records, which can better reflect the current real interest and preference of the object, thereby improving the follow-up The accuracy and recommendation effect of information recommendation.

Fig. 7 is a block diagram of a behavior sequence data processing apparatus according to an exemplary embodiment. Referring to Figure 7, the device includes:

The historical behavior sequence acquisition module 710 is configured to execute the acquisition of the historical behavior sequence of the target object, where the historical behavior sequence includes multiple historical behavior records of the target object;

a time difference determining module 720, configured to execute and determine the time difference between the behavior time in each historical behavior record and the current time;

The location encoding information generation module 730 is configured to generate location encoding information corresponding to each historical behavior record based on the time difference, where the location encoding information represents the degree of distinction between each historical behavior record and other historical behavior records in the multiple historical behavior records , the discrimination corresponding to each historical behavior record is inversely proportional to the time difference corresponding to each historical behavior record;

The first encoding processing module 740 is configured to perform encoding processing on the historical behavior sequence based on the position encoding information to obtain the target behavior sequence feature.

In some embodiments, the above-mentioned apparatus further comprises:

The current behavior data acquisition module is configured to execute the acquisition of the current behavior data of the target object, and the current behavior data represents the behavior data of the recommendation information recommended by the target object to the target object at the current time;

The first encoding processing module 740 is further configured to perform encoding processing on the historical behavior sequence based on the position encoding information and the current behavior data to obtain the target behavior sequence feature.

In some embodiments, the first encoding processing module 740 includes:

The first position encoding unit is configured to perform the replacement of the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position encoding information to obtain the target behavior sequence;

a first feature extraction processing unit, configured to perform feature extraction on the target behavior sequence and current behavior data, to obtain initial behavior sequence features corresponding to the target behavior sequence and behavior feature information corresponding to the current behavior data;

The first attention learning unit is configured to perform attention learning on the initial behavior sequence feature and behavior feature information to obtain the target behavior sequence feature.

In some embodiments, the position coding information generation module 730 includes:

a first logarithmic transformation unit, configured to perform logarithmic transformation on the time difference to obtain the target time difference;

a first equal interval classification unit, configured to perform equal interval classification on the target time difference, and obtain first time difference groups corresponding to multiple categories;

a first one-hot encoding unit, configured to perform one-hot encoding on the first time difference groups corresponding to multiple categories to obtain position encoding information;

or,

The first incremental classification unit is configured to perform incremental classification of the time difference based on the numerical value of the time difference, and obtain a second time difference group corresponding to multiple categories, wherein the time difference interval range of the category corresponding to the time difference corresponding to each historical behavior record Inversely proportional to the time difference corresponding to each historical behavior record;

The second one-hot encoding unit is configured to perform one-hot encoding on the second time difference groups corresponding to the multiple categories to obtain position encoding information.

In some embodiments, the first encoding processing module 740 includes:

The second position encoding unit is configured to perform the replacement of the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position encoding information to obtain the target behavior sequence;

The second feature extraction unit is configured to perform feature extraction on the target behavior sequence to obtain initial behavior sequence features corresponding to the target behavior sequence;

The second attention learning unit is configured to perform attention learning on initial behavior sequence features to obtain target behavior sequence features.

In some embodiments, the first encoding processing module is further configured to perform encoding processing of inputting the position encoding information into the position encoding network of the historical behavior sequence to obtain the target behavior sequence feature.

In some embodiments, the above-mentioned apparatus further comprises:

The training data acquisition module is configured to perform acquisition of sample behavior sequences of multiple sample objects and multi-task annotation results corresponding to the multiple sample objects. The sample behavior sequence of each sample object includes the sample behavior sequence of each sample object before the preset historical time. Multiple sample behavior records;

The sample time difference determination module is configured to execute and determine the sample time difference between the behavior time in each sample behavior record and the preset historical time;

The sample position coding information generation module is configured to generate sample position coding information corresponding to each sample behavior record based on the sample time difference, and the sample position coding information represents that each sample behavior record corresponding to each sample object corresponds to each sample object The degree of distinction between other sample behavior records in the multiple sample behavior records, the discrimination degree corresponding to each sample behavior record is inversely proportional to the sample time difference corresponding to each sample behavior record;

The second encoding processing module is configured to input the sample behavior sequence and the sample position encoding information into the first neural network to be trained for encoding processing to obtain the sample behavior sequence feature;

The second multi-task processing module is configured to perform multi-task processing by inputting the sample sequence features into the second neural network to be trained to obtain multi-task prediction results corresponding to multiple sample objects;

The target loss determination module is configured to determine the target loss according to the multi-task prediction result and the multi-task labeling result;

The network training module is configured to perform training of the first neural network to be trained and the second neural network to be trained based on the target loss to obtain the target encoding network and the multitasking network.

