Content recommendation device, electronic device and computer readable medium

Through the in-depth recommendation model and multi-task learning model, the problem that users find it difficult to obtain content of interest in massive information is solved, personalized recommendation is achieved, and user experience and information acquisition efficiency is improved.

CN114491235BActive Publication Date: 2024-11-26BEIJING LIDA ZHISHENG TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111631506.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-14
Publication Date
2024-11-26
Estimated Expiration
2042-03-14

AI Technical Summary

Technical Problem

In the use scenario of mobile Internet terminals, it is difficult for users to obtain content of interest from massive information. Traditional search engines require users to actively enter keywords, resulting in poor information acquisition efficiency and user experience.

Method used

A content recommendation device is designed to obtain user information through the information module, the primary selection module filters content, the feature module generates user and content features, the scoring module uses the in-depth recommendation model and the multi-task learning model to calculate interest scores, and the recommendation module sorts recommendations based on the score. content.

Benefits of technology

Multi-angle analysis of user interests is realized, the accuracy and user satisfaction of recommended content are improved, and recommendation efficiency is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114491235B_ABST
    Figure CN114491235B_ABST
Patent Text Reader

Abstract

The present application relates to a content recommendation device, an electronic device and a computer-readable medium. The device includes: an information module, which is used to obtain user information based on the user's recommendation request; a preliminary selection module, which is used to obtain multiple preliminary selected contents based on the user information; a feature module, which is used to generate user features, request features, and multiple content features based on the user information, the recommendation request, and the multiple preliminary selected contents; a scoring module, which is used to input the user features, the request features, and the multiple content features into an interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents, and the interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model; a recommendation module, which is used to determine at least one content to be recommended from the multiple preliminary selected contents based on the multiple interest scores. The content recommendation device involved in the present application can analyze the user's interest from multiple angles, and comprehensively generate recommended content based on various information, thereby improving recommendation efficiency and user satisfaction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of computer information processing, and in particular to a content recommendation device, an electronic device, and a computer-readable medium. Background Art

[0002] At present, with the rapid development of computers, the peak daily upload volume of various multimedia data (such as graphic data and video data) from various source channels has exceeded millions or even tens of millions. Especially with the popularization and development of mobile devices, while the amount of information is becoming more and more abundant, users' attention is also overly dispersed, making it difficult to obtain effective information. At the same time, considering the usage scenarios of mobile Internet terminals, search engines and other active search methods that require users to enter keywords are less applicable. It is urgent to solve the problem of the system automatically obtaining user interests and completing push to obtain a good user experience.

[0003] The recommendation system meets the above needs. The recommendation system is an interactive product function. The product provides a carrier for the recommendation system. Users reach and trigger the recommendation system by using the product. The recommendation system provides users with personalized recommendations, thereby improving the user experience. Unlike search engines, the recommendation system is essentially a technical means to help users find information of interest from a large amount of information when user needs are unclear. The recommendation system combines user information and content information, uses machine learning technology to build a user interest model, and provides users with accurate personalized recommendations.

[0004] With the development of deep learning, deep learning has also been greatly used in the field of recommendation systems. A series of recommendation algorithms based on deep learning have gradually come out. In addition to being used in shopping software, recommendation systems are also widely used in scenarios such as information recommendation in news apps and video recommendation in video software.

[0005] The above information disclosed in the Background section is only for enhancement of understanding of the background of the present application and therefore it may contain information that does not constitute the prior art that is already known to a person of ordinary skill in the art. Summary of the invention

[0006] The present application provides a content recommendation device, an electronic device, and a computer-readable medium, which can analyze the user's interest from multiple angles, and comprehensively generate recommended content based on various information, thereby improving recommendation efficiency and user satisfaction.

[0007] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present application.

[0008] According to one aspect of the present application, a content recommendation device is proposed, which includes: an information module, which is used to obtain user information based on a user's recommendation request; a preliminary selection module, which is used to obtain multiple preliminary selected contents based on the user information; a feature module, which is used to generate user features, request features, and multiple content features based on the user information, the recommendation request, and the multiple preliminary selected contents, respectively; a scoring module, which is used to input the user features, the request features, and the multiple content features into an interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents, and the interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model; a recommendation module, which is used to determine at least one content to be recommended from the multiple preliminary selected contents based on the multiple interest scores.

[0009] In an exemplary embodiment of the present application, it also includes: a training module, used to generate an initial model through a deep recommendation model and a multi-task learning model; training the initial model based on multiple user recommendation requests and multiple content recommendation results to generate the interest scoring model.

[0010] In an exemplary embodiment of the present application, obtaining multiple preliminary selected contents based on the user information includes: extracting a pre-stored user portrait and a user vector based on a user identifier in the user information; extracting other users based on the user portrait to generate a list of other users; and generating the multiple preliminary selected contents based on the user portrait, the user vector, and the list of other users.

[0011] In an exemplary embodiment of the present application, generating the plurality of preliminary selected contents based on the user portrait, the user vector, and the other user list further includes: performing exposure and deduplication on the plurality of preliminary selected contents based on the user information.

[0012] In an exemplary embodiment of the present application, user characteristics, request characteristics, and multiple content characteristics are generated based on the user information, the recommendation request, and the multiple preliminary selected contents, respectively, including: organizing the user information, the recommendation request, and the multiple preliminary selected contents according to preset strategies to generate the user characteristics, the request characteristics, and the multiple content characteristics.

[0013] In an exemplary embodiment of the present application, the user features, the request features, and the multiple content features are input into an interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents, including: inputting the user features, the request features, and the multiple content features into the interest scoring model; the deep recommendation model part in the interest scoring model generates a combined feature based on the user features, the request features, and the multiple content features; the multi-task learning model part in the interest scoring model performs calculations based on the combined feature to generate multiple interest scores corresponding to the multiple content features; wherein the interest score corresponding to the content feature represents the user's interest level in the preliminary selected contents.

