Content recommendation system

Abstract

The invention provides a content recommendation system. The content recommendation method comprises the steps of obtaining to-be-processed interaction information; processing the to-be-processed interaction information according to target scene information to obtain target embedded information; and determining target recommendation information based on the target embedded information.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a content recommendation system. Background Technology

[0002] Currently, with the rapid development of the internet and various applications, more and more applications incorporate recommendation systems. These systems automatically calculate user preferences and provide high-quality recommendations. Recommendation systems are widely used in various content recommendation applications, such as online shopping platforms, music recommendation systems, movie recommendation systems, and video-on-demand systems, covering various content types such as text, images, and videos. Recommendation systems can recommend content that a user is interested in based on user input and browsing history. Summary of the Invention

[0003] This application provides a content recommendation system.

[0004] On one hand, this application provides a method comprising the following steps:

[0005] Obtain the interaction information to be processed;

[0006] The interaction information to be processed is processed based on the target scene information to obtain target embedding information;

[0007] Based on the target embedding information, target recommendation information is determined.

[0008] On the other hand, this application provides a system comprising:

[0009] The information acquisition module is configured to acquire interactive information to be processed.

[0010] The embedding extraction module is configured to process the interaction information to be processed based on the target scene information to obtain the target embedding information;

[0011] The content recommendation module is configured to determine target recommendation information based on the target embedding information.

[0012] On the other hand, this application also provides a content recommendation device, the content recommendation device comprising:

[0013] One or more processors;

[0014] Memory; and

[0015] One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the steps of the content recommendation method.

[0016] On the other hand, this application also provides a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to perform the steps in the content recommendation method. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram illustrating a scenario for the method recommended in the embodiments of this application;

[0019] Figure 2 This is a flowchart illustrating one embodiment of the content recommendation method in this application.

[0020] Figure 3 A flowchart illustrating an embodiment of the content recommendation method for determining target recommendation information provided in this application;

[0021] Figure 4 A flowchart illustrating an embodiment of the content recommendation method provided in this application for training a target processing model;

[0022] Figure 5 A schematic diagram of the structure of one embodiment of the content recommendation system provided in this application;

[0023] Figure 6 A schematic diagram of the structure of one embodiment of the content recommendation device provided in this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise expressly defined.

[0026] In this application, the term "exemplary" is used to mean "serving as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0027] To facilitate understanding of the content recommendation method provided in this application embodiment, the application scenarios of the content recommendation method provided in this application embodiment are first described. Optionally, the content recommendation method provided in this application embodiment is mainly applied to users using various content applications (such as shopping applications, audio-visual applications, and content community applications, etc.). For example, when a user is using an audio-visual application, such as when a user selects a video to watch, the content recommendation system in the video application can recommend video content that the user may be interested in by reading the user's historical viewing records, thereby enabling the user to watch the video content they want to watch. As another example, when a user is using a content recommendation application such as Weibo or Xiaohongshu, the content recommendation system in the content recommendation application can read the user's historical notes / Weibo historical reading records and recommend notes / Weibo content that the user may be interested in based on these historical notes or Weibo historical reading records.

[0028] Most existing content recommendation methods predict user interests and then recommend content accordingly. However, most of these methods ignore the value of conformity. The causal entanglement between conformity and interests can interfere with interest modeling during the recommendation process. This means that the recommended content may simply be currently popular and conforming, rather than content that users are genuinely interested in. Furthermore, in actual recommendation processes, there may be a large amount of sparse long-tail content (referring to content with low search volume but wide coverage on the internet), which can also affect the accuracy, robustness, and interpretability of content recommendations.

[0029] Based on this, this application proposes a content recommendation system to solve the technical problems of poor accuracy, robustness and interpretability of recommended content in existing content recommendation systems.

[0030] Based on this, this application proposes a content recommendation system to solve the technical problem of high error rate in the data integration process of the prior art.

[0031] The content recommendation method in this embodiment of the invention is applied to a content recommendation system, which includes a content recommendation device. The content recommendation device is equipped with one or more processors, a memory, and one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the content recommendation method. The content recommendation device can be a smart terminal, such as a mobile phone, tablet computer, network device, and smart computer. Optionally, the content recommendation device can also be a server or a service cluster composed of multiple servers.

[0032] like Figure 1 As shown, Figure 1 This is a schematic diagram of a content recommendation method according to an embodiment of the present application. The content recommendation scenario in this embodiment includes a content recommendation device 100 (the content recommendation device 100 integrates a content recommendation system), and the content recommendation device 100 is equipped with a computer-readable storage medium corresponding to the content recommendation method to execute the steps of the content recommendation method.

[0033] Understandable, Figure 1 The content recommendation device in the content recommendation method scenario shown, or the device included in the content recommendation device, does not constitute a limitation on the embodiments of the present invention. That is, the number or type of the content recommendation device in the content recommendation method scenario, or the number or type of the device included in each device, does not affect the overall implementation of the technical solution in the embodiments of the present invention, and can all be considered as equivalent substitutions or derivatives of the technical solutions claimed in the embodiments of the present invention.

[0034] In this embodiment of the invention, the content recommendation device 100 is mainly used for: acquiring interaction information to be processed;

[0035] The interaction information to be processed is processed based on the target scene information to obtain target embedding information;

[0036] Based on the target embedding information, target recommendation information is determined and output.

[0037] In this embodiment of the invention, the content recommendation device 100 can be an independent content recommendation device, such as a mobile phone, tablet computer, network device, server, and smart computer, or it can be a content recommendation network or content recommendation cluster composed of multiple content recommendation devices.

[0038] This application provides a content recommendation system, which will be described in detail below.

[0039] It will be understood by those skilled in the art that Figure 1 The application environment shown is only one application scenario related to the solution of this application and does not constitute a limitation on the application scenario of this application. Other application environments may include more than one application scenario. Figure 1 The number of more or fewer content recommendation devices shown, or content recommendation network connections, for example Figure 1 Only one content recommendation device is shown in the diagram. It is understood that the scenario of this content recommendation method may also include one or more content recommendation devices, which are not specifically limited here. The content recommendation device 100 may also include a memory for storing recommended content and other data.