In some embodiments, the above-mentioned apparatus further includes:

The first multi-task processing module is configured to perform multi-task processing by inputting the target behavior sequence feature into the multi-task processing network to obtain a multi-task processing result;

The information recommendation module is configured to perform recommending target information to the target object according to the multi-tasking result.

Regarding the apparatus in the above-mentioned embodiment, the specific manner in which each module performs the operation has been described in detail in the embodiment of the method, and will not be described in detail here.

FIG. 8 is a block diagram of an electronic device for processing behavior sequence data according to an exemplary embodiment. The electronic device may be a terminal, and its internal structure diagram may be as shown in FIG. 8 . The electronic device includes a processor, a memory, a network interface, a display screen, and an input device connected by a system bus. Among them, the processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile computer-readable storage medium and an internal memory. The computer-readable storage medium stores an operating system and a computer program. The internal memory provides an environment for the execution of the operating system and computer programs in the computer-readable storage medium. The network interface of the electronic device is used to communicate with an external terminal through a network connection. The computer program, when executed by a processor, implements a method for processing behavior sequence data. The display screen of the electronic device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the electronic device may be a touch layer covered on the display screen, or a button, a trackball or a touchpad set on the shell of the electronic device , or an external keyboard, trackpad, or mouse.

Those skilled in the art can understand that the structure shown in FIG. 8 is only a block diagram of a part of the structure related to the solution of the present disclosure, and does not constitute a limitation on the electronic device to which the solution of the present disclosure is applied. In some embodiments , an electronic device may include more or fewer components than shown in the figures, or combine certain components, or have a different arrangement of components.

In an exemplary embodiment, there is also provided an electronic device, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the present disclosure The behavior sequence data processing method in the example.

In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, which, when the instructions in the computer-readable storage medium are executed by a processor of the electronic device, enables the electronic device to perform the embodiments of the present disclosure Behavior sequence data processing methods in . In some embodiments, the computer-readable storage medium may be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, and the like.

In an exemplary embodiment, there is also provided a computer program product comprising instructions which, when executed on a computer, cause the computer to execute the behavior sequence data processing method in the embodiment of the present disclosure.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium , when the computer program is executed, it may include the processes of the above-mentioned method embodiments. Wherein, any reference to memory, storage, database or other medium used in the various embodiments provided in this application may include non-volatile and/or volatile memory.

Other embodiments of the present disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or techniques in the technical field not disclosed by the present disclosure . The specification and examples are to be regarded as exemplary only, with the true scope and spirit of the disclosure being indicated by the following claims.

It is to be understood that the present disclosure is not limited to the precise structures described above and illustrated in the accompanying drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

All the embodiments of the present disclosure can be implemented independently or in combination with other embodiments, which are all regarded as the protection scope of the present disclosure.

Claims (26)