[0014] The multi-task learning model part in the interest scoring model performs calculations based on the combined features to generate interest scores corresponding to the multiple content features, including: the multi-task learning model part in the interest scoring model performs calculations based on the combined features to generate click probability and propagation probability; and generates interest scores corresponding to the multiple content features based on the click probability, the propagation probability and their corresponding weights.

[0015] The multiple interest scores are used to determine at least one content to be recommended from the multiple preliminary selected contents, including: sorting the multiple preliminary selected contents based on the multiple interest scores; and extracting at least one preliminary selected content as the content to be recommended according to the sorting.

[0016] The interest scoring model is trained and generated on the initial model based on multiple user recommendation requests and multiple content recommendation results, including: generating a training sample set based on multiple user recommendation requests; setting sample labels for samples in the training sample set based on the content recommendation results; training the initial model based on the training sample set data with sample labels; and generating the interest scoring model when the training parameters meet a threshold.

[0017] The recommendation module includes a deep recommendation submodule and a multi-task learning submodule, and the output data of the deep recommendation submodule is used as the input data of the multi-task learning submodule.

[0018] Preferably, the DNN hidden layer in the deep recommendation submodule is set to three layers, each layer has a dropout factor; the expert network of the multi-task learning submodule is set to two, and the expert units are set to sixteen; the DNN hidden layer in the expert network of the multi-task learning submodule is set to two layers; at the same time, the initial model of the recommendation module is generated based on the above settings.

[0019] According to the content recommendation device, electronic device and computer-readable medium of the present application, user information is obtained based on a recommendation request of the user; multiple preliminary selected contents are obtained based on the user information; user features, request features and multiple content features are respectively generated based on the user information, the recommendation request and the multiple preliminary selected contents; the user features, the request features and the multiple content features are input into an interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents, and the interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model; at least one content to be recommended is determined from the multiple preliminary selected contents based on the multiple interest scores, so that the user's interest can be analyzed from multiple angles, and recommended content can be comprehensively generated by combining various information, thereby improving recommendation efficiency and user satisfaction.

[0020] It should be understood that the foregoing general description and the following detailed description are exemplary only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and other objects, features and advantages of the present application will become more apparent by describing in detail the exemplary embodiments thereof with reference to the accompanying drawings. The accompanying drawings described below are only some embodiments of the present application, and it is clear to a person skilled in the art that other accompanying drawings can be obtained from these accompanying drawings without creative effort.

[0022] Figure 1 The present invention is a system block diagram of a content recommendation device according to an exemplary embodiment.

[0023] Figure 2 The present invention is a block diagram of a content recommendation device according to an exemplary embodiment.

[0024] Figure 3 The figure is a flowchart showing a working process of a content recommendation device according to an exemplary embodiment.

[0025] Figure 4 It is a schematic diagram showing the working process of a content recommendation device according to an exemplary embodiment.

[0026] Figure 5 The figure is a flowchart showing a working process of a content recommendation device according to another exemplary embodiment.

[0027] Figure 6 The figure is a schematic diagram showing a working process of a content recommendation device according to another exemplary embodiment.

[0028] Figure 7 It is a block diagram of an electronic device according to an exemplary embodiment.

[0029] Figure 8 It is a block diagram of a computer-readable medium according to an exemplary embodiment. DETAILED DESCRIPTION

[0030] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this application will be comprehensive and complete and will fully convey the concepts of the example embodiments to those skilled in the art. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted.

[0031] In addition, described feature, structure or characteristic can be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present application. However, those skilled in the art will appreciate that the technical scheme of the present application can be put into practice without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, known methods, devices, realizations or operations are not shown or described in detail to avoid blurring the various aspects of the application.

[0032] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0033] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0034] It should be understood that although the terms first, second, third, etc. may be used herein to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another component. Therefore, the first component discussed below can be referred to as the second component without departing from the teachings of the concepts of the present application. As used herein, the term "and / or" includes any one of the associated listed items and all combinations of one or more.

[0035] Those skilled in the art will appreciate that the drawings are merely schematic diagrams of example embodiments, and the modules or processes in the drawings are not necessarily necessary for implementing the present application, and therefore cannot be used to limit the scope of protection of the present application.

[0036] Figure 1The present invention is a system block diagram showing a content recommendation method and device according to an exemplary embodiment.

[0037] like Figure 1 As shown, the system architecture 10 may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0038] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, 103, such as news applications, live video applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0039] The terminal devices 101 , 102 , and 103 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers.

[0040] The server 105 may be a server that provides various services, such as a background management server that provides support for news websites browsed by users using the terminal devices 101, 102, and 103. The background management server may analyze and process the recommendation request from the user, and feed back the processing result (such as at least one content to be recommended) to the terminal device.

[0041] The server 105 may, for example, generate an initial model through a deep recommendation model and a multi-task learning model; the server 105 may, for example, train the initial model based on multiple user recommendation requests and multiple content recommendation results to generate the interest scoring model.

[0042] The server 105 may, for example, send the interest scoring model to the terminal devices 101, 102, 103 so as to perform model calculations on the terminal devices 101, 102, 103. The server 105 may also, for example, perform model calculations on the service side, but the present application is not limited thereto.

[0043] The server 105 may, for example, obtain a user's recommendation request from the terminal devices 101, 102, 103, and obtain user information based on the recommendation request; the server 105 may, for example, obtain multiple preliminary selected contents based on the user information; the server 105 or the terminal devices 101, 102, 103 may, for example, respectively generate user features, request features, and multiple content features based on the user information, the recommendation request, and the multiple preliminary selected contents; the server 105 or the terminal devices 101, 102, 103 may, for example, input the user features, the request features, and the multiple content features into an interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents, and the interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model; the server 105 or the terminal devices 101, 102, 103 may, for example, determine at least one content to be recommended from the multiple preliminary selected contents based on the multiple interest scores.

[0044] Server 105 can be a physical server, or can be composed of multiple servers, for example. A part of server 105 can be used as the service system in this application, for receiving user recommendation requests; and a part of server 105 can also be used as the recommendation system in this application, for generating multiple recommended contents based on user information.