[0040] It should be noted that, Figure 1 The schematic diagram of the content recommendation method shown is merely an example. The scenarios of the content recommendation method described in the embodiments of the present invention are intended to more clearly illustrate the technical solutions of the embodiments of the present invention and do not constitute a limitation on the technical solutions provided in the embodiments of the present invention.

[0041] Based on the scenarios described above for content recommendation methods, various embodiments of the content recommendation method disclosed in this invention are proposed.

[0042] like Figure 2 As shown, Figure 2 This is a flowchart illustrating one embodiment of the content recommendation method in this application. The content recommendation method includes the following steps 201 to 203:

[0043] 201. Obtain the interaction information to be processed;

[0044] The content recommendation method in this embodiment is applied to a content recommendation device. The type and number of content recommendation devices are not specifically limited; that is, the content recommendation device can be one or more smart terminals or servers equipped with a target processing model capable of content recommendation. In one specific embodiment, the content recommendation device is a smart computer. The target processing model is a model algorithm in which an artificial intelligence system performs causal inference and graph contrastive learning based on the user's interaction information to be processed, thereby outputting content recommendation results. Optionally, in one specific embodiment, the target processing model may include a causal embedding module, a decoupling learning module, and a target prediction module. The causal embedding module is a deep learning model that determines the first embedding information corresponding to the interaction information to be processed. For example, the causal embedding module can be any one or more of a neural network model, a convolutional neural network model, or a recurrent neural network model. The decoupling learning module is a graph contrastive learning module that separates the interest embedding information and the popularity embedding information in the first embedding information for decoupling and contrastive learning. The target prediction module is a module that determines the interaction probability of each candidate recommendation information based on the target embedding information and outputs a prediction of the target recommendation information.

[0045] Optionally, in specific application scenarios, more and more manufacturers are integrating target processing models into content recommendation applications (such as shopping applications, video applications, and content community applications). When users use content recommendation applications, these target processing models perform causal inference and decoupled comparative learning based on users' historical browsing records and other interaction information. This allows them to identify users' interests and conformity in specific scenarios, and then recommend target recommendation information that matches users' true interests and intentions and is not affected by users' conformity, thus making content recommendations more accurate.

[0046] Optionally, during operation, the content recommendation device responds to user content recommendation requests, whereby the content recommendation device is an operational instruction that drives the content recommendation device to recommend content of interest to a specific user. Optionally, the triggering method for this content recommendation request is not specifically limited here; that is, the content recommendation request can be actively triggered by the user, for example, by clicking a content recommendation button or refreshing the recommendation information. Furthermore, the content recommendation request can also be automatically triggered by the content recommendation device; for example, the content recommendation device may have a preset timed recommendation process that actively triggers content recommendation requests within a preset time period to obtain target recommendation information and update the current recommendation information.

[0047] In this system, after receiving a content recommendation request, the content recommendation device also obtains the user identifier of the user currently using the content recommendation application, and obtains the user's pending interaction information based on this user identifier. This pending interaction information is an interaction chart recording the user's historical interaction information. For example, if the content recommendation application is a video / audio application, the pending interaction information can be a movie viewing interaction chart constructed based on the user's historical movie viewing history. As another example, if the content recommendation application is a shopping application, the pending interaction information can be a user shopping interaction chart constructed based on shopping information recently clicked by the user. In a specific embodiment, the pending interaction information can include the interacting user, interaction item, and interaction operation. The interaction item can be any one or more of video / audio content, shopping information, notes, articles, and microblogs. Optionally, the interaction item is associated with the application type of the content recommendation application; different content recommendation applications have different interaction items. The interaction operation represents the interactive action performed by the user based on the interaction item. For example, the interaction operation can be a click interaction. The interaction chart is an interaction record data table recording the user's interaction information during historical interactions. In one specific embodiment, the fields in the interactive chart include at least an interactive user field, an interactive item field, and an interactive operation field.

[0048] 202. Process the interaction information to be processed according to the target scene information to obtain target embedding information;

[0049] Optionally, after acquiring the interaction information to be processed, the content recommendation device further processes the interaction information based on the target scenario information to obtain target embedding information representing the causal relationships of interaction interest and conformity. This identifies and decouples the causal relationships of interest and conformity during user interaction, thereby eliminating the influence of conformity on user behavior. This allows the system to dynamically adjust personalized recommendation results based on factors such as the recommendation scenario and scenario popularity during subsequent content recommendations, improving recommendation accuracy.

[0050] Optionally, after acquiring the interaction information to be processed, the content recommendation device also calculates the popularity of each interaction item in the interaction information. Here, the popularity of the interaction item is a quantitative parameter characterizing the popularity of each interaction item in the interaction information to be processed. Popularity is a metric parameter characterizing the popularity or breadth of a specific thing within a specified range (e.g., time, location, and category range). In a specific embodiment, the calculation of popularity can be simplified and abstracted as the ratio between the number of users interacting with a specific thing or the number of effective interactions and the total number of exposures of that specific thing. The popularity of the interaction item is the popularity parameter corresponding to the popularity of the interaction item.

[0051] Optionally, the content recommendation device obtains the interaction item corresponding to a specific interaction operation in the interaction information to be processed, reads the number of user interactions for the interaction item and the total number of exposures of the candidate recommendation information, and calculates the ratio of the number of user interactions to the total number of exposures to obtain the popularity of the interaction item.

[0052] Optionally, after acquiring the interaction information to be processed, the content recommendation device further generates first embedding information of the interaction information to be processed based on the causal embedding module and the interaction information to be processed. That is, the content recommendation device inputs the interaction information to be processed into the causal embedding module in the target processing model, and uses the causal embedding module to perform causal inference on the interaction information to be processed to obtain the first embedding information of the interaction information to be processed.

[0053] Optionally, after acquiring the interaction information to be processed, the content recommendation device uses a causal embedding module to map the causal relationships in the interaction information to vector space points to obtain first embedding information. The first embedding information is embedding vector information representing the causal relationships between the interaction information to be processed and subsequent interactions with interactive items. This causal embedding module is a neural network model trained based on causal relationship samples to minimize the distance between embedding vectors or maximize similarity, accurately capturing the causal correlation in vector space.