A method for processing behavior sequence data, comprising: Obtain the historical behavior sequence of the target object, and the historical behavior sequence includes a plurality of historical behavior records of the target object; Determine the time difference between the behavior time in each historical behavior record and the current time; Based on the time difference, position coding information corresponding to each of the historical behavior records is generated, and the position coding information represents the degree of distinction between each of the historical behavior records and other historical behavior records in the plurality of historical behavior records, The degree of discrimination corresponding to each of the historical behavior records is inversely proportional to the time difference corresponding to each of the historical behavior records; The historical behavior sequence is encoded based on the position encoding information to obtain the target behavior sequence feature. The behavior sequence data processing method according to claim 1, wherein the method further comprises: obtaining the current behavior data of the target object, the current behavior data representing the behavior data of the recommendation information recommended by the target object to the target object at the current time; The encoding processing of the historical behavior sequence based on the position encoding information, and obtaining the target behavior sequence features include: The historical behavior sequence is encoded based on the position encoding information and the current behavior data, to obtain the target behavior sequence feature. The behavior sequence data processing method according to claim 2, wherein the encoding processing of the historical behavior sequence based on the location coding information and the current behavior data, and obtaining the target behavior sequence feature comprises: Replacing the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position coding information to obtain the target behavior sequence; Perform feature extraction on the target behavior sequence and the current behavior data to obtain initial behavior sequence features corresponding to the target behavior sequence and behavior feature information corresponding to the current behavior data; Perform attention learning on the initial behavior sequence feature and the behavior feature information to obtain the target behavior sequence feature. The behavior sequence data processing method according to claim 1, wherein the generating, based on the time difference, a location code letter corresponding to each of the historical behavior records comprises: logarithmically transform the time difference to obtain the target time difference; Equal interval classification is performed on the target time difference to obtain a first time difference group corresponding to a plurality of categories; performing one-hot encoding on the first time difference groups corresponding to the multiple categories to obtain the position encoding information; or, The time difference is incrementally classified based on the numerical value of the time difference, and a second time difference group corresponding to multiple categories is obtained, wherein the time difference interval range of the category corresponding to the time difference corresponding to each historical behavior record is the same as that of each The time difference corresponding to the historical behavior record is inversely proportional; One-hot encoding is performed on the second time difference groups corresponding to the multiple categories to obtain the position encoding information. The behavior sequence data processing method according to claim 1, wherein the encoding processing of the historical behavior sequence based on the position coding information to obtain the target behavior sequence feature comprises: Replacing the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position coding information to obtain the target behavior sequence; Perform feature extraction on the target behavior sequence to obtain initial behavior sequence features corresponding to the target behavior sequence; Perform attention learning on the initial behavior sequence features to obtain the target behavior sequence features. The behavior sequence data processing method according to any one of claims 1 to 5, wherein the encoding processing is performed on the historical behavior sequence based on the position coding information, and the characteristics of the target behavior sequence obtained include: The position coding information is input to the position coding network for encoding the historical behavior sequence, so as to obtain the target behavior sequence feature. The behavior sequence data processing method according to claim 6, wherein the method further comprises: Obtain sample behavior sequences of multiple sample objects and multi-task annotation results corresponding to the multiple sample objects, where the sample behavior sequence of each sample object includes multiple sample behavior records of each sample object before a preset historical time; Determine the sample time difference between the behavior time in each sample behavior record and the preset historical time; Based on the sample time difference, sample location coding information corresponding to each of the sample behavior records is generated, and the sample location coding information represents the corresponding value of each of the sample behavior records corresponding to each sample object and each of the sample objects. The degree of distinction between other sample behavior records in the plurality of sample behavior records, the degree of distinction corresponding to each of the sample behavior records is inversely proportional to the sample time difference corresponding to each of the sample behavior records; Inputting the sample behavior sequence and the sample position coding information into the first neural network to be trained for encoding processing to obtain the sample behavior sequence feature; Inputting the sample sequence features into the second neural network to be trained to perform multi-task processing to obtain multi-task prediction results corresponding to the multiple sample objects; determining the target loss according to the multi-task prediction result and the multi-task labeling result; The first neural network to be trained and the second neural network to be trained are trained based on the target loss to obtain the target encoding network and multitasking network. The behavior sequence data processing method according to claim 7, wherein the method further comprises: Inputting the target behavior sequence feature into the multi-task processing network for multi-task processing to obtain a multi-task processing result; Recommend target information to the target object according to the multitasking result. A behavior sequence data processing device, characterized in that it includes: A historical behavior sequence acquisition module, configured to execute the acquisition of a historical behavior sequence of a target object, the historical behavior sequence including a plurality of historical behavior records of the target object; a time difference determination module, configured to execute and determine the time difference between the behavior time in each historical behavior record and the current time; A location coding information generation module, configured to generate location coding information corresponding to each of the historical behavior records based on the time difference, the location coding information representing each of the historical behavior records and the plurality of historical behavior records The degree of distinction between other historical behavior records in , the degree of distinction corresponding to each of the historical behavior records is inversely proportional to the time difference corresponding to each of the historical behavior records; The first encoding processing module is configured to perform