[0045] It should be noted that the content recommendation method provided in the embodiment of the present application can be executed by the server 105 and / or the terminal devices 101, 102, 103, and accordingly, the content recommendation device can be set in the server 105 and / or the terminal devices 101, 102, 103. The web page provided for users to browse news is generally located in the terminal devices 101, 102, 103.

[0046] The following are embodiments of the device of the present application.

[0047] Figure 2 FIG. 1 is a block diagram of a content recommendation device according to another exemplary embodiment. Figure 2 As shown, the content recommendation device 60 includes: an information module 602 , a preliminary selection module 604 , a feature module 606 , a scoring module 608 , and a recommendation module 610 . The content recommendation device 60 may further include: a training module 612 .

[0048] The information module 602 is used to obtain user information based on the user's recommendation request;

[0049] The preliminary selection module 604 is used to obtain multiple preliminary selection contents based on the user information; the preliminary selection module 604 is also used to extract pre-stored user portraits and user vectors based on the user identifier in the user information; extract other users based on the user portrait to generate a list of other users; and generate the multiple preliminary selection contents based on the user portrait, the user vector, and the list of other users.

[0050] The feature module 606 is used to generate user features, request features, and multiple content features based on the user information, the recommendation request, and the multiple preliminary selected contents; the feature module 606 is also used to organize the user information, the recommendation request, and the multiple preliminary selected contents according to preset strategies to generate the user features, the request features, and the multiple content features.

[0051] The scoring module 608 is used to input the user characteristics, the request characteristics, and the multiple content characteristics into the interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents, and the interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model; the scoring module 608 is also used to input the user characteristics, the request characteristics, and the multiple content characteristics into the interest scoring model; the deep recommendation model part in the interest scoring model generates a combined feature based on the user characteristics, the request characteristics, and the multiple content characteristics; the multi-task learning model part in the interest scoring model performs calculations based on the combined features to generate multiple interest scores corresponding to the multiple content characteristics.

[0052] The recommendation module 610 is used to determine at least one content to be recommended from the multiple preliminary selected contents based on the multiple interest scores. The recommendation module 610 is also used to sort the multiple preliminary selected contents based on the multiple interest scores; and extract at least one preliminary selected content as the content to be recommended according to the sorting.

[0053] The training module 612 is used to generate an initial model through a deep recommendation model and a multi-task learning model; the initial model is trained based on multiple user recommendation requests and multiple content recommendation results to generate the interest scoring model.

[0054] According to the content recommendation device of the present application, user information is obtained based on a recommendation request of the user; multiple preliminary selected contents are obtained based on the user information; user features, request features, and multiple content features are respectively generated based on the user information, the recommendation request, and the multiple preliminary selected contents; the user features, the request features, and the multiple content features are input into an interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents, and the interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model; based on the multiple interest scores, at least one content to be recommended is determined from the multiple preliminary selected contents, so that the user's interest can be analyzed from multiple angles, and recommended content can be comprehensively generated by combining various information, thereby improving recommendation efficiency and user satisfaction.

[0055] The following combination Figure 3-Figure 6 The specific working mode of the device described in this application is described in detail. It should be pointed out that Figure 3-Figure 6 The method shown is only a typical implementation of the present application and does not represent a specific restriction on the protection scope of the present application.

[0056] Figure 3 2 is a flowchart showing a typical working process of a content recommendation device according to an exemplary embodiment. The working process 20 at least includes steps S202 to S210.

[0057] like Figure 2 As shown, in S202, user information is obtained based on the user's recommendation request. When the user pulls down to refresh the client homepage of the preset application, a recommendation request can be generated. The user information is obtained based on the client request, and the user information can be the user's ID, registration time, age, gender, and other information. The recommendation request can also include the user's current browsing page, login time, context features, device model, system, request time, request IP, location, and other information.

[0058] In S204, a plurality of preliminary selected contents are obtained based on the user information. A pre-stored user portrait and a user vector are extracted based on the user identifier in the user information; other users are extracted based on the user portrait to generate a list of other users; and the plurality of preliminary selected contents are generated based on the user portrait, the user vector, and the list of other users.

[0059] In one embodiment, the method further includes: performing exposure deduplication on the plurality of preliminarily selected contents based on the user information, that is, the contents that have been shown to the user cannot be pushed to the user again, and the exposure deduplication filters out the contents that have been shown to the user.

[0060] In S206, based on the user information, the recommendation request, and the multiple preliminary selected contents, user features, request features, and multiple content features are generated respectively. The user information, the recommendation request, and the multiple preliminary selected contents are sorted according to a preset strategy to generate the user features, the request features, and the multiple content features.

[0061] Sorting is a very important part of the recommendation system and is also the focus of major manufacturers' recommendation systems. Sorting integrates item-side, user-side and context features, and uses a relatively complex model to score a small number of items obtained at the recall layer, with the goal of achieving personalized and accurate recommendations. In addition to focusing on accuracy, the ranking model usually also pays more attention to response speed, which requires that the model structure should not be too complex and the number of features should also be controlled.

[0062] In an embodiment of the present application, key features are extracted through statistical analysis and organization of multiple user information, multiple recommendation requests, and multiple preliminary selected contents, and preset strategies are formulated based on the key features. Each time data is received, the data is organized according to the preset strategy to generate user features, request features, and content features.

[0063] In one embodiment, some features sorted according to the preset strategy are shown in the following table. Due to space limitations, only part of the feature data is shown for illustration, and the feature data used in this application is not limited to this.

[0064]

[0065]

[0066] In S208, the user features, the request features, and the multiple content features are input into an interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents. The interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model.

[0067] In one embodiment, the user features, the request features, and the multiple content features can be input into an interest scoring model; the deep recommendation model part of the interest scoring model generates a combined feature based on the user features, the request features, and the multiple content features; the multi-task learning model part of the interest scoring model performs calculations based on the combined feature to generate multiple interest scores corresponding to the multiple content features; wherein the interest score corresponding to the content feature represents the user's interest in the initially selected content.