[0054] Optionally, the content recommendation device further utilizes the target scene information corresponding to the interaction information to be processed to generate scene mask information corresponding to the interaction information to be processed. This scene mask information is a scene indicator variable used to correct the popularity of each interaction item.

[0055] Optionally, after determining the popularity of each interactive item, the content recommendation device further utilizes scene mask information to decouple the popularity of interactive items from the scene, since the popularity of interactive items is considered from the popularity of all items. However, the popularity of different interactive items is not the same in different scenarios. Therefore, the content recommendation device further utilizes scene mask information to decouple the popularity of interactive items from the scene, so as to remove the influence of items in non-target scenes and improve the accuracy of subsequent content recommendations.

[0056] Optionally, after acquiring the interaction information to be processed and identifying the interaction items within that information, the content recommendation device also acquires the corresponding interaction item information, parses it, and determines the scene identifier information corresponding to that interaction item. This scene identifier information represents the interaction scene to which the interaction item belongs. For example, in a film and television recommendation scenario, this scenario can be further subdivided into multiple sub-domains based on specific segmentation logic, such as a movie recommendation sub-domain and a children's recommendation sub-domain. The popularity of interaction items might then consider cross-domain popularity, thus interfering with the accuracy of content recommendation.

[0057] Optionally, after obtaining scene identification information, the content recommendation device accesses a preset scene database to retrieve the target scene information corresponding to the scene identification information. Furthermore, it determines the scene mask information corresponding to the target scene information based on the mask generation module. That is, after obtaining the target scene information, the content recommendation device establishes scene mask variables related to the target scene information to obtain scene mask information, where the scene mask information is an indicator variable representing the characteristics or attributes of the target scene information.

[0058] Optionally, after accessing the scene database, the content recommendation device determines the candidate scene information associated with the interactive item information in the scene database, as well as the candidate identifier information of the candidate scene information. The candidate scene information whose candidate identifier information matches the scene identifier information to be recommended is determined as the target scene information of the interactive item information. Furthermore, a scene mask information for the target scene information is generated based on the target scene information and the candidate scene information. Here, the candidate scene information consists of multiple scene information associated with the interactive item information. For example, each interactive item information can be divided into different interactive scenes according to different classification methods, and each scene corresponds to one candidate scene information.

[0059] Optionally, after determining the candidate scene information and target scene information associated with the interactive item information, the content recommendation device obtains a first scene value for the candidate scene information and a second scene value for the target scene information. The first scene value is a variable value or other unique representation parameter characterizing the candidate scene information. The second scene value is a variable value or other unique representation parameter characterizing the target scene information.

[0060] Specifically, after determining the first scene value and the second scene value, the content recommendation device performs a bitwise OR operation on the first scene value and the second scene value to obtain the scene mask information corresponding to the target scene information.

[0061] Optionally, after obtaining the scene mask information, the content recommendation device further calculates the scene popularity of candidate recommendation information based on the scene mask information and the popularity of the interaction item. That is, the content recommendation device uses the target processing model to correct the popularity of the interaction item using the scene mask information, thereby limiting the popularity of the interaction item to the target interaction scene, and then calculates the scene popularity of each candidate recommendation item corresponding to the popularity of the interaction item. Here, the candidate recommendation information refers to interaction items that belong to the same interaction item type as the interaction item in the interaction information to be processed and belong to the target interaction scene. Optionally, the formula for calculating the scene popularity is:

[0062] Optionally, after obtaining the scene popularity, the content recommendation device further generates target embedding information for the interaction information to be processed based on the scene popularity, the first embedding information, the initial scene mask information, and the target processing model. This target embedding information includes target user embedding information and item embedding information for each interaction item. That is, the content recommendation device inputs the scene popularity and the first embedding information into the decoupling learning module in the target processing model. The decoupling recognition module then performs causal decoupling learning on the first embedding information based on the scene popularity, thereby obtaining target embedding information that can represent the user's true interest intent.

[0063] Optionally, the content recommendation device inputs the scene popularity and the first embedding information into the target processing model for embedding calculation to obtain the target embedding information. The calculation method of the target embedding information is as follows:

[0064]

[0065] Among them, e u and e i To embed information into the target, α k ≥0 represents the weight hyperparameters embedded in the target embedding information at the k-th layer of the target processing model. and This is the first embedded information.

[0066] Optionally, in other embodiments, before generating target embedding information of the interaction information to be processed based on scene popularity, first embedding information, and target processing model, the content recommendation device further performs decoupled comparative learning training on the preset network model to generate a target processing model, and uses the target processing model to determine the target embedding information of the interaction information to be processed.

[0067] 203. Based on the target embedding information, determine the target recommendation information.

[0068] Optionally, after obtaining the target embedding information, the content recommendation device also determines the target recommendation information to be recommended based on the target embedding information, and outputs the target recommendation information to the user after determining the target recommendation information, thereby reducing interference from users' herd behavior, improving the personalization and accuracy of the content recommendation process, and meeting the user's content recommendation needs.

[0069] Optionally, the content recommendation device can calculate the embedding similarity information between the target user embedding information and the item embedding information in the target embedding information, determine the target recommendation information in the candidate recommendation information based on the embedding similarity information, and output the target recommendation information to the user after determining the target recommendation information.

[0070] Optionally, the content recommendation device can also utilize the structural causal module in the target processing model to calculate the target causal relationship and interaction causal information between the user and the interactive item in the target embedding information. The target causal relationship includes interest causal relationship, conformity causal relationship, and comprehensive causal relationship. Optionally, the causal relationship expression between the user and the interactive item is as follows:

[0071]

[0072] Among them, e interrst ,e conformity ,e total As independent noise, To establish the causal relationship of interests between users and interactive projects. To represent the causal relationship of user interest in interactions with interactive items, s ui This refers to the comprehensive causal relationship between the user and the interactive project.

