CN113469752A

CN113469752A - Content recommendation method and device, storage medium and electronic equipment

Info

Publication number: CN113469752A
Application number: CN202110832808.6A
Authority: CN
Inventors: 王莹; 肖军波; 安伟亭; 葛覃; 翟仕召; 宋烈金
Original assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Wodong Tianjun Information Technology Co Ltd
Current assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Wodong Tianjun Information Technology Co Ltd
Priority date: 2021-07-22
Filing date: 2021-07-22
Publication date: 2021-10-01

Abstract

The present disclosure relates to the field of computer technologies, and in particular, to a content recommendation method and apparatus, a storage medium, and an electronic device. The content recommendation method comprises the steps of responding to an access request of a target user, and obtaining a user characteristic vector of the target user; performing preference evaluation based on the user characteristic vector of the target user and the content characteristic vector of the candidate content to obtain a preference result of the target user on the candidate content; selecting preference content from the candidate content according to the preference result so as to recommend the preference content to the target user. The content recommendation method provided by the disclosure can solve the problems of low efficiency, low accuracy and incomplete coverage in the existing activity circle person recommendation scheme.

Description

Content recommendation method and device, storage medium and electronic equipment

Technical Field

The present disclosure relates to the field of computer technologies, and in particular, to a content recommendation method, a content recommendation apparatus, a computer-readable storage medium, and an electronic device.

Background

In an e-commerce scenario, in order to attract user traffic and attention, operators often set up marketing campaigns for parts of commodities and circle groups of people who may be interested in the marketing campaigns.

At present, the main activity recommendation people circling scheme mainly depends on the past experience of operation one day in advance, and people who may be interested in activities are released by manually selecting labels to circle. Therefore, the current recommendation circle strategy has strong dependence on manpower, low efficiency, strong dependence on service experience of operators and low accuracy, and meanwhile, the current circle recommendation algorithm is based on that users possibly interested in the part of the label which is away from the circle selection part one day ahead of time, so that the range of active release crowd is limited.

It is to be noted that the information disclosed in the above background section is only for enhancement of understanding of the background of the present disclosure, and thus may include information that does not constitute prior art known to those of ordinary skill in the art.

Disclosure of Invention

The present disclosure aims to provide a content recommendation method, a content recommendation apparatus, a computer-readable storage medium, and an electronic device, and aims to solve the problems of low efficiency, low accuracy, and incomplete coverage in the existing activity circle person recommendation scheme.

Additional features and advantages of the disclosure will be set forth in the detailed description which follows, or in part will be obvious from the description, or may be learned by practice of the disclosure.

According to an aspect of an embodiment of the present disclosure, there is provided a content recommendation method including: responding to an access request of a target user, and acquiring a user characteristic vector of the target user; performing preference evaluation based on the user characteristic vector of the target user and the content characteristic vector of the candidate content to obtain a preference result of the target user on the candidate content; selecting preference content from the candidate content according to the preference result so as to recommend the preference content to the target user.

According to some embodiments of the present disclosure, based on the foregoing scheme, the obtaining the user feature vector of the target user includes: acquiring basic information of the target user; and extracting the user feature vector of the target user from a preset user feature database based on the basic information.

According to some embodiments of the present disclosure, based on the foregoing solution, the method further comprises pre-constructing the user characteristic database, the pre-constructing the user characteristic database comprising: periodically acquiring user attribute information and user behavior information of historical users; based on the user attribute information and the user behavior information, performing feature extraction on the historical user by using a user side sub-model of a double-tower model to obtain a user feature vector of the historical user; and storing the historical users and the user feature vectors of the historical users to construct the user feature database.

According to some embodiments of the present disclosure, based on the foregoing scheme, the performing, by using a user-side sub-model of a double-tower model, feature extraction on the historical user based on the user attribute information and the user behavior information to obtain a user feature vector of the historical user includes: performing feature processing on the user attribute information to obtain user attribute features; performing characteristic conversion on the user behavior information to obtain user content characteristics; and performing feature splicing on the user attribute features and the user content features to obtain the user feature vector of the historical user.

According to some embodiments of the present disclosure, based on the foregoing scheme, the performing feature conversion on the user behavior information to obtain a user content feature includes: performing feature mapping on the user behavior information to obtain a content category vector based on the content category information corresponding to the user behavior information; and carrying out mean processing on the content category vectors to obtain the user content characteristics.

According to some embodiments of the present disclosure, based on the foregoing solution, the method further comprises: and extracting the content feature vector of the candidate content from a preset content feature database.

According to some embodiments of the present disclosure, based on the foregoing solution, the method further includes pre-constructing the content feature database, including: acquiring content information of the candidate content; performing feature extraction on the candidate content by using a content side sub-model of a double-tower model based on the content information to obtain a content feature vector of the candidate content; storing the candidate content and the content feature vector of the candidate content to construct the content feature database.

According to some embodiments of the present disclosure, based on the foregoing scheme, the content information includes content category information, content brand information, and content attribute information; the performing feature extraction on the candidate content by using a content side sub-model of a double-tower model based on the content information to obtain a content feature vector of the candidate content includes: performing feature mapping on the content category information to obtain a content category vector, and performing weighting processing on the content category vector to obtain content category features; performing feature mapping on the content brand information to obtain a content brand vector, and performing weighting processing on the content brand vector to obtain content brand features; and carrying out feature processing on the content attribute information to obtain content attribute features; and performing feature splicing on the content category features, the content brand features and the content attribute features to obtain content feature vectors of the candidate content.