encoding processing on the historical behavior sequence based on the position encoding information to obtain a target behavior sequence feature. The behavior sequence data processing device according to claim 9, wherein the device further comprises: The current behavior data acquisition module is configured to execute and acquire the current behavior data of the target object, the current behavior data representing the behavior data of the recommendation information recommended by the target object to the target object at the current time; The first encoding processing module is further configured to perform encoding processing on the historical behavior sequence based on the position encoding information and the current behavior data to obtain the target behavior sequence feature. The behavior sequence data processing device according to claim 10, wherein the first encoding processing module comprises: The first position encoding unit is configured to perform the replacement of the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position encoding information to obtain the target behavior sequence; A first feature extraction processing unit, configured to perform feature extraction on the target behavior sequence and the current behavior data, to obtain initial behavior sequence features corresponding to the target behavior sequence and behavior feature information corresponding to the current behavior data ; The first attention learning unit is configured to perform attention learning on the initial behavior sequence feature and the behavior feature information to obtain the target behavior sequence feature. The behavior sequence data processing device according to claim 9, wherein the position coding information generating module comprises: a first logarithmic transformation unit, configured to perform logarithmic transformation on the time difference to obtain a target time difference; a first equal interval classification unit, configured to perform equal interval classification on the target time difference to obtain first time difference groups corresponding to multiple categories; a first one-hot encoding unit, configured to perform one-hot encoding on the first time difference groups corresponding to the multiple categories to obtain the position encoding information; or, The first incremental classification unit is configured to perform incremental classification on the time difference based on the numerical value of the time difference, and obtain a second time difference group corresponding to a plurality of categories, wherein the time difference corresponding to each of the historical behavior records is determined. The time difference interval range of the corresponding category is inversely proportional to the time difference corresponding to each of the historical behavior records; The second one-hot encoding unit is configured to perform one-hot encoding on the second time difference groups corresponding to the multiple categories to obtain the position encoding information. The behavior sequence data processing device according to claim 9, wherein the first encoding processing module comprises: The second position encoding unit is configured to perform the replacement of the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position encoding information to obtain the target behavior sequence; A second feature extraction unit configured to perform feature extraction on the target behavior sequence to obtain initial behavior sequence features corresponding to the target behavior sequence; The second attention learning unit is configured to perform attention learning on the initial behavior sequence feature to obtain the target behavior sequence feature. The behavior sequence data processing device according to any one of claims 9 to 13, wherein the first encoding processing module is further configured to perform a process of inputting the position encoding information to the historical behavior sequence into a position encoding network. The encoding process is performed to obtain the target behavior sequence feature. The behavior sequence data processing device according to claim 14, wherein the device further comprises: The training data acquisition module is configured to perform acquisition of sample behavior sequences of multiple sample objects and multi-task annotation results corresponding to the multiple sample objects, and the sample behavior sequence of each sample object includes the sample behavior sequence of each sample object at a preset historical time Multiple previous sample behavior records; a sample time difference determination module, configured to execute and determine the sample time difference between the behavior time in each sample behavior record and the preset historical time; A sample position coding information generation module, configured to generate, based on the sample time difference, sample position coding information corresponding to each of the sample behavior records, the sample position coding information representing each of the samples corresponding to each sample object The degree of distinction between behavior records and other sample behavior records in the plurality of sample behavior records corresponding to each of the sample objects, the degree of discrimination corresponding to each of the sample behavior records and the sample corresponding to each of the sample behavior records The time difference is inversely proportional; The second encoding processing module is configured to input the sample behavior sequence and the sample position encoding information into the first neural network to be trained for encoding processing to obtain the sample behavior sequence features; A second multi-task processing module configured to perform multi-task processing by inputting the sample sequence features into the second neural network to be trained to obtain multi-task prediction results corresponding to the plurality of sample objects; a target loss determination module configured to determine a target loss according to the multi-task prediction result and the multi-task labeling result; A network training module configured to perform training of the first neural network to be trained and the second neural network to be trained based on the target loss to obtain the target encoding network and multitasking network. The behavior sequence data processing device according to claim 15, wherein the device further comprises: a first multi-task processing module, configured to perform multi-task processing by inputting the target behavior sequence feature into a multi-task processing network to obtain a multi-task processing result; an information recommendation module, configured to perform recommending target information to the target object according to the multitasking result. An electronic device, comprising: processor; memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the following steps: Obtain the historical behavior sequence of the target object, and the historical behavior sequence includes a plurality of historical behavior records of the target object; Determine the time difference between the behavior time in each historical behavior record and the current time; Based on the time difference, position coding information corresponding to each of the historical behavior records is generated, and the position coding information represents the degree of discrimination between each of the historical behavior records and other historical behavior records in the plurality of historical behavior records, The degree of discrimination corresponding to each of the historical behavior records is inversely proportional to the time difference corresponding to each of the historical behavior records; The historical behavior sequence is encoded based on the position encoding information to obtain the target behavior sequence feature. 