[0073] Optionally, after obtaining the target causal relationship, the content recommendation device also uses an additive model to calculate the interaction causal information of the target causal relationship. Optionally, this interaction causal information is a causal score characterizing the relationship between the user and the interactive item. Optionally, the calculation method for this interaction causal information is as follows:

[0074]

[0075] Optionally, after determining the interaction causal information, the content recommendation device also determines the target recommendation information based on the interaction causal information. That is, the content recommendation device compares each interaction causal information and determines the candidate recommendation information corresponding to the interaction item with the highest interaction causal information as the target recommendation information.

[0076] Optionally, in other embodiments, after generating target embedding information using the target processing model, the content recommendation device further acquires the target user embedding information and the embedding information of each item within the target embedding information, and calculates the inner product parameters between the target embedding information and each item embedding information. These inner product parameters are used as recommendation evaluation parameters between the target embedding information and each item embedding information. Candidate recommendation information corresponding to the item embedding information is then sorted from highest to lowest according to the recommendation evaluation parameters. Based on a preset recommendation quantity, several candidate recommendation information items ranked highest by the recommendation evaluation parameters are determined as target recommendation information, and this target recommendation information is recommended to the user. The calculation method for the recommendation evaluation parameters is as follows:

[0077]

[0078] in, These are the recommended evaluation parameters.

[0079] Optionally, in other embodiments, the content recommendation device can also perform normalized weighted summation processing on the above-mentioned interaction causal relationship, embedded similarity information and recommendation evaluation parameters according to preset recommendation weight information to obtain target evaluation parameters, sort the corresponding candidate recommendation information according to the target evaluation parameters, determine the top-ranked candidate recommendation information as target recommendation information according to preset recommendation quantity, and recommend the target recommendation information to the user.

[0080] In this embodiment, the content recommendation device acquires interaction information to be processed; processes the interaction information to be processed according to target scene information to obtain target embedding information; determines target recommendation information based on the target embedding information, and outputs the target recommendation information. This achieves the goal of acquiring interaction information to be processed and using a scene mask corresponding to the interaction information, a causal embedding module, and a decoupled contrastive learning module to identify interest-based causal embeddings and conformity-based causal embeddings in the interaction information to obtain target embedding information that separates interest-based causal relationships and popularity-based causal relationships in the interaction information. That is, the target embedding information can consider the influence of scene popularity and interest as much as possible to adjust the popularity bias in the recommendation system. Based on the target embedding information, the device outputs the inference results of the corresponding recommendation task and generates target recommendation information, thereby improving the personalization and accuracy of the content recommendation process to meet users' content recommendation needs.

[0081] like Figure 3 As shown, Figure 3 This is a flowchart illustrating an embodiment of the content recommendation method for determining target recommendation information provided in this application. Optionally, in this embodiment, the content recommendation method further includes steps 301 to 303:

[0082] 301. Obtain the target user embedding information and the embedding information of each item from the target embedding information;

[0083] 302. Calculate the embedding similarity information between each item's embedding information and the target user's embedding information;

[0084] 303. Determine the target project embedding information in the project embedding information based on the embedding similarity information, and obtain the target recommendation information corresponding to the target project embedding information.

[0085] Based on the above embodiments, in this embodiment, after obtaining the target embedding information, the content recommendation device also determines the target recommendation information to be recommended based on the target embedding information, and outputs the target recommendation information to the user after determining the target recommendation information.

[0086] Optionally, after obtaining the target embedding information, the content recommendation device further parses the target embedding information to obtain the target user embedding information and item embedding information. The target user embedding information is the embedding vector feature representation of the target user to be recommended. The item embedding information is the embedding vector feature representation of the interactive items historically interacted with by the target user. This item embedding information and target user embedding information can reflect the interest-based causal relationship and conformity-based causal relationship between the target user and each interactive item; that is, the item embedding information and target user embedding information can take into account the scene popularity influence of each candidate recommendation information, thereby improving the accuracy of content recommendation.

[0087] Optionally, after obtaining the target user embedding information and the item embedding information, the content recommendation device further calculates the embedding similarity between the target user embedding information and each item embedding information. The embedding similarity can be measured by Euclidean distance, cosine similarity, and the similarity distance between Manhattan distance.

[0088] For example, in an optional specific embodiment, the content recommendation device can use the target processing model to calculate the Euclidean distance between the target user embedding information and the embedding information of each item, that is, calculate the straight-line distance between the target user embedding information and the item embedding information, and use the Euclidean distance as the embedding similarity information between the target user embedding information and the item embedding information.

[0089] For example, in an optional specific embodiment, the content recommendation device can also use the target processing model to calculate the cosine similarity between the target user embedding information and the embedding information of each item, that is, calculate the angle between the feature vectors of the target user embedding information and the embedding information of each item, and use the cosine similarity as the embedding similarity information between the target user embedding information and the embedding information of each item.

[0090] Optionally, after obtaining the embedding similarity information between the target user embedding information and the embedding information of each item, the content recommendation device further utilizes the target prediction module in the target processing model to perform recommendation inference on the embedding information of each item based on the embedding similarity information, in order to obtain the target recommendation information to be recommended. The target prediction module can be any one or more graph neural network models, such as LightGCN and graph convolutional models. Optionally, the method for determining the target recommendation information is as follows:

[0091] Argmax(similarity(e_im,e_un)for min(1,…,M,M refers to the number of embedded information items), where e_un is the embedded information of the target user for the current user n.

[0092] That is, after obtaining the embedding similarity information between the target user embedding information and the embedding information of each item, the content recommendation device also determines the target item embedding information in the item embedding information based on the embedding similarity information, obtains the candidate recommendation information associated with the target item embedding information, determines the candidate recommendation information as the target recommendation information, and outputs the target recommendation information.

[0093] Optionally, in one specific embodiment, the content recommendation device determines the project embedding information with the highest embedding similarity information among the project embedding information as the target project embedding information. Optionally, in another specific embodiment, after obtaining the embedding similarity information corresponding to each project embedding information, the content embedding device further sorts the project embedding information from largest to smallest according to the numerical value of the project embedding information to obtain the sorting number corresponding to each project embedding information (wherein, the smaller the sorting number, the larger the corresponding embedding similarity information), and determines the project embedding information with a sorting number less than a preset sorting number threshold as the target project embedding information. That is, the content recommendation device determines the embedding information of several projects with the highest embedding similarity information as the target project embedding information, obtains the candidate recommendation information associated with the target project embedding information, determines the candidate recommendation information as the target recommendation information, and outputs the target recommendation information. The preset sorting number threshold can be customized according to actual recommendation needs, and this embodiment does not specifically limit it.