According to some embodiments of the present disclosure, based on the foregoing solution, the performing preference evaluation based on the user feature vector of the target user and the content feature vector of the candidate content to obtain a preference result of the target user for the candidate content includes: determining a target content from the candidate contents; and based on the user characteristic vector of the target user and the content characteristic vector of the target content, performing preference evaluation by using a preference evaluation sub-model of the double-tower model to obtain a preference result.

According to some embodiments of the present disclosure, based on the foregoing solution, the performing, by using a preference evaluation sub-model of a two-tower model, a preference evaluation to obtain the preference result based on the user feature vector of the target user and the content feature vector of the target content includes: calculating the cross feature of the content feature vector and the user feature vector by using a multi-layer perceptron; determining a preference value of the target content as the preference result based on the cross feature.

According to some embodiments of the present disclosure, based on the foregoing solution, the selecting, according to the preference result, preferred content from the candidate content for recommending the preferred content to the target user includes: sorting the candidate contents according to the preference value of each candidate content in the preference result to obtain a sorting result; selecting preference content based on the sorting result, and recommending the preference content to the target user.

According to some embodiments of the present disclosure, based on the foregoing solution, the method further comprises constructing the double-tower model, the constructing the double-tower model comprising: generating a training set according to historical recommendation data; and carrying out model training through the training set to obtain parameters of an input layer, a representation layer and a matching layer so as to obtain the double-tower model.

According to some embodiments of the present disclosure, based on the foregoing solution, the method further comprises: storing an input layer and a representation layer corresponding to a user tower in the double-tower model as a user side sub-model; storing an input layer and a representation layer corresponding to a content tower in the double-tower model as a content side sub-model; and

and storing the matching layer in the double-tower model as a preference evaluation submodel.

According to a second aspect of the embodiments of the present disclosure, there is provided a content recommendation apparatus including: the response module is used for responding to an access request of a target user and acquiring a user characteristic vector of the target user; the evaluation module is used for carrying out preference evaluation on the basis of the user characteristic vector of the target user and the content characteristic vector of the candidate content to obtain a preference result of the target user on the candidate content; and the recommending module is used for selecting the preference content from the candidate contents according to the preference result so as to recommend the preference content to the target user.

According to a third aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements a content recommendation method as in the above embodiments.

According to a fourth aspect of the embodiments of the present disclosure, there is provided an electronic apparatus, including: one or more processors; a storage device for storing one or more programs which, when executed by the one or more processors, cause the one or more processors to implement the content recommendation method as in the above embodiments.

Exemplary embodiments of the present disclosure may have some or all of the following benefits:

in the technical solutions provided by some embodiments of the present disclosure, a user feature vector characterization is used for a user, a content feature vector characterization is used for candidate content, when an access request of a target user is received, preference evaluation can be performed according to the user feature vector of the target user and content feature vectors of all candidate content to obtain a preference result of the target user for each candidate content, and finally, the preference content of the target user is determined according to the preference result for recommendation. On one hand, the content recommendation method provided by the disclosure responds in real time to perform preference evaluation to obtain a recommendation result when receiving an access request, avoids manual selection of a tag for selection, and improves the content recommendation efficiency; on the other hand, the user and the content are subjected to vector representation for preference evaluation instead of recommendation by using the existing label, so that the accuracy of content recommendation is improved; on the other hand, content recommendation is carried out according to the accessed target users, the accessed users can be preferentially processed, people groups are prevented from being selected in advance, and the crowd range of content delivery is enlarged.

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

Drawings

The accompanying drawings, which are incorporated in 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 disclosure. It is to be understood that the drawings in the following description are merely exemplary of the disclosure, and that other drawings may be derived from those drawings by one of ordinary skill in the art without the exercise of inventive faculty. In the drawings:

fig. 1 schematically illustrates a flow chart of a content recommendation method in an exemplary embodiment of the present disclosure;

FIG. 2 schematically illustrates a structural schematic of a double tower model in an exemplary embodiment of the disclosure;

FIG. 3 is a schematic diagram illustrating a process for building a user profile database according to an exemplary embodiment of the present disclosure;

FIG. 4 schematically illustrates a flow diagram for building a content feature database in an exemplary embodiment of the disclosure;

FIG. 5 is a flow chart diagram schematically illustrating a content recommendation method in an exemplary embodiment of the present disclosure;

fig. 6 schematically illustrates a composition diagram of a content recommendation device in an exemplary embodiment of the present disclosure;

FIG. 7 schematically illustrates a schematic diagram of a computer-readable storage medium in an exemplary embodiment of the disclosure;

fig. 8 schematically shows a structural diagram of a computer system of an electronic device in an exemplary embodiment of the disclosure.

Detailed Description

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

Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the disclosure. One skilled in the relevant art will recognize, however, that the subject matter of the present disclosure can be practiced without one or more of the specific details, or with other methods, components, devices, steps, and so forth. In other instances, well-known methods, devices, implementations, or operations have not been shown or described in detail to avoid obscuring aspects of the disclosure.

The block diagrams shown in the figures are functional entities only and do not necessarily correspond to physically separate entities. I.e. these functional entities may be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor means and/or microcontroller means.

The flow charts shown in the drawings are merely illustrative and do not necessarily include all of the contents and operations/steps, nor do they necessarily have to be performed in the order described. For example, some operations/steps may be decomposed, and some operations/steps may be combined or partially combined, so that the actual execution sequence may be changed according to the actual situation.

In an e-commerce scenario, in order to attract user traffic and attention, operators often formulate marketing campaign content for a portion of commodities and circle groups of people who may be interested in the marketing campaign content.