[0068] The specific content of "inputting the user characteristics, the request characteristics, and the multiple content characteristics into the interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents" will be described in Figure 5 The corresponding embodiments are described in detail.

[0069] In S210, at least one content to be recommended is determined from the multiple preliminary selected contents based on the multiple interest scores. The multiple preliminary selected contents are sorted based on the multiple interest scores; and at least one preliminary selected content is extracted as the content to be recommended according to the sorting. The N items with the highest scores are taken out according to the comprehensive scores of each content by the user and returned to the client for display to the user.

[0070] According to the content recommendation method of the present application, user information is obtained based on a recommendation request of the user; multiple preliminary selected contents are obtained based on the user information; user features, request features, and multiple content features are respectively generated based on the user information, the recommendation request, and the multiple preliminary selected contents; the user features, the request features, and the multiple content features are input into an interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents, and the interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model; based on the multiple interest scores, at least one content to be recommended is determined from the multiple preliminary selected contents, so that the user's interest can be analyzed from multiple angles, and recommended content can be comprehensively generated based on various information, thereby improving recommendation efficiency and user satisfaction.

[0071] It should be clearly understood that the present application describes how to form and use specific examples, but the principles of the present application are not limited to any details of these examples. On the contrary, based on the teaching of the content disclosed in the present application, these principles can be applied to many other embodiments.

[0072] As a more preferred implementation, S204 can be performed according to the following steps.

[0073] Preferably, the characteristic data required in step S204 includes: request information, user side information, and / or article side information;

[0074] A preferred characteristic data is shown in the following table:

[0075]

[0076]

[0077]

[0078]

[0079] The step S204 further comprises:

[0080] Step S204-1: storing the user behavior data in the subscription message database into the data warehouse and partitioning them according to specific time features;

[0081] Step S204-2: searching the data warehouse big wide table for the activity characteristics of the user on each channel every day during a specific time period in the past, and storing them in a first data structure;

[0082] Step S204-3: Based on the number of user behaviors in each category in the past 15 days, a custom UDF function is used to perform weighted cumulative calculation to obtain a weighted score;

[0083] The calculation formula of the UDF function is: agg_score = ∑ the number of behaviors in the past i days * 0.95i;

[0084] Step S204-4: performing outlier processing on the weight score, wherein the outlier processing refers to assigning a value of 0 to a weight value that is null, and retaining two decimal places to a weight value that is not null;

[0085] Preferably, the step S204-4 further includes concatenating the user's features such as clicks, exposures, shares, favorites, ctr and sharing rates in the past 1, 2, 3 and 15 days into one field for storage;

[0086] Step S204-4 concatenates features to avoid dimensionality explosion;

[0087] Step S204-5: Generate daily features of the content, specifically including: find out the exposure, click, share, favorite and comment count of the content in the past 1, 3, 5 and 7 days from the data warehouse, and calculate the comprehensive score through the second UDF function. The formula of the second UDF function is:

[0088] Step S204-5-a. First normalize the CTR, propagation rate, CPM, binary rate, and sharing rate. The normalization method is: (current value - minimum value) / (maximum value - minimum value);

[0089] Step S204-5-b. Calculate the comprehensive normalized score = (propagation rate normalized * 0) + (CPM normalized * 45) + (CTR normalized * 45) + (binary rate normalized * 5) + (sharing rate normalized * 5);

[0090] Step S204-5-c. Calculate the weight score of each content = (the current exposure UV of the content + click UV + sharing UV + external UV + secondary sharing UV) / exposure UV;

[0091] Step S204-5-d. The final comprehensive score of the content = weight score * normalized score;

[0092] The features required by the recommendation algorithm model include: request information, user side information, and / or article side information;

[0093] The subscription message database is preferably kafka;

[0094] Preferably, the data warehouse stored in step S204-1 is developed based on Alibaba Cloud DataWorks framework;

[0095] Preferably, the specific time feature selects hours and days as features respectively;

[0096] Preferably, the specific time period is 10-20 days; preferably, the specific time period is 15 days;

[0097] The channel refers to the classification and display of different types of content in the app in order to improve the user's reading experience. Each type of display is called a channel, for example, the news channel, beauty channel, technology channel, etc. in the app software;

[0098] Preferably, the activity characteristics include the number of exposures, clicks, shares, comments and favorites;

[0099] Preferably, the fields of the first data structure include: user ID, channel ID (such as beautiful articles, health, hot spots), behavior type (such as exposure, click, share), number of corresponding behaviors in the past day, number of corresponding behaviors in the past 2 days, number of corresponding behaviors in the past 3 days, number of corresponding behaviors in the past 15 days;

[0100] The step S204 further includes real-time feature generation, further including:

[0101] Step S204-6: performing minute-level data aggregation of articles, specifically comprising: the real-time feature generation calculation framework selects the user behavior log stream from the subscription message database, aggregates the data of content exposure, clicks, sharing, collection and comments for one minute, and stores the results in the second database;

[0102] Step S204-7: Sampling to determine sample labels, specifically including: Determine the label of each sample in the sampling stage. Features and labels constitute a complete sample. Usually, the content that users have exposed and clicked is a positive sample, and the samples that have been exposed but not clicked are negative samples. This is also the case in our business. However, in the negative sample sampling, in addition to the exposed but not clicked content as negative samples, the last clicked content in the request is also considered as negative samples even if it has not been exposed. After the label is determined, the previous features are found from hbase and the data warehouse according to the content ID and user ID to splice into a complete sample.

[0103] Preferably, the aggregation result in step S204-6 is in the form of: content ID, behavior type (exposure, click, share, comment or favorite), number of behaviors in one minute (for example, if it is clicked 100 times in one minute, the value of this column is 100), statistical start time, statistical end time (indicating which one-minute content is counted), and the result is stored in hbase;

[0104] Preferably, the real-time feature generation computing framework selects the flink framework.

