Material recall and neural network training method, device, equipment and storage medium

Abstract

The application provides a material recall method and device, a neural network training method and device, an electronic device, and a computer readable storage medium. The material recall method comprises: obtaining user basic features and user behavior features of a target user, wherein the user basic features are used to represent attributes of the target user, and the user behavior features are used to represent materials interacted by the target user; inputting the user behavior features into a multi-interest extraction network for feature extraction to obtain a plurality of user interest features of the target user, wherein the plurality of user interest features are used to represent a plurality of interests of the target user; inputting the user basic features and the plurality of user interest features into a plurality of neural networks to obtain user multi-interest features of the target user; and performing material recall for the target user based on the user multi-interest features. Through the application, the accuracy of material recall can be improved.

Description

Technical Field

[0001] This application relates to the field of recommendation system technology, and in particular to a material recall method and apparatus, a neural network training method and apparatus, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Currently, recommendation systems can be used in online social networking scenarios to recommend materials that users may be interested in. In particular, for microblogging platforms like Weibo, recommendation systems can be used to recommend posts related to users' interests.

[0003] Users' interests are often diverse. For example, a user may have multiple interests, including entertainment, astrology, current events, sports and fitness, reading, and music. Therefore, a user might interact with blog posts in categories such as entertainment, astrology, current events, sports and fitness, reading, and music.

[0004] Therefore, it is necessary to accurately recommend materials that match users' diverse interests. Summary of the Invention

[0005] This application provides a material recall method and apparatus, a neural network training method and apparatus, an electronic device, and a computer-readable storage medium, which can improve the accuracy of material recall.

[0006] The technical solution of this application is implemented as follows:

[0007] In a first aspect, this application provides a material recall method. The material recall method includes: obtaining basic user characteristics and user behavior characteristics of a target user, wherein the basic user characteristics are used to characterize the attributes of the target user, and the user behavior characteristics are used to characterize the materials interacted with by the target user; inputting the user behavior characteristics into a multi-interest extraction network for feature extraction to obtain multiple user interest features of the target user, wherein the multiple user interest features represent multiple interests of the target user; inputting the basic user characteristics and the multiple user interest features into multiple neural networks to obtain the target user's multi-interest features; and performing material recall for the target user based on the multi-interest features.

[0008] In a second aspect, this application provides a neural network training method. The method includes: acquiring basic user features and user behavior features of sample users, as well as material features of sample materials, wherein the basic user features are used to characterize the attributes of the sample users, and the user behavior features are used to characterize the materials interacted with by the sample users; inputting the user behavior features into a multi-interest extraction network for feature extraction to obtain multiple user interest features of the sample users, wherein the multiple user interest features represent multiple interests of the sample users; inputting the basic user features and the multiple user interest features into multiple neural networks to obtain the multi-interest features of the sample users; and adjusting the multi-interest extraction network and / or the multiple neural networks based on the multi-interest features of the sample users and the material features of the sample materials.

[0009] In a third aspect, this application provides a material recall device. The material recall device includes: a first acquisition module configured to acquire basic user characteristics and user behavior characteristics of a target user, wherein the basic user characteristics are used to characterize the attributes of the target user, and the user behavior characteristics are used to characterize the materials interacted with by the target user; a first extraction module configured to input the user behavior characteristics into a multi-interest extraction network for feature extraction to obtain multiple user interest features of the target user, wherein the multiple user interest features represent multiple interests of the target user; a first processing module configured to input the basic user characteristics and the multiple user interest features into multiple neural networks to obtain the target user's multi-interest features; and a recall module configured to perform material recall for the target user based on the multi-interest features.

[0010] In a fourth aspect, this application provides a neural network training apparatus. The neural network training apparatus includes: a second acquisition module configured to acquire basic user features and user behavior features of a sample user, and material features of sample materials, wherein the basic user features of the sample user are used to characterize the attributes of the sample user, and the user behavior features of the sample user are used to characterize the materials interacted with by the sample user; a second extraction module configured to input the user behavior features into a multi-interest extraction network for feature extraction to obtain multiple user interest features of the sample user, wherein the multiple user interest features are used to represent multiple interests of the sample user; a second processing module configured to input the basic user features and the multiple user interest features into multiple neural networks to obtain the multi-interest features of the sample user; and an adjustment module configured to adjust the multi-interest extraction network and / or the multiple neural networks based on the multi-interest features of the sample user and the material features of the sample materials.