At present, the main scheme of circle people mainly depends on the past experience of operation one day in advance, and manual selection of labels and circle selection can be carried out on people who are interested in activities. This approach has the following problems:

1. the strategy selection has strong manual dependence, each activity content to be released needs to be manually operated with a label, the interested crowd is released by the strategy selection, and the efficiency is low; especially, when the activity magnitude is large, the corresponding requirements cannot be met based on manual circle selection.

2. The strategy selection has strong dependence on the service experience of operators, but activities of each day are infinite and are very different from one another, each activity and user cannot be accurately and individually mined by utilizing the existing label, and the accuracy of the group to be selected is not high;

3. the current people circling algorithm is based on that a tag is away from a selected part of a circle of possibly interested users one day in advance, on one hand, the tag is generally single, the problem of user omission can be caused, and in addition, the probability that the user visits the app the next day is random, so that the range of active release crowds is further limited.

Therefore, aiming at the problems in the prior art, the disclosure provides a content recommendation method, which utilizes user information and content information to perform modeling based on a double-tower model, automatically represents the total amount of users and contents, and realizes the selection of potential users for active content.

Implementation details of the technical solution of the embodiments of the present disclosure are set forth in detail below.

Fig. 1 schematically illustrates a flow chart of a content recommendation method in an exemplary embodiment of the present disclosure. As shown in fig. 1, the content recommendation method includes steps S1 to S3:

step S1, responding to the access request of the target user, and acquiring the user characteristic vector of the target user;

step S2, performing preference evaluation based on the user characteristic vector of the target user and the content characteristic vector of the candidate content to obtain a preference result of the target user for the candidate content;

and step S3, selecting the preference content from the candidate contents according to the preference result, so as to recommend the preference content to the target user.

In the technical solutions provided by some embodiments of the present disclosure, a user feature vector characterization is used for a user, a content feature vector characterization is used for candidate content, when an access request of a target user is received, preference evaluation can be performed according to the user feature vector of the target user and content feature vectors of all candidate content to obtain a preference result of the target user for each candidate content, and finally, the preference content of the target user is determined according to the preference result for recommendation. On one hand, the content recommendation method provided by the disclosure responds in real time to perform preference evaluation to obtain a recommendation result when receiving an access request, so that manual selection operation is avoided, and the content recommendation efficiency is improved; on the other hand, the user and the content are subjected to vector representation for preference evaluation instead of recommendation by using the existing label, so that the accuracy of content recommendation is improved; on the other hand, content recommendation is carried out according to the accessed target users, the accessed users can be preferentially processed, people groups are prevented from being selected in advance, and the crowd range of content delivery is enlarged.

Hereinafter, each step of the content recommendation method in the present exemplary embodiment will be described in more detail with reference to the drawings and examples.

Step S1, in response to the access request of the target user, obtaining the user feature vector of the target user.

In one embodiment of the present disclosure, in order to recommend preferred content to a target user, user information and content information need to be characterized. According to the method, a double-tower model is adopted to carry out model training in advance, and three submodels, namely a user side submodel, a content measurement model and a preference evaluation submodel are respectively derived after the training is finished, so that the representation vectors are directly and automatically generated for visiting users and all candidate contents respectively, the contents in which the users are most interested are generated, and the problems of low content recommendation efficiency, low precision and incomprehensible coverage of the users are solved.

Therefore, before step S1, the method further comprises step S0: constructing the double-tower model, wherein the constructing the double-tower model comprises the following steps: generating a training set according to historical recommendation data; and carrying out model training through the training set to obtain parameters of an input layer, a representation layer and a matching layer so as to obtain the double-tower model.

Fig. 2 schematically illustrates a structural schematic diagram of a double tower model in an exemplary embodiment of the present disclosure. In an e-commerce scenario, a marketing campaign is typically formulated for a portion of the goods, and content delivery is performed around the population that may be of interest to the goods, so the content candidates may be marketing campaigns related to the goods.

Referring to fig. 2, during model training, Embedding characterization is performed on the user-related information and the activity-related information, and finally preference scoring is performed according to the characterized user Embedding and activity Embedding. The model is divided into three layers, namely an input layer, a representation layer and a matching layer from bottom to top.

When the training set is constructed, user information and activity information of the historical recommendation data, such as the input layer in the double tower model shown in fig. 2, are mainly extracted. The user information mainly comprises attribute information (such as gender, age, purchasing power and the like) of a user, category behavior information (such as item clicking behavior, item searching behavior, item concerned behavior, item adding behavior and item placing behavior) of the user, the activity information can be extracted by formulating an activity detail page of the marketing activity and mainly comprises category information, brand information and attribute information (such as effective date, popular category number and the like) of the activity, a training set is constructed based on marking of the activity by the user, and training is input through a double-tower model structure.

During training, through carrying out feature mapping on input information, finally obtaining user embedding corresponding to user information and camp embedding corresponding to activity content information. The method is completed by a representation layer of a double-tower model, a user information representation and an activity content information representation are divided into two independent towers for feature mapping, a multi-layer NN model is respectively constructed, and finally, a multi-dimensional embedding is output and is respectively used as low-dimensional semantic representations of a user and activity content.

It should be noted that the category behavior information of the user and the category information of the activity are the same category system, and a general three-level category classification method may be adopted, or the classification may be set as needed. Therefore, since the categories of commodities involved in the user behavior and the categories of commodities involved in the activity are in the same system, in order to reduce the training parameters, the category parameters of the user and the activity in the model may be shared parameters.