[0094] In this embodiment, the content recommendation device acquires the target user embedding information and the embedding information of each item from the target embedding information; calculates the embedding similarity information between each item embedding information and the target user embedding information; acquires the target item embedding information with the highest embedding similarity information among the item embedding information, and the target recommendation information corresponding to the target item embedding information, and outputs the target recommendation information. This achieves accurate content recommendation by determining the embedding similarity information between the user and each interactive item during content recommendation, thus considering the impact of scene popularity on interaction as much as possible, and accurately identifying the user's true interest intent, thereby improving the accuracy and stability of content recommendation.

[0095] like Figure 4 As shown, Figure 4 This is a flowchart illustrating an embodiment of the content recommendation method for training a target processing model, as provided in this application. Optionally, in this embodiment, the content recommendation method further includes steps 401 to 404:

[0096] 401. Perform a first comparative learning based on the first embedded information and the scene popularity to obtain first loss information;

[0097] 402. Perform a second comparative learning based on the first embedded information and the scene mask information to obtain second loss information;

[0098] 403. Generate target loss information based on the main task loss information, the first loss information, the second loss information, and the target hyperparameters;

[0099] 404. Train the preset network model based on the target loss information to obtain the target processing model.

[0100] Based on the above embodiments, in this embodiment, in order to avoid interference with the accuracy of content recommendation due to herd mentality, before generating the target embedding information of the interaction information to be processed based on the scene popularity, the first embedding information and the target processing model, the content recommendation device also performs decoupled comparative learning training on the preset network model to generate the target processing model, and uses the target processing model to determine the target embedding information of the interaction information to be processed.

[0101] Optionally, the target processing model includes a decoupling learning module. This module pre-sets multiple user-item contrastive learning tasks and uses these tasks to perform embedding information separation learning on the first embedding information, separating it into popular embedding information and interest embedding information. The popular embedding information is an embedding vector representing a user's interaction with interactive items based on conformity behavior. The interest embedding information is an embedding vector representing a user's interaction with interactive items based on interests. Specifically, for interactive items, the popular embedding information is related to the popularity of the interactive item, while the interest embedding information is related to the item content of the interactive item.

[0102] Optionally, after separating the first embedding information into popular embedding information and interest embedding information, the decoupled learning module also uses the interest embedding information and scene popularity in the first embedding information to perform a first comparative learning, thereby obtaining the first loss information.

[0103] That is, the decoupled learning module represents the positive interest embedding information as follows: The inverse interest embedding information is represented as: Furthermore, the dot product function is used to calculate the similarity between each pair of representations in each pair of interest embeddings. In addition, to ensure that the interactions between long-tail items in the interactive projects are interest-based, the decoupling learning module also obtains the scene popularity of each interactive project, and uses the first popularity weight to weight the scene popularity to obtain the first weighted popularity. This first weighted popularity is then compared with the positive interest embeddings and similarity in the interest embedding information to obtain the first loss information based on interest. The calculation method of the first loss information is as follows:

[0104]

[0105] Among them, w(pop) i ) represents the weighted popularity of interactive item i. Forward interest embedding information and reverse interest embedding information.

[0106] In the first loss information, for high-popularity items, the weighted popularity is close to w(-1) rather than 0, so that the contrastive learning task can learn interest preferences to a certain extent. By integrating the popularity signal into the contrastive loss learning of interest embedding pairs, it is possible to learn the separated interest representation directly from the interaction information.

[0107] Optionally, the decoupled contrastive learning module can also perform a second contrastive learning based on the popular embedding information and scene mask information in the first embedding information to obtain the second loss information.

[0108] That is, the decoupled contrastive learning module performs popular embedding separation on the first embedding information to obtain popular embedding information, and divides the popular embedding information into positive popular embedding information. and reverse flow embedded information Furthermore, this decoupled contrastive learning module also obtains a second popularity weight and uses this second popularity weight to weight the scene popularity of each interaction item, obtaining a second weighted popularity. It then performs contrastive learning based on positive popularity embedding information, negative popularity embedding information, and the second weighted popularity to obtain second loss information. This ensures that interactions with highly popular meme items are more attributable to conformity, thus enabling better learning to separate conformity embeddings. The sum of the first popularity weight and the second popularity weight is 1. Optionally, the formula for calculating the second loss information is as follows:

[0109]

[0110] Among them, L con For the second loss information, (1-w(pop) i () represents the second weighted popularity.

[0111] Optionally, to achieve scene decoupling through comparative learning for different scenarios, this decoupling learning module can also utilize scene masks to perform comparative learning separately for different scenarios. This restricts the calculation of the first and second loss information to items within the same scenario, reducing interference from different scenarios on the comparative learning. Optionally, the method for scene decoupling of the second loss information using scene masks is as follows:

[0112]

[0113] Among them, I imFor scene mask, L con This is the final second loss information.

[0114] Similarly, the decoupled contrastive learning module can also use this scene mask to decouple the first loss information from the scene, obtaining the decoupled first loss information. Optionally, the scene decoupling method for the first loss information is as follows:

[0115]

[0116] Among them, L int This is the first loss information after decoupling.

[0117] Optionally, the decoupled contrastive learning module can also simultaneously optimize the main task loss information, the first loss information, and the second loss information under a multi-task training strategy to obtain the target loss information. That is, the decoupled contrastive learning module can generate target loss information based on the main task loss information, the first loss information, the second loss information, and the target hyperparameters, and train the preset network model based on this target loss information to obtain the target processing model. The formula for the target loss information is as follows:

[0118] L total =L rec +λL int +μL con ,

[0119] Among them, L total For target loss information, L rec The loss information is the main task information, and λ and μ are the target hyperparameters.

[0120] The target hyperparameter is a model hyperparameter used to balance the loss information of the main task, the first loss information, and the second loss information.