18. The electronic device of claim 17, wherein the processor is configured to execute the instructions to implement the following steps: obtaining the current behavior data of the target object, the current behavior data representing the behavior data of the recommendation information recommended by the target object to the target object at the current time; The encoding processing of the historical behavior sequence based on the position encoding information, and obtaining the target behavior sequence features include: The historical behavior sequence is encoded based on the position encoding information and the current behavior data, to obtain the target behavior sequence feature. The electronic device according to claim 18, wherein the encoding processing of the historical behavior sequence based on the location coding information and the current behavior data, and obtaining the target behavior sequence feature comprises: Replacing the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position coding information to obtain the target behavior sequence; Perform feature extraction on the target behavior sequence and the current behavior data to obtain initial behavior sequence features corresponding to the target behavior sequence and behavior feature information corresponding to the current behavior data; Perform attention learning on the initial behavior sequence feature and the behavior feature information to obtain the target behavior sequence feature. The electronic device according to claim 17, wherein the generating, based on the time difference, a location code letter corresponding to each of the historical behavior records comprises: logarithmically transform the time difference to obtain the target time difference; Perform equal interval classification on the target time difference to obtain first time difference groups corresponding to multiple categories; performing one-hot encoding on the first time difference groups corresponding to the multiple categories to obtain the position encoding information; or, The time difference is incrementally classified based on the numerical value of the time difference, and a second time difference group corresponding to a plurality of categories is obtained, wherein the time difference interval range of the category corresponding to the time difference corresponding to each historical behavior record is the same as that of each The time difference corresponding to the historical behavior record is inversely proportional; One-hot encoding is performed on the second time difference groups corresponding to the multiple categories to obtain the position encoding information. The electronic device according to claim 17, wherein the encoding processing of the historical behavior sequence based on the position encoding information to obtain the characteristics of the target behavior sequence comprises: Replacing the behavior time of each historical behavior record in the historical behavior sequence with the corresponding position coding information to obtain the target behavior sequence; Perform feature extraction on the target behavior sequence to obtain initial behavior sequence features corresponding to the target behavior sequence; Perform attention learning on the initial behavior sequence features to obtain the target behavior sequence features. The electronic device according to any one of claims 17 to 21, wherein the encoding processing of the historical behavior sequence based on the position encoding information to obtain the characteristics of the target behavior sequence comprises: The position coding information is input to the position coding network for encoding the historical behavior sequence, so as to obtain the target behavior sequence feature. 23. The electronic device of claim 22, wherein the processor is configured to execute the instructions to implement the following steps: Obtain sample behavior sequences of multiple sample objects and multi-task annotation results corresponding to the multiple sample objects, where the sample behavior sequence of each sample object includes multiple sample behavior records of each sample object before a preset historical time; Determine the sample time difference between the behavior time in each sample behavior record and the preset historical time; Based on the sample time difference, sample location coding information corresponding to each of the sample behavior records is generated, and the sample location coding information represents the corresponding value of each of the sample behavior records corresponding to each sample object and each of the sample objects. The degree of distinction between other sample behavior records in the plurality of sample behavior records, the degree of distinction corresponding to each of the sample behavior records is inversely proportional to the sample time difference corresponding to each of the sample behavior records; Inputting the sample behavior sequence and the sample position coding information into the first neural network to be trained for encoding processing to obtain the sample behavior sequence feature; Inputting the sample sequence features into the second neural network to be trained for multi-task processing to obtain multi-task prediction results corresponding to the multiple sample objects; determining the target loss according to the multi-task prediction result and the multi-task labeling result; The first neural network to be trained and the second neural network to be trained are trained based on the target loss to obtain the target encoding network and multitasking network. 24. The electronic device of claim 23, wherein the processor is configured to execute the instructions to implement the following steps: Inputting the target behavior sequence feature into the multi-task processing network for multi-task processing to obtain a multi-task processing result; Recommend target information to the target object according to the multitasking result. A non-volatile computer-readable storage medium, characterized in that, when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device can implement the following steps: Obtain the historical behavior sequence of the target object, and the historical behavior sequence includes a plurality of historical behavior records of the target object; Determine the time difference between the behavior time in each historical behavior record and the current time; Based on the time difference, generate each of the historical behavior records corresponding to The position coding information, the position coding information represents the degree of distinction between each of the historical behavior records and other historical behavior records in the plurality of historical behavior records, and the degree of distinction corresponding to each of the historical behavior records is the same as that of each historical behavior record. The time difference corresponding to the historical behavior record is inversely proportional; The historical behavior sequence is encoded based on the position encoding information to obtain the target behavior sequence feature. A computer program product, comprising a computer program, characterized in that, when the computer program is executed by a processor, the following steps are implemented: Obtain the historical behavior sequence of the target object, and the historical behavior sequence includes a plurality of historical behavior records of the target object; Determine the time difference between the behavior time in each historical behavior record and the current time; Based on the time difference, position coding information corresponding to each of the historical behavior records is generated, and the position coding information represents the degree of discrimination between each of the historical behavior records and other historical behavior records in the plurality of historical behavior records, The degree of discrimination corresponding to each of the historical behavior records is inversely proportional to the time difference corresponding to each of the historical behavior records; The historical behavior sequence is encoded based on the position encoding information to obtain the target behavior sequence feature.

Images