[0105] Aggregation refers to: accumulating and summing the number of user behaviors of the content within a certain period of time;

[0106] Based on the minute-level aggregated data in HBase, we can expand and calculate the values ​​of content exposure, clicks, shares, favorites, number of comments, CTR, and sharing rate in the past 1, 2, 4, 8, and 12 hours. These data will be stored in a separate table in HBase to facilitate the subsequent sampling of samples to obtain corresponding features.

[0107] Figure 4 is a schematic diagram showing a typical working process of a content recommendation device according to another exemplary embodiment. Figure 4 It is combined with a specific application scenario Figure 3 A detailed description of the process shown.

[0108] In a specific application scenario, the recommendation request is passed from the client to the background server interface. After receiving the request from the client, the server calls the recommendation system interface and passes on the features required by the recommendation algorithm model.

[0109] After receiving the request from the server, the recommendation system will obtain the user portrait and the user vector pre-trained offline and stored in redis according to the user ID. The user vector is a scoring matrix generated based on the user's exposure, clicks, comments, sharing, collections and other behaviors, and then expressed by the user vector generated by matrix decomposition. The recommendation system will also obtain a list of other users with similar interests (similar user portraits) to the user, and recall multiple preliminary content based on the above information.

[0110] After receiving the above recalled content sets, the server first performs exposure elimination, and then retrieves the features of the remaining content from the background redis and MySQL according to the content ID. Then, the server requests the interest scoring model service interface of the recommendation system in protobuf format. The interest scoring model returns the user's comprehensive score for each content, and then the server retrieves the N highest-scoring items and returns them to the client for display to the user.

[0111] As a more preferred implementation, the multi-channel recall content may be performed according to the multi-channel recall step S205.

[0112] The multi-way recall step S205 further includes: the user and article vectors are obtained by matrix decomposition of a scoring matrix, and the matrix decomposition method and formula partly correspond to the following steps S205-1 and S205-2.

[0113] Step S205-1: using a user collaborative filtering algorithm, and using a content similarity algorithm to calculate a collaborative content list for each content based on the click logs of all users in the site;

[0114] Step S205-2: user hot spot recall, wherein the user hot spot recall refers to calculating the heat value of the hottest content in the category that the user is most interested in according to the heat value calculation formula;

[0115] Step S205-3: Algorithm model recall, wherein the algorithm model recall refers to vectorizing the user and the content respectively through a matrix decomposition algorithm;

[0116] Step S205-4: cold start exposure, wherein the cold start exposure refers to processing the newly added content using a multi-level traffic pool model;

[0117] Preferably, the calculation formula of the similarity algorithm in step S205-1 is:

[0118]

[0119] Where N(a) represents the number of users who clicked on content a, N(b) represents the number of users who clicked on content b, and N(a)∩N(b) represents the number of users who clicked on both a and b.

[0120] The heat value calculation formula in step S205-2 is:

[0121] For the case where the values ​​of sharing, spread rate, etc. satisfy the first characteristic vector, the heat value = ctr*A+the overall spread rate of the day*B,

[0122] Otherwise, heat value = ctr*C + overall transmission rate of the day*D;

[0123] The said spread rate refers to the number of off-site clicks / shares;

[0124] A, B, C, and D are natural numbers greater than 1.

[0125] The first feature vector refers to a combination of CTR, number of shares and propagation rate that satisfies certain conditions; preferably, the first feature vector refers to CTR>E%, number of shares greater than F and propagation rate greater than G.

[0126] The E is a real number greater than or equal to 5, the F is a natural number greater than 20, and the G is a real number greater than or equal to 1.

[0127] Further, preferably, the heat value calculation formula is:

[0128] If the content meets the requirement of cumulative CTR>8%, number of shares greater than 30 and spread rate greater than 1, then the popularity value = ctr*10000+total spread rate of the day*600.

[0129] Otherwise, the heat value = ctr*1000+the overall transmission rate of the day*60;

[0130] The said spread rate refers to the number of off-site clicks / shares;

[0131] Preferably, for some content with strong timeliness, such as good morning, good afternoon, good night, or live broadcast of games, these contents are taken offline in time by specifying the exposure time period. There are also some outlier processing, such as ctr with the following specifications:

[0132] CTR over 20% is set to 20%;

[0133] The abnormal data with CTR exceeding 100% is set to 10%;

[0134] CTR below 3% set to 3%

[0135] The matrix decomposition algorithm of step S205-3 specifically includes:

[0136] Step S205-3-1: The user and content are vectorized respectively. When recalling the user, the user vector and the content vector are calculated by the inner product to find the similarity. The scoring matrix decomposition form is as follows:

[0137]

[0138] Among them, the method of generating each element in the M (m×n) scoring matrix is ​​to generate a final weight of the user for each content according to the weight specified in the user-content behavior correspondence table, where m is a natural number representing the number of users and n is a natural number representing the number of contents:;

[0139] Step S205-3-2: With the goal of minimizing the square difference between the user's score and the score obtained by matrix multiplication, the square difference loss is used as the loss function, and a linear regression algorithm is used to perform matrix decomposition calculations.

[0140] Preferably, the loss function in step S205-3-2 is

[0141] The specific calculation method is to use the gradient descent method to calculate the loss function for p i and q j Derivative:

[0142]

[0143]

[0144] At the same time, since the training is not a single sample, it needs to be updated iteratively, p i and q j The iteration formula is:

[0145]

[0146]

[0147] Preferably, during the training process, the model update frequency of the article is hourly, and the model update frequency of the user is daily.