[0121] Optionally, by integrating the popularity weights and loss information mentioned above, it is possible to effectively learn the separated causal embeddings directly from sparse interaction data, thereby improving the personalization and accuracy of content recommendation.

[0122] In this embodiment, the content recommendation device obtains first loss information by performing a first comparative learning based on the first embedding information and the scene popularity; it then obtains second loss information by performing a second comparative learning based on the first embedding information and the scene mask information; finally, it generates target loss information based on the main task loss information, the first loss information, the second loss information, and the target hyperparameters; and finally, it trains a preset network model based on the target loss information to obtain a target processing model. This integrates popularity weights and loss information, and uses scene masks to limit the loss information to a specified scene, thereby adjusting the popularity bias in the recommendation system to improve the personalization and accuracy of recommendations.

[0123] To better implement the content recommendation method in the embodiments of this application, based on the content recommendation method, the embodiments of this application also provide a content recommendation system, such as... Figure 5 As shown, Figure 5 This is a schematic diagram of the structure of a content recommendation system provided in an embodiment of this application. Optionally, the content recommendation system 500 includes:

[0124] The information acquisition module 501 is configured to acquire interactive information to be processed.

[0125] The embedding extraction module 502 is configured to process the interaction information to be processed based on the target scene information to obtain target embedding information;

[0126] The content recommendation module 503 is configured to determine target recommendation information based on the target embedding information.

[0127] In one possible implementation of this embodiment, the content recommendation system processes the interaction information to be processed based on the target scene information to obtain target embedding information, including:

[0128] Determine the popularity of the interactive items in the interactive information to be processed;

[0129] Generate the first embedded information corresponding to the interaction information to be processed;

[0130] Generate scene mask information for the interaction information to be processed based on the target scene information corresponding to the interaction information to be processed.

[0131] The target embedding information is determined based on the popularity of the interactive item, the first embedding information, and the scene mask information.

[0132] In one possible implementation of this embodiment, the content recommendation system generates scene mask information for the interaction information to be processed based on the target scene information corresponding to the interaction information to be processed, including:

[0133] Obtain the interaction item information from the interaction information to be processed, and the scene identifier information corresponding to the interaction item information;

[0134] Based on the scene identification information, determine the target scene information corresponding to the interactive item information;

[0135] The scene mask information corresponding to the target scene information is determined based on the mask generation module.

[0136] In one possible implementation of this embodiment, the content recommendation system generates scene mask information for the interaction information to be processed based on the target scene information corresponding to the interaction information to be processed, including:

[0137] Obtain the interaction item information from the interaction information to be processed, and the scene identifier information corresponding to the interaction item information;

[0138] Based on the scene identification information, the target scene information corresponding to the interactive item information is determined, and scene mask information corresponding to the target scene information is generated.

[0139] In one possible implementation of this embodiment, the content recommendation system determines the target scene information corresponding to the interactive item information based on the scene identifier information, and generates scene mask information corresponding to the target scene information, including:

[0140] Access the scene database to obtain candidate scene information associated with the interactive project information, as well as candidate identifier information of the candidate scene information;

[0141] Candidate scene information whose candidate identifier information is the same as the scene identifier information is determined as the target scene information of the interactive project information;

[0142] The scene mask information of the target scene information is generated based on the target scene information and the candidate scene information.

[0143] In one possible implementation of this embodiment, the content recommendation system generates scene mask information for the target scene information based on the target scene information and the candidate scene information, including:

[0144] Obtain a first scene value of the candidate scene information and a second scene value of the target scene information;

[0145] Perform a bitwise OR operation on the first scene value and the second scene value to obtain the scene mask information of the target scene information.

[0146] In one possible implementation of this embodiment, the content recommendation system determines the target embedding information based on the popularity of the interactive item, the first embedding information, and the scene mask information, including:

[0147] The scene popularity of the candidate recommendation information is calculated based on the scene mask information and the popularity of the interactive items.

[0148] Based on the popularity of the scene, the first embedding information, and the target processing model, the target embedding information of the interaction information to be processed is generated. The target embedding information includes target user embedding information and embedding information of each item.

[0149] In one possible implementation of this embodiment, the content recommendation system determines target recommendation information based on the target embedding information and outputs the target recommendation information, including:

[0150] Obtain the target user embedding information and the embedding information of each item from the target embedding information;

[0151] Calculate the embedding similarity information between each item embedding information and the target user embedding information;

[0152] Based on the embedding similarity information, the target project embedding information in the project embedding information is determined, and the target recommendation information corresponding to the target project embedding information is obtained.

[0153] In one possible implementation of this embodiment, before the content recommendation system generates the target embedding information of the interaction information to be processed based on the scene popularity, the first embedding information, and the target processing model, it further includes:

[0154] A first comparative learning is performed based on the first embedded information and the scene popularity to obtain first loss information;

[0155] A second comparison learning is performed based on the first embedded information and the scene mask information to obtain second loss information;

[0156] Target loss information is generated based on the main task loss information, the first loss information, the second loss information, and the target hyperparameters.

[0157] The target processing model is obtained by training a preset network model based on the target loss information.

[0158] In one possible implementation of this embodiment, the content recommendation system performs a first comparative learning based on the first embedding information and the scene popularity to obtain first loss information, including:

[0159] The first embedding information is subjected to interest embedding separation to obtain interest embedding information, which includes forward interest embedding information and reverse interest embedding information.

[0160] The popularity of the scene is weighted using the first popularity weight to obtain the weighted popularity.

[0161] First contrastive learning is performed based on the positive interest embedding information, the reverse interest embedding information, and the weighted popularity to obtain first loss information.

[0162] In one possible implementation of this embodiment, the content recommendation system performs a second comparative learning based on the first embedding information and the scene mask information to obtain second loss information, including:

[0163] The first embedding information is subjected to popular embedding separation to obtain popular embedding information, which includes forward popular embedding information and reverse popular embedding information;

[0164] The popularity of the scene is weighted using the second popularity weight to obtain the weighted popularity.

[0165] A second comparison learning is performed based on the scene mask information, the forward flow embedding information, the backward flow embedding information, and the weighted flow to obtain the second loss information.