[0011] In a fifth aspect, this application provides an electronic device. The electronic device includes a memory and a processor. The memory is configured to store executable instructions. The processor is configured to execute the executable instructions stored in the memory to implement the material recall method as described in the first aspect, or the neural network training method as described in the second aspect.

[0012] In a sixth aspect, this application provides a computer-readable storage medium. The computer-readable storage medium stores executable instructions. These executable instructions, when executed by a processor, implement the material recall method as described in the first aspect, or the neural network training method as described in the second aspect.

[0013] In a seventh aspect, this application provides a computer program product. The computer program product includes a computer program or instructions. When executed by a processor, the computer program or instructions implement the material recall method as described in the first aspect, or the neural network training method as described in the second aspect.

[0014] The technical solution of this application has the following beneficial effects:

[0015] In the technical solution of this application, basic user characteristics and multiple user interest features are input into multiple neural networks. Each neural network processes the basic user characteristics and multiple user interest features to obtain corresponding multi-interest user features. In this way, since multiple user interest features are related to multiple user interests, the multi-interest user features obtained through multiple neural networks, corresponding to each of the multiple interests, can include user interest representations from different perspectives. This makes the representation of user interests more comprehensive and accurate, and based on these multi-interest user features, materials that better match user interests can be identified for material recall, thereby improving the accuracy of material recall. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the model architecture of the multi-interest recommendation model provided in the embodiments of this application.

[0017] Figure 2 This is a flowchart illustrating the material recall method provided in the embodiments of this application.

[0018] Figure 3 This is a schematic diagram illustrating the specific implementation process of the material recall method provided in the embodiments of this application.

[0019] Figure 4 This is a flowchart illustrating the neural network training method provided in the embodiments of this application.

[0020] Figure 5 This is a schematic diagram illustrating the specific implementation process of the neural network training method provided in the embodiments of this application.

[0021] Figure 6 This is a schematic diagram of the material recall device provided in the embodiments of this application.

[0022] Figure 7 This is a schematic diagram of the structure of the neural network training device provided in the embodiments of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0025] If the application documents contain similar descriptions such as "first / second", the following explanation shall be added: In the following description, the terms "first / second / third" are used only to distinguish similar objects and do not represent a specific order of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0027] Currently, personalized recommendation systems can be used in online social networking scenarios to recommend content that users are interested in. In particular, for microblogging platforms like Weibo, recommendation systems can be used to recommend microblogs related to a user's interests. Specifically, recommendation systems can recommend blog posts (i.e., content) that users are likely to be interested in based on their basic profile and historical interaction behavior.

[0028] Recommendation systems mainly consist of two stages: recall and ranking. In the recall stage, a recall algorithm identifies candidate items from the item pool that are relevant to the user's interests. In the ranking stage, a ranking algorithm scores each candidate item, and based on the scores, the highest-scoring items are recommended to the user.

[0029] Users' interests are often diverse. For example, a user might have multiple interests, including entertainment, astrology, current events, sports and fitness, reading, and music. Therefore, a user might interact with blog posts from multiple categories, such as entertainment, astrology, current events, sports and fitness, reading, and music.

[0030] Currently, some recommendation systems can recall blog posts of the same category based on a user's single interest representation. However, this cannot meet the needs of users with diverse interests.

[0031] Furthermore, some recommendation systems can recall blog posts of different categories based on a user's single interest representation. However, the accuracy of such recommendation systems in recalling materials based on users' diverse interests needs improvement.

[0032] Therefore, it is necessary to accurately recommend materials that match users' diverse interests.

[0033] Therefore, this application proposes a multi-interest recommendation model. Figure 1 This is a schematic diagram of the model architecture of the multi-interest recommendation model provided in the embodiments of this application.

[0034] like Figure 1 As shown, the multi-interest recommender model includes: a multi-interest extractor layer on the user side, a fusion layer, k neural networks, a fusion layer and a fully connected layer, a neural network on the material side, a multiplier in the service module, and a label-aware attention network in the training module. It is understood that k is an integer greater than 1.

[0035] The multi-interest recommendation model includes two types of inputs on the user side: user attributes and interactive materials. Basic user characteristics are obtained based on user attributes. User behavioral characteristics are obtained based on interactive materials. On the material side, the multi-interest recommendation model takes materials as input. Basic material characteristics are obtained based on these materials.