And finally, the matching layer of the double-tower model evaluates activity preference according to user embedding of the user and camp embedding corresponding to the content, and performs backward propagation to optimize network parameters by comparing with an actual recommendation result.

And after the training is finished, obtaining parameters of the double-tower model after the input layer, the presentation layer and the matching layer are adjusted, and obtaining the double-tower model. The training process is similar to the subsequent use process, and the specific process of the model will be elaborated when in use, and will not be described more extensively here.

In practical application, in order to improve the performance of the online service and reduce the inference pressure of the online service, the model can be respectively derived into three sub-models.

Specifically, an input layer and a presentation layer corresponding to a user tower in the double-tower model are stored as a user side sub-model, namely a user Embedding generating sub-model, and Embedding of the user is generated mainly based on relevant characteristics of the user; storing an input layer and a representation layer corresponding to a content tower in the double-tower model as a content side submodel, namely an activity Embedding generating submodel, and generating the activity Embedding mainly based on the relevant characteristics of the activity; and storing the matching layer in the double-tower model as a preference evaluation submodel, namely a user activity preference scoring submodel, and obtaining the preference degree of the user to the activity mainly based on the user Embedding and the activity Embedding generated by the first two submodels.

In step S1, in response to an access request of a target user, a user feature vector of the target user is acquired.

The target user refers to a user accessing the APP. When a user accesses the APP, the user is used as a target user, an access request is generated and sent to the server, and when the server receives the access request of the target user, the user characteristic vector corresponding to the target user is obtained

Based on the method, each visiting user is evaluated in real time and recommended with the candidate content, the problem of user omission can be avoided, and compared with the traditional recommendation method in which users are selected based on the label circle to release the content, the release crowd range is enlarged.

In an embodiment of the present disclosure, the obtaining the user feature vector of the target user includes: acquiring basic information of the target user; and extracting the user feature vector of the target user from a preset user feature database based on the basic information.

In order to reduce the pressure of server online recommendation, a user feature database can be constructed by characterizing user information in advance, and when a target user visits, a user feature vector of the user is extracted from the database.

Fig. 3 schematically illustrates a flowchart of building a user feature database in an exemplary embodiment of the disclosure. Referring to fig. 3, pre-constructing the user characteristic database includes:

step S31, periodically acquiring the user attribute information and the user behavior information of the historical user;

step S32, based on the user attribute information and the user behavior information, extracting the characteristics of the historical user by using a user side sub-model of a double-tower model to obtain a user characteristic vector of the historical user;

step S33, storing the historical users and the user feature vectors of the historical users to construct the user feature database.

The following describes each step in detail:

in step S31, the user information management module in the data management module may collect and manage the relevant data of the user, and periodically pull the user attribute information and the user behavior information for the full amount of user pool periodically, which may be hourly, daily, or weekly, and may be specifically set according to the requirement, so as to update the user information in real time.

The user attribute information may be portrait attributes of the user, such as attributes of the user's gender, age, geographic location, occupation, and the like; the user behavior information is a sequence of offline behaviors of the user, such as historical click behaviors, search behaviors, focus behaviors, buy-in behaviors, order-out behaviors, and the like of the user.

In step S32, the full-scale user is characterized using the user-side sub-model of the two-tower model. Specifically, the extracting the features of the historical user by using the user-side submodel of the double-tower model based on the user attribute information and the user behavior information to obtain the user feature vector of the historical user includes:

step S321, performing feature processing on the user attribute information to obtain a user attribute feature.

Specifically, the user attribute information generally exists in the form of a character string, and as shown in fig. 2, the user attribute information such as the gender, age, geographical position, and occupation of the user is subjected to feature mapping by the Embedding Layer of the user-side submodel input Layer, and the character string is connected to form a feature vector, which is recorded as a user attribute feature as u_p。

Step S322, performing feature conversion on the user behavior information to obtain user content features.

In an embodiment of the present disclosure, the performing feature conversion on the user behavior information to obtain a user content feature includes: performing feature mapping on the user behavior information to obtain a content category vector based on the content category information corresponding to the user behavior information; and carrying out mean processing on the content category vectors to obtain the user content characteristics.

The user behavior information is an offline behavior sequence of the user, each behavior corresponds to commodity information, such as a click item type behavior, a search item type behavior, an attention item type behavior, an additional purchase item type behavior and an order type behavior, and the user behaviors are subjected to feature processing and are marked as u_l,bThe method is mainly used for extracting the behavior interest of the user.

After the behavior characteristics of the user are obtained, the content class corresponding to the user behavior is subjected to feature mapping based on the Embedding Layer, content class Embedding vectors of the behavior characteristics of the user are generated, and finally, average-value processing is carried out on the Embedding vectors to obtain abstract representation of the user behavior, namely the user content characteristics. This is because the categories of goods involved in user behavior are unordered and vectors can be averaged.

It should be noted that, in the present disclosure, the execution steps of step S321 and step S322 are not limited, the user attribute feature and the user content feature are obtained in no order, and the user attribute feature may be obtained by processing first, or the user content feature may be obtained first, or certainly, the user attribute feature and the user content feature may be processed simultaneously.

And step S323, performing feature splicing on the user attribute features and the user content features to obtain user feature vectors of the historical users.

In one embodiment of the present disclosure, the process of feature splicing the user attribute features and the user content features is denoted as f_u. Converting the spliced features through multi-layer full connection at the user side to obtain the user' S Embedding, which is recorded as S_uNamely:

S_u＝f_u(u_p,u_l,b) (1)

wherein u is_pAs a user attribute feature, u_l,bIs the behavior characteristic of the user.