[0148] Preferably, the multi-level flow pool in step S205-4 specifically includes:

[0149] Step S205-4-1: Initial exposure traffic pool, take the content with the number of requests published in the last 24 hours less than 700 (the reason for judging the number of requests here is that the exposure is delayed. Although the user has requested, the exposure will be delayed because the content is not exposed on the current screen) or the number of exposures less than 100, and sort them in reverse order according to the source index. The calculation formula of the source index is:

[0150]

[0151] The reverse order output means outputting the content in the order of source index values ​​from large to small;

[0152] Wherein, a is a weight coefficient. Preferably, the default value is 1; further, as a preferred setting, if the source reading number is less than 100, a=0.001; if the source reading number is less than 500, a=0.05; if the source reading number is greater than 2000, a=source reading number / 10000;

[0153] b is the weighting coefficient. Preferably, the value of b defaults to 1. For high-quality content, the value of b is a natural number greater than 5;

[0154] c is the classification weighting coefficient. Preferably, the default value is 1. When the channel is classified into animation, automobile, literature, games, etc., c = 0.5;

[0155] exp(-kt) is the Newton cooling function, k is the cooling coefficient; it is preferably any value within the range of greater than 0 and less than 1.

[0156] Furthermore, k is set to 0.05, and t is the number of hours between the article publishing time and the current time;

[0157] s represents the source weighting coefficient. Different weighting coefficients are formulated based on the real data performance of different source station contents in the station. The coefficient value ranges from 0 to 100. Preferably, the coefficient value is proportional to the credibility of the source station. For example, when the source is a national media client such as People's Daily, s takes a value of 100.

[0158] When the article is uploaded by the editor, the source index is directly and randomly set to any value between 20000 and 30000;

[0159] Step S205-4-2: Secondary traffic pool. The content entering the secondary traffic pool is continuously requested in reverse order according to the CTR. After reaching a specific number of requests or exposure times, the content with a CTR greater than the first characteristic value will enter the tertiary traffic pool. Otherwise, it will enter the secondary pending pool. Every minute, the content with a CTR greater than the first characteristic value in the secondary pending pool will be transferred to the tertiary traffic pool.

[0160] Step S205-4-3: The content entering the third-level traffic pool is recommended to the user according to the ctr sorting. When the number of requests or exposures reaches a certain number, it is determined whether the ctr is greater than the second characteristic value. If so, the cold start end mark is marked and the cold start is ended.

[0161] Furthermore, the specific request times of step S205-4-2 is a natural number greater than 1400, and the exposure times is a natural number greater than 200; the first characteristic value is a percentage greater than 25% and less than 70%;

[0162] In step S205-4-3, the number of requests is a natural number greater than 4200, and the number of exposures is a natural number greater than 600; the second characteristic value is a percentage greater than 25% and less than 70%;

[0163] The second eigenvalue is greater than the first eigenvalue.

[0164] Step S205-4-2 and step S205-4-3 specifically include:

[0165] Step S205-4-2: Secondary traffic pool: the content entering the secondary traffic pool is continuously requested in reverse order according to the CTR. After reaching 1,400 requests or 200 exposures, the content with a CTR greater than 25% will enter the tertiary traffic pool, otherwise it will enter the secondary pending pool. Every minute, the content with a CTR greater than 25% in the secondary pending pool will be transferred to the tertiary traffic pool;

[0166] Step S205-4-3: The content entering the third-level traffic pool is sorted by ctr and pushed to users for exposure. When the number of requests reaches 4,200 or 600 exposures are obtained, it is determined whether the ctr is greater than 25%. If it is satisfied, the cold start end mark is marked and the cold start is ended. The content displayed in all modules in the site is the content at the end of the cold start.

[0167] Preferably, for a typical scenario of hotspot recall, the user's interest score for each category (such as beautiful articles, health, home, etc.) is combined with the user's behavior and scoring rules to finally generate the user's interest score for different categories of content. The user interest score is calculated based on the user's click, exposure, sharing, comment, like and collection behaviors.

[0168] Preferably, according to the scoring rules, based on the user behavior log, the interest score of each user for each category is calculated. For example, as a preferred example, the scoring rules are as follows:

[0169]

[0170]

[0171] Preferably, according to the above scoring rules, the interest scores of a certain user for different categories of content are obtained as follows:

[0172]

[0173] Preferably, the above results are stored in a Redis database.

[0174] Further, such as Figure 1 As shown, after obtaining the user's interest score for each category, the popular content in the category that the user is most interested in can be recalled and stored in the content heat value database. Preferably, the content heat value database selects a MySQL database.

[0175] Preferably, the content heat value is calculated using the following calculation formula:

[0176] If the content meets the requirement of CTR>8%, the number of shares is greater than 30, and the spread rate (spread rate = number of clicks in Moments / number of shares) is greater than 1, then the popularity value = ctr*10000+the overall spread rate of the day*600.

[0177] Otherwise, the heat value = ctr*1000+the overall transmission rate of the day*60

[0178] For some content with strong timeliness, such as good morning, good afternoon, good night, or live broadcast of games, these contents can be taken offline in time by specifying the exposure time period. There are also some outlier processing, such as ctr with the following specifications:

[0179] For CTR over 20%, set to 20%.

[0180] The abnormal data with CTR exceeding 100% is set to 10%.

[0181] For CTR below 3%, set it to 3%.

[0182] Preferably, the popularity of all content is calculated every two minutes. When an online user requests, the system first searches for the category that the user is most interested in, and then returns the top N content with the largest popularity value under the category. N is a natural number greater than 1.

[0183] In the embodiments of the present application, transmitting data in protobuf format can speed up the transmission speed, especially for large amounts of data; the protobuf format also supports cross-platform and multi-language, and can also expand the application scope of the solution in the present application.

[0184] Figure 5 is a flowchart showing a typical working process of a content recommendation device according to another exemplary embodiment. Figure 5 The process 40 shown is a detailed description of "generating the interest scoring model by jointly training the deep recommendation model and the multi-task learning model".

[0185] like Figure 5 As shown, in S402, an initial model is generated by a deep recommendation model and a multi-task learning model. The DNN hidden layer in the deep recommendation model is set to three layers, each layer has a dropout factor; the output data of the deep recommendation model (DeepFM) is used as the input data of the multi-task learning model; the expert network of the multi-task learning model (mmoe) is set to three, and the expert units are set to sixteen; the DNN hidden layer in the expert network of the multi-task learning model is set to two layers; the initial model is generated based on the above settings.