[0166] In this embodiment, the content recommendation system acquires interaction information to be processed; processes the interaction information to be processed according to target scene information to obtain target embedding information; determines target recommendation information based on the target embedding information, and outputs the target recommendation information. This achieves the goal of acquiring interaction information to be processed and using a scene mask corresponding to the interaction information, a causal embedding module, and a decoupled contrastive learning module to identify interest-based causal embeddings and conformity-based causal embeddings in the interaction information to obtain target embedding information that separates interest-based causal relationships and popularity-based causal relationships in the interaction information. That is, the target embedding information can consider the influence of scene popularity and interest as much as possible to adjust the popularity bias in the recommendation system. Based on the target embedding information, the system outputs the inference results of the corresponding recommendation task and generates target recommendation information, thereby improving the personalization and accuracy of the content recommendation process to meet users' content recommendation needs.

[0167] This invention also provides a content recommendation device, such as... Figure 6 As shown, Figure 6 This is a schematic diagram of one embodiment of the content recommendation device provided in this application.

[0168] The content recommendation device integrates any one of the content recommendation systems provided in the embodiments of the present invention, and the content recommendation device includes:

[0169] One or more processors;

[0170] Memory; and

[0171] One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor in the steps of the content recommendation method in any of the embodiments described above.

[0172] Specifically, a content recommendation device may include components such as a processor 601 with one or more processing cores, a memory 602 with one or more computer-readable storage media, a power supply 603, and an input unit 604. Those skilled in the art will understand that... Figure 6 The content recommendation device structure shown does not constitute a limitation on the content recommendation device. It may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0173] The processor 601 is the control center of the content recommendation device. It connects various parts of the device via various interfaces and lines, and performs various functions and processes data by running or executing software programs and / or modules stored in the memory 602, and by calling data stored in the memory 602, thereby providing overall monitoring of the content recommendation device. Optionally, the processor 601 may include one or more processing cores; preferably, the processor 601 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 601.

[0174] The memory 602 can be used to store software programs and modules. The processor 601 executes various functional applications and data processing by running the software programs and modules stored in the memory 602. The memory 602 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, video playback function, etc.), etc.; the data storage area may store data created based on content recommendations for device usage, etc. In addition, the memory 602 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 602 may also include a memory controller to provide the processor 601 with access to the memory 602.

[0175] The content recommendation device also includes a power supply 603 that supplies power to the various components. Preferably, the power supply 603 can be logically connected to the processor 601 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 603 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0176] The content recommendation device may also include an input unit 604, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0177] Although not shown, the content recommendation device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 601 in the content recommendation device loads the executable files corresponding to the processes of one or more applications into the memory 602 according to the following instructions, and the processor 601 runs the applications stored in the memory 602 to realize various functions, as follows:

[0178] Obtain the interaction information to be processed;

[0179] The interaction information to be processed is processed based on the target scene information to obtain target embedding information;

[0180] Based on the target embedding information, target recommendation information is determined and output.

[0181] In this embodiment, the content recommendation device acquires interaction information to be processed; processes the interaction information to be processed according to target scene information to obtain target embedding information; determines target recommendation information based on the target embedding information, and outputs the target recommendation information. This achieves the goal of acquiring interaction information to be processed and using a scene mask corresponding to the interaction information, a causal embedding module, and a decoupled contrastive learning module to identify interest-based causal embeddings and conformity-based causal embeddings in the interaction information to obtain target embedding information that separates interest-based causal relationships and popularity-based causal relationships in the interaction information. That is, the target embedding information can consider the influence of scene popularity and interest as much as possible to adjust the popularity bias in the recommendation system. Based on the target embedding information, the device outputs the inference results of the corresponding recommendation task and generates target recommendation information, thereby improving the personalization and accuracy of the content recommendation process to meet users' content recommendation needs.

[0182] Therefore, embodiments of the present invention provide a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc. A computer program is stored thereon, which is loaded by a processor to execute the steps in any of the content recommendation methods provided in the embodiments of the present invention. For example, the computer program loaded by the processor may execute the following steps:

[0183] Obtain the interaction information to be processed;

[0184] The interaction information to be processed is processed based on the target scene information to obtain target embedding information;

[0185] Based on the target embedding information, target recommendation information is determined and output.

[0186] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the detailed descriptions of other embodiments above, which will not be repeated here.

[0187] In practice, each of the above units or structures can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units or structures, please refer to the previous method embodiments, which will not be repeated here.

[0188] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0189] The above provides a detailed description of a content recommendation method provided by the embodiments of this application. Specific embodiments have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method, characterized in that, The content recommendation method includes: Obtain the interaction information to be processed; The interaction information to be processed is processed based on the target scene information to obtain target embedding information; Based on the target embedding information, target recommendation information is determined.

2. The method according to claim 1, characterized in that, The step of processing the interaction information to be processed based on the target scene information to obtain target embedding information includes: Determine the popularity of the interactive items in the interactive information to be processed; Generate the first embedded information corresponding to the interaction information to be processed; Generate scene mask information for the interaction information to be processed based on the target scene information corresponding to the interaction information to be processed. The target embedding information is determined based on the popularity of the interactive item, the first embedding information, and the scene mask information.

3. The method according to claim 2, characterized in that, The step of generating scene mask information for the interaction information to be processed based on the target scene information corresponding to the interaction information to be processed includes: Obtain the interaction item information from the interaction information to be processed, and the scene identifier information corresponding to the interaction item information; Based on the scene identification information, the target scene information corresponding to the interactive item information is determined, and scene mask information corresponding to the target scene information is generated.

4. The method according to claim 3, characterized in that, The step of determining the target scene information corresponding to the interactive item information based on the scene identifier information and generating scene mask information corresponding to the target scene information includes: Access the scene database to obtain candidate scene information associated with the interactive project information, as well as candidate identifier information of the candidate scene information; Candidate scene information whose candidate identifier information is the same as the scene identifier information is determined as the target scene information of the interactive project information; The scene mask information of the target scene information is generated based on the target scene information and the candidate scene information.