[0036] The multi-interest extraction layer is used to extract multiple interests from multiple user behavioral features, resulting in n interest features. These interest features can correspond to the user's n interests. It's understood that n is an integer greater than 1.

[0037] The fusion layer is used to fuse each of the n interest features with the user's basic features.

[0038] The k neural networks are distinct from each other. On one hand, they can be of the same type but with different parameters. On the other hand, they can be of different types. For each feature obtained by fusing interest features with user base features, the k neural networks process the fused feature separately. Because the k neural networks are different, the resulting k features all correspond to the interest features related to the fused feature, and can represent the relationship between the user and the corresponding interest at different depths and from different perspectives.

[0039] The fusion layer and fully connected layer consist of a fusion layer and a fully connected layer. The fusion layer is used to fuse the k features obtained from passing each interest feature through k neural networks. The fully connected layer is used to process the fused features.

[0040] The neural network on the material side is used to process the basic characteristics of the material to obtain the material features.

[0041] In the service module, the features output by the fusion layer and the fully connected layer on the user side are multiplied with the materials in the material library. Based on the result of the multiplication operation, candidate materials are determined in the material library, and the candidate materials are used to realize material recall.

[0042] In the training module, based on the features output by the fusion layer and fully connected layer on the user side and the features output by the neural network on the material side, the loss can be obtained by calculating the weights of the attention network using labels. This loss can then be used as a basis for training the multi-interest recommendation model.

[0043] Based on the above multi-interest recommendation model, this application provides a material recall method and apparatus, a neural network training method and apparatus, an electronic device, and a computer-readable storage medium, which can improve the accuracy of material recall.

[0044] This application provides a material recall method. The material recall method can be implemented using a combination of terminal devices and a server, or it can be implemented using only a server. The terminal device runs a terminal application, and the server provides support for the terminal application. The terminal device and the server are connected and communicate wirelessly or via a wired connection.

[0045] Material recall methods can begin with a user request from the target user. For example, a user request can be initiated by the target user through a proactive action on a terminal device. Alternatively, a user request can be initiated automatically by a terminal application on the terminal device based on preset conditions.

[0046] Figure 2 This is a schematic flowchart of the material recall method provided in an embodiment of this application. Figure 2 As shown, the material recall method in this application embodiment includes S210 to S240.

[0047] In S210, the basic user characteristics and user behavior characteristics of the target user are obtained.

[0048] Here, basic user characteristics are used to characterize the attributes of the target user, and user behavior characteristics are used to characterize the materials that the target user interacts with.

[0049] Understandably, the target users are those who need recommended content. For example, the target users could be users of the Weibo app. The content could be blog posts or other types of information.

[0050] User basic characteristics are derived from the target user's attribute information. The target user's attribute information refers to the target user's own attributes. For example, attribute information may include at least one of the following: social attributes and interests / preferences. Social attributes may include username, gender, age, device ID, location, frequently used applications, etc. Interest / preference attribute information may include the target user's interest tags. A target user may have multiple interests, and therefore may simultaneously possess multiple interest tags. Interest tags can be single-level or multi-level tags. For example, a target user's interest tags related to astrology may be divided into three levels: first-level tag "astrology and fortune telling," second-level tag "zodiac sign," and third-level tag "Aquarius." Another example is a target user's interest tags related to sports, which may be divided into three levels: first-level tag "fitness and exercise," second-level tag "running," and third-level tag "marathon." Yet another example is a target user's interest tags related to history, which may be divided into four levels: first-level tag "history," second-level tag "Chinese history," third-level tag "Han Dynasty history," and fourth-level tag "the reigns of Emperors Wen and Jing." Finally, a target user's interest tags related to entertainment may be divided into two levels: first-level tag "celebrity," and second-level tag "a specific celebrity." In one embodiment, the basic characteristics of the target user can be obtained by encoding and pooling all the attribute information of the target user. For example, the pooling process for the attribute information can be flat pooling, max pooling, or other types of pooling.