[0186] In S404, a training sample set is generated based on multiple user recommendation requests. The recommendation requests of multiple users in history can be obtained from the log, and user information, recommendation requests, and multiple content information are extracted based on the records in the log. The above data are feature sorted to generate user features, request features, and content features. A training sample set is generated based on the above data, and each user is used as a training sample.

[0187] In S406, sample labels are set for samples in the training sample set based on the content recommendation results. Log record data is extracted from the log, and labels are set for samples in the training sample set according to the content of subsequent operations of the user. More specifically, multiple labels can be set for the samples, such as click labels and propagation labels.

[0188] In S408, the initial model is trained based on the training sample set data with sample labels, and when the training parameters meet the threshold, the interest scoring model is generated.

[0189] Specifically, during each training, an initial model is established, and each sample in the training sample set is input into the initial model to obtain a predicted label. The predicted label is compared with the corresponding real label to determine whether the predicted label is consistent with the real label, and the number of predicted labels consistent with the real label is counted, and the proportion of the number of predicted labels consistent with the real label in the number of all predicted labels is calculated. If the proportion is greater than or equal to the preset proportion value, the initial model converges to obtain a trained interest scoring model. If the proportion is less than the preset proportion value, the parameters in the initial model are adjusted, and the predicted labels of each object are re-predicted by the adjusted initial model until the proportion is greater than or equal to the preset proportion value. Among them, the method of adjusting the parameters in the initial model can be performed using a stochastic gradient descent algorithm, a gradient descent algorithm, or a normal equation.

[0190] In the initial model, the MMoe model cooperates with the DeepFM model, and the output of DeepFM is used as the input of the MMoe model. In the actual model training, it is found that due to the large number of parameters in the depth part of MMoe and DeepFM, it is very easy to have overfitting problems. In the initial model, the depth of the DNN side is initially set to 3 layers, the hidden units are set to [128, 64, 32], the number of expert networks in the MMoe part is 3, the expert units are 16, and the hidden units are [128, 64, 32]. This parameter configuration shows that the auc and acc of the training data are 0.86 and 0.92 respectively during model training, but the performance on the test set is extremely poor (auc is 0.52, acc is 0.6), which is a typical overfitting performance.

[0191] In an embodiment of the present application, L2 regularization penalty and dropout are added to the network layers of DNN and mmoe, wherein the dropout ratio is set to 0.6, and the L2 regularization penalty factor is 0.01. At this time, the auc and acc of the model training are increased to 0.65 and 0.73 respectively, and then the number of expert networks in mmoe is further adjusted to 2 during the training process, and the hidden units are changed to [16,8]. The dropout ratio in the network layer is increased from 0.6 to 0.7. At this time, the performance of the model training for the test set data becomes auc of 0.75 and acc of 0.89. After going online, the ctr is increased by 4.5% without increasing the features.

[0192] Figure 6 FIG. 1 is a schematic diagram showing a typical working process of a content recommendation device according to another exemplary embodiment. Figure 6 Yes Figure 3 The process S208 shown, "inputting the user feature, the request feature, and the multiple content features into the interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents" is described in detail.

[0193] In one embodiment, the user features, the request features, and the multiple content features can be input into an interest scoring model; the deep recommendation model part of the interest scoring model generates a combined feature based on the user features, the request features, and the multiple content features; the multi-task learning model part of the interest scoring model performs calculations based on the combined feature to generate multiple interest scores corresponding to the multiple content features; wherein the interest score corresponding to the content feature represents the user's interest in the initially selected content.

[0194] More specifically, the multi-task learning model part in the interest scoring model calculates and generates the click probability and the propagation probability based on the combined features; and generates the interest scores corresponding to the multiple content features based on the click probability, the propagation probability and their corresponding weights.

[0195] Among them, the FM model is a machine learning algorithm based on matrix decomposition. The DeepFM model can process many features that traditional models cannot handle. For example, it can directly learn the implicit information of high-dimensional sparse IDs from massive training samples and represent it through Embedding. In addition, for text, sequence features and image features, the deep network has corresponding structures or units to process them.

[0196] In actual calculations, user features, request features, and content features are discretized and input into the DeepFM part of the initial model. After hidden layer calculations, combined features are proposed and input into the MMoe model part. The MMoe model pre-sets various parameters obtained by model training, and the user's click probability and forwarding probability are calculated. Based on these two probabilities and the weights corresponding to the two probabilities, the final interest score value can be obtained.

[0197] Those skilled in the art will appreciate that all or part of the steps for implementing the above embodiments are implemented as a computer program executed by a CPU. When the computer program is executed by the CPU, the above functions defined by the above method provided in the present application are executed. The program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0198] In addition, it should be noted that the above figures are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present application, and are not intended to be limiting. It is easy to understand that the processes shown in the above figures do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be performed synchronously or asynchronously, for example, in multiple modules.

[0199] Figure 7 It is a block diagram of an electronic device according to an exemplary embodiment.

[0200] Refer to the following Figure 7 hereinafter describes an electronic device 700 according to this embodiment of the present application. Figure 7 The electronic device 700 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0201] like Figure 7 As shown, the electronic device 700 is in the form of a general computing device. The components of the electronic device 700 may include, but are not limited to: at least one processing unit 710, at least one storage unit 720, a bus 730 connecting different system components (including the storage unit 720 and the processing unit 710), a display unit 740, etc.

[0202] The storage unit stores a program code, which can be executed by the processing unit 710, so that the processing unit 710 performs the steps described in this specification according to various exemplary embodiments of the present application. For example, the processing unit 710 can perform the following steps: Figure 3 , Figure 5 Follow the steps shown in .

[0203] The storage unit 720 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 7201 and / or a cache memory unit 7202 , and may further include a read-only memory unit (ROM) 7203 .

[0204] The storage unit 720 may also include a program / utility 7204 having a set (at least one) of program modules 7205, such program modules 7205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each of which or some combination may include the implementation of a network environment.