5. The method according to claim 4, characterized in that, The step of generating scene mask information for the target scene information based on the target scene information and the candidate scene information includes: Obtain a first scene value of the candidate scene information and a second scene value of the target scene information; Perform a bitwise OR operation on the first scene value and the second scene value to obtain the scene mask information of the target scene information.

6. The method according to claim 2, characterized in that, Determining the target embedding information based on the popularity of the interactive item, the first embedding information, and the scene mask information includes: The scene popularity of candidate recommendation information is calculated based on the scene mask information and the popularity of the interactive items. The target embedding information of the interaction information to be processed is generated based on the popularity of the scene, the first embedding information, and the target processing model.

7. The method according to any one of claims 1-6, characterized in that, The step of determining target recommendation information based on the target embedding information includes: Obtain the target user embedding information and the embedding information of each item from the target embedding information; Calculate the embedding similarity information between each item embedding information and the target user embedding information; Based on the embedding similarity information, the target project embedding information in the project embedding information is determined, and the target recommendation information corresponding to the target project embedding information is obtained.

8. The method according to claim 6, characterized in that, Before generating the target embedding information of the interaction information to be processed based on the scene popularity, the first embedding information, and the target processing model, the method further includes: A first comparative learning is performed based on the first embedded information and the scene popularity to obtain first loss information; A second comparison learning is performed based on the first embedded information and the scene mask information to obtain second loss information; Target loss information is generated based on the main task loss information, the first loss information, the second loss information, and the target hyperparameters. The parameters of the current network model are adjusted based on the target loss information to obtain the target processing model.

9. The method according to claim 8, characterized in that, The step of performing a first comparative learning based on the first embedding information and the scene popularity to obtain first loss information includes: Perform a first separation operation on the first embedded information to obtain the interest embedded information in the first embedded information; The popularity of the scene is weighted using the first popularity weight to obtain the first weighted popularity; First contrastive learning is performed based on the interest embedding information and the weighted popularity to obtain first loss information.

10. The method according to claim 8, characterized in that, The step of performing a second comparative learning based on the first embedded information and the scene mask information to obtain second loss information includes: Perform popular embedding separation on the first embedding information to obtain popular embedding information in the first embedding information; The popularity of the scene is weighted using the second popularity weight to obtain the second weighted popularity. A second comparison learning is performed based on the scene mask information, the popular embedding information, and the second weighted popularity to obtain the second loss information.

11. A system, characterized in that, The system includes: The information acquisition module is configured to acquire interactive information to be processed. The embedding extraction module is configured to process the interaction information to be processed based on the target scene information to obtain the target embedding information; The content recommendation module is configured to determine target recommendation information based on the target embedding information; Further, the step of processing the interaction information to be processed based on the target scene information to obtain target embedding information includes: Determine the popularity of the interactive items in the interactive information to be processed; Generate the first embedded information corresponding to the interaction information to be processed; Generate scene mask information for the interaction information to be processed based on the target scene information corresponding to the interaction information to be processed. The target embedding information is determined based on the popularity of the interactive item, the first embedding information, and the scene mask information; Further, the step of generating scene mask information for the interaction information to be processed based on the target scene information corresponding to the interaction information to be processed includes: Obtain the interaction item information from the interaction information to be processed, and the scene identifier information corresponding to the interaction item information; Based on the scene identification information, determine the target scene information corresponding to the interactive item information, and generate scene mask information corresponding to the target scene information; Further, the step of determining the target scene information corresponding to the interactive item information based on the scene identifier information, and generating scene mask information corresponding to the target scene information, includes: Access the scene database to obtain candidate scene information associated with the interactive project information, as well as candidate identifier information of the candidate scene information; Candidate scene information whose candidate identifier information is the same as the scene identifier information is determined as the target scene information of the interactive project information; Generate scene mask information for the target scene based on the target scene information and the candidate scene information; Further, the step of generating scene mask information for the target scene information based on the target scene information and the candidate scene information includes: Obtain a first scene value of the candidate scene information and a second scene value of the target scene information; Perform a bitwise OR operation on the first scene value and the second scene value to obtain the scene mask information of the target scene information; Further, determining the target embedding information based on the popularity of the interactive item, the first embedding information, and the scene mask information includes: The scene popularity of candidate recommendation information is calculated based on the scene mask information and the popularity of the interactive items. Based on the scene popularity, the first embedding information, and the target processing model, the target embedding information of the interaction information to be processed is generated. Further, determining the target recommendation information based on the target embedding information includes: Obtain the target user embedding information and the embedding information of each item from the target embedding information; Calculate the embedding similarity information between each item embedding information and the target user embedding information; Based on the embedding similarity information, the target project embedding information in the project embedding information is determined, and the target recommendation information corresponding to the target project embedding information is obtained; Furthermore, before generating the target embedding information of the interaction information to be processed based on the scene popularity, the first embedding information, and the target processing model, the method further includes: A first comparative learning is performed based on the first embedded information and the scene popularity to obtain first loss information; A second comparison learning is performed based on the first embedded information and the scene mask information to obtain second loss information; Target loss information is generated based on the main task loss information, the first loss information, the second loss information, and the target hyperparameters. Based on the target loss information, the parameters of the current network model are adjusted to obtain the target processing model; Further, the step of performing a first comparative learning based on the first embedding information and the scene popularity to obtain first loss information includes: Perform a first separation operation on the first embedded information to obtain the interest embedded information in the first embedded information; The popularity of the scene is weighted using the first popularity weight to obtain the first weighted popularity; A first comparison learning is performed based on the interest embedding information and the weighted popularity to obtain first loss information; Further, the step of performing a second comparative learning based on the first embedding information and the scene mask information to obtain second loss information includes: Perform popular embedding separation on the first embedding information to obtain popular embedding information in the first embedding information; The popularity of the scene is weighted using the second popularity weight to obtain the second weighted popularity. A second comparison learning is performed based on the scene mask information, the popular embedding information, and the second weighted popularity to obtain the second loss information.

12. A device, characterized in that, The content recommendation devices include: One or more processors; Memory; and One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the steps of the method of any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, It contains a computer program that is loaded by a processor to perform the steps of the method according to any one of claims 1 to 10.

Images