The user feature vector used for representing the user information is obtained by splicing the user attribute features and the user content features, wherein the user attribute features and the content information designed in the user behavior are included, and the basic attribute and the behavior preference of the user are subjected to feature mapping.

In step S33, the total number of historical users and the calculated user feature vector of the historical user are stored in Redis, and a user feature database is constructed, and an identifier is added for each user, so as to search the database for a matching historical user by the identifier.

By regularly pulling the user information and representing the user information to construct the user characteristic database, the user characteristic vector of the target user can be directly obtained when the access request of the target user is received, the pressure of an online recommendation server is reduced, and the recommendation efficiency is improved.

Step S2, performing preference evaluation based on the user feature vector of the target user and the content feature vector of the candidate content, to obtain a preference result of the target user for the candidate content.

In one embodiment of the present disclosure, the candidate content may also be characterized in advance, thereby reducing the server pressure of online real-time recommendation. Thus, the method further comprises: and extracting the content feature vector of the candidate content from a preset content feature database.

Fig. 4 schematically illustrates a flow chart of building a content feature database in an exemplary embodiment of the disclosure. Referring to fig. 4, the pre-constructing the content feature database includes:

step S41, acquiring content information of the candidate content;

step S42, extracting the characteristics of the candidate content by using a content side sub-model of a double-tower model based on the content information to obtain a content characteristic vector of the candidate content;

step S43, storing the candidate content and the content feature vector of the candidate content to construct the content feature database.

In step S41, specifically, the content information includes content item information, content brand information, and content attribute information.

Taking the candidate content as an example of recommending the activity to the user, the content information refers to information such as the category, brand, and attribute of the product corresponding to the activity. Since business creates new activities on an irregular basis every day, in order to support the demand, all the category, brand, and attribute information of the activity detail page can be extracted when the business reports new activity contents.

Parsed content data may be collected by a content information management module in the data management module in preparation for subsequent user characterization and content characterization. The system is used for receiving and analyzing the information of the activity class, the brand, the start-stop time and the like of the service newspaper in real time.

In step S42, the performing feature extraction on the candidate content by using a content-side sub-model of a double-tower model based on the content information to obtain a content feature vector of the candidate content includes:

step S421, performing feature mapping on the content category information to obtain a content category vector, and performing weighting processing on the content category vector to obtain content category features;

step S422, performing feature mapping on the content brand information to obtain a content brand vector, and performing weighting processing on the content brand vector to obtain content brand features; and step S423, performing characteristic processing on the content attribute information to obtain content attribute characteristics;

in an embodiment of the present disclosure, when representing content information, according to different categories of the content information, feature mapping is performed on the content category information, the content brand information, and the content attribute information by using corresponding Embedding layers to obtain corresponding Embedding vectors, which are denoted as c_i。

Referring to fig. 2, after the activity attribute information is mapped to a character string, the character string may be spliced to obtain the content attribute feature. The commodity types or brand information related to the activities are different, in order to accurately represent the activities, weighted-posing weighting processing needs to be carried out on vectors obtained by mapping, different commodities in the same activity have different weights, content representation is more accurate, and the accuracy of content recommendation can be further improved.

It should be noted that the present disclosure does not limit the execution sequence of step S421 to step S423, and may be executed simultaneously or separately.

Step S424, performing feature splicing on the content category features, the content brand features, and the content attribute features to obtain content feature vectors of the candidate content.

Performing characteristic splicing on the content category characteristics and the content brand characteristics and the content attribute characteristics, performing multilayer full-connection conversion on spliced characteristic vectors at the active side, and recording the process as f_aObtaining the Embedding of the content, denoted as S_aNamely:

wherein, w_iIs the weight of the ith item of content, c_iIs the feature vector of the ith item of content, and a is the total number of content.

In step S43, the entire amount of candidate contents and the calculated content feature vector of the content are stored in Redis, and a content feature database is constructed. And new content information is acquired and represented when the content is updated, so that the real-time accuracy of content representation can be ensured, the pressure of an online recommendation server can be reduced, and the recommendation efficiency is improved.

In one embodiment of the disclosure, after determining the user feature vector of the target user and acquiring the content feature vector of the full amount of candidate content, preference evaluation is performed based on the user feature vector and the content feature vector.

Therefore, the performing preference evaluation based on the user feature vector of the target user and the content feature vector of the candidate content to obtain a preference result of the target user for the candidate content includes: determining a target content from the candidate contents; and based on the user characteristic vector of the target user and the content characteristic vector of the target content, performing preference evaluation by using a preference evaluation sub-model of the double-tower model to obtain a preference result.

Wherein, the preference evaluation is performed by using a preference evaluation submodel of a double-tower model based on the user feature vector of the target user and the content feature vector of the target content to obtain the preference result, and the preference evaluation comprises: calculating the cross feature of the content feature vector and the user feature vector by using a multi-layer perceptron; determining a preference value of the target content as the preference result based on the cross feature.

Specifically, the current real-time visiting user's embed and all active embed in the activity pool are respectively obtained by querying from redis, and the score evaluation of 1v1 is performed. Therefore, when performing preference evaluation, a target content needs to be selected, and the user embedding of the target user and the camp embedding corresponding to the target content are subjected to preference evaluation, so as to finally obtain a preference result of the target user for the target content.