[0205] Bus 730 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0206] The electronic device 700 may also communicate with one or more external devices 700' (e.g., keyboards, pointing devices, Bluetooth devices, etc.) to enable a user to communicate with the electronic device 700, and / or any device (e.g., routers, modems, etc.) that the electronic device 700 can communicate with one or more other computing devices. Such communication may be performed via an input / output (I / O) interface 750. In addition, the electronic device 700 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 760. The network adapter 760 may communicate with other modules of the electronic device 700 via a bus 730. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0207] Through the above description of the implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by combining software with necessary hardware. Figure 8 As shown, the technical solution according to the implementation mode of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes a number of instructions to enable a computing device (which can be a personal computer, a server, or a network device, etc.) to execute the above method according to the implementation mode of the present application.

[0208] The software product may be any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0209] The computer readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by an instruction execution system, an apparatus, or a device or used in combination with it. The program code contained on the readable storage medium may be transmitted with any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

[0210] Program code for performing the operations of the present application may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., using an Internet service provider to connect through the Internet).

[0211] The computer-readable medium carries one or more programs. When the one or more programs are executed by a device, the computer-readable medium implements the following functions: obtaining user information based on a user's recommendation request; obtaining multiple preliminary selected contents based on the user information; generating user features, request features, and multiple content features based on the user information, the recommendation request, and the multiple preliminary selected contents respectively; inputting the user features, the request features, and the multiple content features into an interest scoring model to obtain multiple interest scores corresponding to the multiple preliminary selected contents, wherein the interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model; determining at least one content to be recommended from the multiple preliminary selected contents based on the multiple interest scores. The computer-readable medium may also implement the following functions: generating an initial model through a deep recommendation model and a multi-task learning model; training the initial model based on multiple user recommendation requests and multiple content recommendation results to generate the interest scoring model.

[0212] Those skilled in the art will appreciate that the above modules can be distributed in the device according to the description of the embodiment, or can be changed accordingly and only used in one or more devices different from the embodiment. The modules of the above embodiments can be combined into one module, or further divided into multiple sub-modules.

[0213] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the embodiment of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) to execute the method according to the embodiment of the present application.

[0214] The exemplary embodiments of the present application are specifically shown and described above. It should be understood that the present application is not limited to the detailed structures, configurations or implementations described herein; on the contrary, the present application is intended to cover various modifications and equivalent configurations included in the spirit and scope of the appended claims.

Claims

1. A content recommendation device, characterized in that: include: An information module, used to obtain user information based on the user's recommendation request; A preliminary selection module, used for obtaining a plurality of preliminary selection contents based on the user information; A feature module, configured to generate a user feature, a request feature, and a plurality of content features based on the user information, the recommendation request, and the plurality of pre-selected contents; A scoring module, configured to input the user feature, the request feature, and the plurality of content features into an interest scoring model to obtain a plurality of interest scores corresponding to the plurality of preliminary selected contents, wherein the interest scoring model is generated by joint training of a deep recommendation model and a multi-task learning model; A recommendation module, configured to determine at least one content to be recommended from the plurality of preliminarily selected contents based on the plurality of interest scores; Among them, the method for generating user features, request features, and multiple content features based on the user information, the recommendation request, and the multiple preliminary selected contents is implemented by using a multi-way recall module, and the specific steps further include: user and article vectors are obtained by matrix decomposition of a scoring matrix, and the matrix decomposition method and formula part correspond to the following steps S205-1 and S205-2; step S205-1: using a user collaborative filtering algorithm, and using a content similarity algorithm to calculate a collaborative content list for each content based on the click logs of all users in the site; step S205-2: user hot spot recall, the user hot spot recall refers to the calculation of the heat value formula for the hottest content in the category that the user is most interested in, and the heat of the content; step S205-3: algorithm model recall, the algorithm model recall refers to vectorizing users and content respectively through a matrix decomposition algorithm; step S205-4: cold start exposure, the cold start exposure refers to the use of a multi-level traffic pool model to process newly added content; The step of inputting the user features, the request features, and the multiple content features into the interest scoring model to obtain the multiple interest scores corresponding to the multiple preliminary selected contents includes: inputting the user features, the request features, and the multiple content features into the interest scoring model; the deep recommendation model part in the interest scoring model generates a combined feature based on the user features, the request features, and the multiple content features; the multi-task learning model part in the interest scoring model performs calculations based on the combined feature to generate the multiple interest scores corresponding to the multiple content features; wherein the interest score corresponding to the content feature represents the user's interest level in the preliminary selected contents.

2. The device according to claim 1, characterized in that Also includes: The training module is used to generate the initial model through the deep recommendation model and the multi-task learning model; The initial model is trained based on multiple user recommendation requests and multiple content recommendation results to generate the interest scoring model.

3. The device according to any one of claims 1 or 2, characterized in that: The obtaining of a plurality of preliminary selected contents based on the user information further includes: extracting a pre-stored user portrait and a user vector based on a user identifier in the user information; extracting other users based on the user portrait to generate a list of other users; and generating the plurality of preliminary selected contents based on the user portrait, the user vector, and the list of other users.

4. The device according to claim 2, characterized in that Also includes: The recommendation module passes through a deep recommendation submodule and a multi-task learning submodule, and uses the output data of the deep recommendation submodule as input data of the multi-task learning submodule.

5. The device according to claim 4, characterized in that Also includes: The DNN hidden layer in the deep recommendation model is set to three layers, each with a dropout factor; The output data of the deep recommendation model is used as the input data of the multi-task learning model; the expert network of the multi-task learning model is set to two, and the expert units are set to sixteen; the DNN hidden layer in the expert network of the multi-task learning model is set to two layers; and the initial model is generated based on the above settings.

6. The device according to claim 2, characterized in that The interest scoring model is trained and generated on the initial model based on multiple user recommendation requests and multiple content recommendation results, including: generating a training sample set based on multiple user recommendation requests; setting sample labels for samples in the training sample set based on the content recommendation results; training the initial model based on the training sample set data with sample labels; and generating the interest scoring model when the training parameters meet a threshold.

7. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the apparatus as described in any one of claims 1 to 6.

8. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the apparatus according to any one of claims 1 to 6 is implemented.