In view of the fact that the double-tower model used in the method is not easy to introduce user-activity cross features, the method is high in real-time requirement on the line, time delay cannot be too large, and measurement functions such as dot product and cosine are often used on the line in the past to be not satisfactory, therefore, the preference evaluation part of the method adopts a light MLP structure, can introduce higher-order cross features on the premise of ensuring calculation time, and can accurately capture preference factors of users on activities, such as x²、y²、x²y, and so on. And the user Embedding and the content Embedding are ensured to be in the same characteristic space through an end-to-end training mode. Calculating the score through MLP and recording as S_u,aThe process is denoted as f_u,aNamely:

S_u,a＝f_u,a(S_u,S_a) (3)

In an embodiment of the present disclosure, the selecting, according to the preference result, preferred content from the candidate content for recommending the preferred content to the target user includes: sorting the candidate contents according to the preference value of each candidate content in the preference result to obtain a sorting result; selecting preference content based on the sorting result, and recommending the preference content to the target user.

When preference evaluation is carried out, preference scores S of the target user to each candidate content are obtained_u,aWhen recommending, the preference scores of all candidate activities can be sorted, and the content with the highest score is selected as the final result to be returned to the user, and is marked as A_uNamely:

fig. 5 schematically illustrates a flow chart of a content recommendation method in an exemplary embodiment of the present disclosure. Referring to fig. 5, the content recommendation method is mainly divided into three parts, namely, an activity representation, an activity recommendation and a user representation.

For the activity characterization part, 510 the activity information management database collects real-time activity information; then, step S501 is executed to analyze the category of the active brand, that is, to analyze the active attributes of the service newspaper, such as the category of the active brand, the start-stop time, and the like; and then, executing a step S502, representing the activity information by using a double-tower model to obtain an activity Embedding vector, and storing the activity Embedding vector into a 520 activity redis database.

For the user characterization part, 580 a user information management database collects and manages related data of users, the data mainly comprises user attribute information 570 and historical behavior information 580 generated by the users on the platform, and 580 the user information management database is also responsible for collecting real-time activity information for subsequent characterization of user behaviors; then, the user behavior information 560, the user attribute information 570, the user 512 and the user 511 are subjected to feature processing through the steps S510 and S509 to convert the information into user feature data which can be read by a computer, so that the user feature data 550 is obtained and stored in the offline database 540; and then executing step S508, performing user characterization on the user feature data of each user in the offline database to obtain a user Embedding vector, and storing the user Embedding vector into 530 a user redis database.

For the activity recommending part, the 590 activity filtering module is configured to receive an access request from a user, then perform step S504 and step S506, respectively, that is, obtain an activity Embedding vector from 520 activity redis, and obtain a user Embedding vector from 530 user redis, then perform step S505, perform preference evaluation using the user Embedding vector and the activity Embedding vector, return the result to the 590 activity filtering module, and finally determine an activity with a preference value Top to recommend to the user.

Based on the method, the method respectively derives three sub-models by combining the portrait attribute and the category behavior sequence of the user and the category attribute of the activity: the user representation model, the activity representation model and the user-activity preference real-time scoring model are used for realizing the purpose of automatically generating the representation vectors for visiting users and all candidate activities respectively, generating the activities of most interest of the users and solving the problems of low efficiency, low precision and incomprehensive coverage of the users.

Fig. 6 schematically illustrates a composition diagram of a content recommendation apparatus in an exemplary embodiment of the disclosure, and as shown in fig. 6, the content recommendation apparatus 600 may include a response module 601, an evaluation module 602, and a recommendation module 603. Wherein:

a response module 601, configured to respond to an access request of a target user, and obtain a user feature vector of the target user;

an evaluation module 602, configured to perform preference evaluation based on the user feature vector of the target user and the content feature vector of the candidate content, to obtain a preference result of the target user for the candidate content;

a recommending module 603, configured to select a preferred content from the candidate contents according to the preference result, so as to recommend the preferred content to the target user.

According to an exemplary embodiment of the present disclosure, the response module 601 includes a subscriber unit, which is configured to obtain basic information of the target user; and extracting the user feature vector of the target user from a preset user feature database based on the basic information.

According to an exemplary embodiment of the present disclosure, the content recommendation device 600 further includes a user characteristic module (not shown in the figure) for pre-constructing the user characteristic database, including: periodically acquiring user attribute information and user behavior information of historical users; based on the user attribute information and the user behavior information, performing feature extraction on the historical user by using a user side sub-model of a double-tower model to obtain a user feature vector of the historical user; and storing the historical users and the user feature vectors of the historical users to construct the user feature database.

According to an exemplary embodiment of the present disclosure, the content recommendation apparatus 600 is further configured to perform feature processing on the user attribute information to obtain a user attribute feature; performing characteristic conversion on the user behavior information to obtain user content characteristics; and performing feature splicing on the user attribute features and the user content features to obtain the user feature vector of the historical user.

According to an exemplary embodiment of the present disclosure, the content recommendation apparatus 600 is further configured to perform feature mapping on the user behavior information based on content category information corresponding to the user behavior information to obtain a content category vector; and carrying out mean processing on the content category vectors to obtain the user content characteristics.

According to an exemplary embodiment of the present disclosure, the evaluation module 602 includes a content unit, and the content unit is configured to extract a content feature vector of the candidate content from a preset content feature database.

According to an exemplary embodiment of the present disclosure, the content recommendation device 600 further includes a content feature module (not shown in the figure) for obtaining content information of the candidate content; performing feature extraction on the candidate content by using a content side sub-model of a double-tower model based on the content information to obtain a content feature vector of the candidate content; storing the candidate content and the content feature vector of the candidate content to construct the content feature database.

According to an exemplary embodiment of the present disclosure, the content information includes content item information, content brand information, and content attribute information; the content feature module is further configured to perform feature mapping on the content category information to obtain a content category vector, and perform weighting processing on the content category vector to obtain content category features; performing feature mapping on the content brand information to obtain a content brand vector, and performing weighting processing on the content brand vector to obtain content brand features; and carrying out feature processing on the content attribute information to obtain content attribute features; and performing feature splicing on the content category features, the content brand features and the content attribute features to obtain content feature vectors of the candidate content.

According to an exemplary embodiment of the disclosure, the evaluation module 602 further includes an evaluation unit, configured to determine a target content from the candidate contents; and based on the user characteristic vector of the target user and the content characteristic vector of the target content, performing preference evaluation by using a preference evaluation sub-model of the double-tower model to obtain a preference result.

According to an exemplary embodiment of the present disclosure, the evaluation unit is further configured to calculate cross features of the content feature vector and the user feature vector using a multi-layer perceptron; determining a preference value of the target content as the preference result based on the cross feature.

According to an exemplary embodiment of the present disclosure, the recommending module 603 is configured to rank the candidate contents according to preference values of the candidate contents in the preference result to obtain a ranking result; selecting preference content based on the sorting result, and recommending the preference content to the target user.

According to an exemplary embodiment of the present disclosure, the content recommendation apparatus 600 further includes a model training module (not shown in the figure) for generating a training set according to the historical recommendation data; and carrying out model training through the training set to obtain parameters of an input layer, a representation layer and a matching layer so as to obtain the double-tower model.

According to an exemplary embodiment of the present disclosure, the model training module is further configured to store an input layer and a representation layer corresponding to a user tower in the double-tower model as a user-side sub-model; storing an input layer and a representation layer corresponding to a content tower in the double-tower model as a content side sub-model; and storing the matching layer in the double-tower model as a preference evaluation submodel.

The details of each module in the content recommendation apparatus 600 are already described in detail in the corresponding content recommendation method, and therefore are not described herein again.

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

In an exemplary embodiment of the present disclosure, there is also provided a storage medium capable of implementing the above-described method. Fig. 7 schematically illustrates a schematic diagram of a computer-readable storage medium in an exemplary embodiment of the disclosure. As shown in fig. 7, a program product 700 for implementing the above method according to an embodiment of the present disclosure is described, which may employ a portable compact disc read only memory (CD-ROM) and include program code, and may be run on a terminal device, such as a cell phone. However, the program product of the present disclosure is not limited thereto, and in this document, a readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

In an exemplary embodiment of the present disclosure, an electronic device capable of implementing the above method is also provided. Fig. 8 schematically shows a structural diagram of a computer system of an electronic device in an exemplary embodiment of the disclosure.

It should be noted that the computer system 800 of the electronic device shown in fig. 8 is only an example, and should not bring any limitation to the functions and the scope of the application of the embodiments of the present disclosure.

As shown in fig. 8, a computer system 800 includes a Central Processing Unit (CPU)801 that can perform various appropriate actions and processes according to a program stored in a Read-Only Memory (ROM) 802 or a program loaded from a storage section 808 into a Random Access Memory (RAM) 803. In the RAM 803, various programs and data necessary for system operation are also stored. The CPU 801, ROM 802, and RAM 803 are connected to each other via a bus 804. An Input/Output (I/O) interface 805 is also connected to bus 804.

The following components are connected to the I/O interface 805: an input portion 806 including a keyboard, a mouse, and the like; an output section 807 including a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), and a speaker; a storage portion 808 including a hard disk and the like; and a communication section 809 including a Network interface card such as a LAN (Local Area Network) card, a modem, or the like. The communication section 809 performs communication processing via a network such as the internet. A drive 810 is also connected to the I/O interface 805 as necessary. A removable medium 811 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is mounted on the drive 810 as necessary, so that a computer program read out therefrom is mounted on the storage section 808 as necessary.

In particular, the processes described below with reference to the flowcharts may be implemented as computer software programs, according to embodiments of the present disclosure. For example, embodiments of the present disclosure include a computer program product comprising a computer program embodied on a computer readable medium, the computer program comprising program code for performing the method illustrated in the flow chart. In such an embodiment, the computer program can be downloaded and installed from a network through the communication section 809 and/or installed from the removable medium 811. When the computer program is executed by a Central Processing Unit (CPU)801, various functions defined in the system of the present disclosure are executed.

It should be noted that the computer readable medium shown in the embodiments of the present disclosure may be a computer readable signal medium or a computer readable storage medium or any combination of the two. A computer readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples of the computer readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a Read-Only Memory (ROM), an Erasable Programmable Read-Only Memory (EPROM), a flash Memory, an optical fiber, a portable Compact Disc Read-Only Memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present disclosure, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. In contrast, in the present disclosure, a computer-readable signal medium may include a propagated data signal with computer-readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wired, etc., or any suitable combination of the foregoing.

The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams or flowchart illustration, and combinations of blocks in the block diagrams or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

The units described in the embodiments of the present disclosure may be implemented by software, or may be implemented by hardware, and the described units may also be disposed in a processor. Wherein the names of the elements do not in some way constitute a limitation on the elements themselves.

As another aspect, the present disclosure also provides a computer-readable medium, which may be contained in the electronic device described in the above embodiments; or may exist separately without being assembled into the electronic device. The computer readable medium carries one or more programs which, when executed by an electronic device, cause the electronic device to implement the method described in the above embodiments.

Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein may be implemented by software, or by software in combination with necessary hardware. Therefore, the technical solution according to the embodiments of the present disclosure may be embodied in the form of a software product, which may be stored in a non-volatile storage medium (which may be a CD-ROM, a usb disk, a removable hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which may be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiments of the present disclosure.

Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of the disclosure following, in general, the principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains.

It will be understood that the present disclosure is not limited to the precise arrangements described above and shown in the 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.

Claims

1. A content recommendation method, comprising:

responding to an access request of a target user, and acquiring a user characteristic vector of the target user;

performing preference evaluation based on the user characteristic vector of the target user and the content characteristic vector of the candidate content to obtain a preference result of the target user on the candidate content;

selecting preference content from the candidate content according to the preference result so as to recommend the preference content to the target user.

2. The content recommendation method according to claim 1, wherein said obtaining the user feature vector of the target user comprises:

acquiring basic information of the target user;

and extracting the user feature vector of the target user from a preset user feature database based on the basic information.

3. The content recommendation method according to claim 2, further comprising pre-constructing the user characteristic database, the pre-constructing the user characteristic database comprising:

periodically acquiring user attribute information and user behavior information of historical users;

based on the user attribute information and the user behavior information, performing feature extraction on the historical user by using a user side sub-model of a double-tower model to obtain a user feature vector of the historical user;

and storing the historical users and the user feature vectors of the historical users to construct the user feature database.

4. The content recommendation method according to claim 3, wherein said performing feature extraction on the historical user by using a user-side sub-model of a double-tower model based on the user attribute information and the user behavior information to obtain a user feature vector of the historical user comprises:

performing feature processing on the user attribute information to obtain user attribute features; and

performing characteristic conversion on the user behavior information to obtain user content characteristics;

and performing feature splicing on the user attribute features and the user content features to obtain the user feature vector of the historical user.

5. The content recommendation method according to claim 4, wherein said performing feature transformation on the user behavior information to obtain user content features comprises:

performing feature mapping on the user behavior information to obtain a content category vector based on the content category information corresponding to the user behavior information;

and carrying out mean processing on the content category vectors to obtain the user content characteristics.

6. The content recommendation method according to claim 1, further comprising:

and extracting the content feature vector of the candidate content from a preset content feature database.

7. The content recommendation method according to claim 6, further comprising pre-building the content feature database, including:

acquiring content information of the candidate content;

performing feature extraction on the candidate content by using a content side sub-model of a double-tower model based on the content information to obtain a content feature vector of the candidate content;

storing the candidate content and the content feature vector of the candidate content to construct the content feature database.

8. The content recommendation method according to claim 7, wherein the content information includes content category information, content brand information, and content attribute information;

the performing feature extraction on the candidate content by using a content side sub-model of a double-tower model based on the content information to obtain a content feature vector of the candidate content includes:

performing feature mapping on the content category information to obtain a content category vector, and performing weighting processing on the content category vector to obtain content category features;

performing feature mapping on the content brand information to obtain a content brand vector, and performing weighting processing on the content brand vector to obtain content brand features; and

performing feature processing on the content attribute information to obtain content attribute features;

and performing feature splicing on the content category features, the content brand features and the content attribute features to obtain content feature vectors of the candidate content.

9. The content recommendation method according to claim 1, wherein performing preference evaluation based on the user feature vector of the target user and the content feature vector of the candidate content to obtain a preference result of the target user for the candidate content comprises:

determining a target content from the candidate contents;

and based on the user characteristic vector of the target user and the content characteristic vector of the target content, performing preference evaluation by using a preference evaluation sub-model of the double-tower model to obtain a preference result.

10. The content recommendation method according to claim 9, wherein the performing preference evaluation using a preference evaluation submodel of a two-tower model based on the user feature vector of the target user and the content feature vector of the target content to obtain the preference result comprises:

calculating the cross feature of the content feature vector and the user feature vector by using a multi-layer perceptron;

determining a preference value of the target content as the preference result based on the cross feature.

11. The content recommendation method according to claim 1, wherein said selecting preferred content from the candidate content according to the preference result for recommending the preferred content to the target user comprises:

sorting the candidate contents according to the preference value of each candidate content in the preference result to obtain a sorting result;

selecting preference content based on the sorting result, and recommending the preference content to the target user.

12. The content recommendation method according to any one of claims 3, 7 and 9, further comprising constructing the double tower model, wherein constructing the double tower model comprises:

generating a training set according to historical recommendation data;

and carrying out model training through the training set to obtain parameters of an input layer, a representation layer and a matching layer so as to obtain the double-tower model.

13. The content recommendation method according to claim 12, characterized in that the method further comprises:

storing an input layer and a representation layer corresponding to a user tower in the double-tower model as a user side sub-model;

storing an input layer and a representation layer corresponding to a content tower in the double-tower model as a content side sub-model; and

14. A content recommendation apparatus characterized by comprising:

the response module is used for responding to an access request of a target user and acquiring a user characteristic vector of the target user;

the evaluation module is used for carrying out preference evaluation on the basis of the user characteristic vector of the target user and the content characteristic vector of the candidate content to obtain a preference result of the target user on the candidate content;

and the recommending module is used for selecting the preference content from the candidate contents according to the preference result so as to recommend the preference content to the target user.

15. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out a content recommendation method according to any one of claims 1 to 13.

16. An electronic device, comprising:

one or more processors;

storage means for storing one or more programs which, when executed by the one or more processors, cause the one or more processors to implement the content recommendation method according to any one of claims 1 to 13.