[0051] User behavior characteristics are derived from the target user's historical interaction behavior. Specifically, a target user may interact with one or more materials within a certain period. Target users can interact with materials in various forms. For example, a target user's interaction with a material might be clicking on it. Another example is browsing the material or browsing it for a certain duration. Yet another example is liking or saving the material. To obtain the target user's historical interaction behavior, it is necessary to acquire data on the materials the target user interacted with over a certain period. For example, historical interaction behavior can target materials interacted with by the target user over periods of months, weeks, days, hours, etc. Each material in the historical interaction behavior has a content tag. Content tags can be single-level or multi-level tags. For example, a material related to constellations might have three levels of content tags: a first-level tag for music, a second-level tag for classical music, and a third-level tag for Beethoven. All content tags of each material interacted with by the target user within a certain period can constitute an interaction sequence. In one embodiment, user behavior characteristics may include the tag characteristics of the materials interacted with by the target user. For example, by encoding and pooling all the tags in the interaction sequence corresponding to each material interacting with the target user, the user behavior characteristics of the target user can be obtained. For example, the pooling process for the interaction sequence can be flat pooling, max pooling, or other types of pooling. Thus, for each material interacting with, corresponding user behavior characteristics can be obtained. Since user behavior characteristics are obtained based on the material's tags, user behavior characteristics can be considered as the material's tag characteristics.

[0052] In practical applications, after obtaining data on materials interacted with by target users over a period of time, the obtained data can be "cleaned." For example, the number of interactions between the target user and each material can be counted, and materials with interaction counts exceeding a certain value can be filtered out. Another example is the number of occurrences of each tag on the materials interacted with by the target user, and materials corresponding to tags with occurrence counts below a certain value can be filtered out. It should be noted that the above "data cleaning" can also be implemented in other ways, and this application embodiment does not specifically limit this approach.

[0053] It is understood that user basic characteristics and user behavior characteristics can be obtained by the terminal device or the server. In one embodiment, user basic characteristics can be obtained by the terminal device and sent to the server. In this case, the terminal device obtains the target user's attribute information and obtains the user basic characteristics based on the attribute information. In another embodiment, user basic characteristics can be obtained by the server. In this case, the terminal device can obtain the target user's attribute information and send the attribute information to the server. Then, the server obtains the user basic characteristics based on the attribute information. In one embodiment, user behavior characteristics can be obtained by the terminal device and sent to the server. In this case, the terminal device obtains the target user's interaction sequence and obtains the user behavior characteristics based on the interaction sequence. In yet another embodiment, user behavior characteristics can be obtained by the server. For example, the terminal device can obtain the target user's interaction sequence and send the interaction sequence to the server. Then, the server obtains the user behavior characteristics based on the interaction sequence. Alternatively, the server can obtain the target user's interaction sequence itself based on the target user's historical interaction behavior and obtain the user behavior characteristics based on the interaction sequence.

[0054] It should be noted that the user basic characteristics and user behavior characteristics obtained in S210 can be in vector form. Specifically, the user basic characteristics are vectors obtained based on all attribute information of the target user. Furthermore, the user behavior characteristics corresponding to each interacting material are vectors obtained based on all content tags of that material.

[0055] Furthermore, it should be noted that the target user's basic characteristics and behavioral characteristics can be obtained before or after a user request is made. In some cases, the basic characteristics and behavioral characteristics can be obtained in advance and stored on the terminal device or server, and then retrieved directly from the terminal device or server in response to a user request. In other cases, the basic characteristics and behavioral characteristics can be obtained in response to a user request.

[0056] In S220, user behavior features are input into a multi-interest extraction network for feature extraction, resulting in multiple user interest features of the target user.

[0057] Here, multiple user interest features are used to represent the multiple interests of the target user.

[0058] Specifically, for all user behavior features obtained in S210, the multi-interest network clusters these user behavior features to obtain multiple user interest features. These multiple user interest features are used to represent the diverse interests of the target user. For example, for materials interacting with by the target user, the multiple user interest features obtained through the multi-interest extraction network can correspond to different interests, such as social, music, and entertainment.

[0059] Through S220, the multi-interest extraction network can identify multiple interests related to materials interacting with the target user, and aggregate the information contained in the user behavior characteristics according to different interests.

[0060] It should be noted that the user interest features obtained in S220 can be in vector form.

[0061] In S230, the user's basic features and multiple user interest features are input into multiple neural networks to obtain the target user's multi-interest features.

[0062] Here, based on the basic user features obtained in S210 and the multiple user interest features obtained in S220, multiple user interest features can be obtained through multiple neural networks.

[0063] In one embodiment, S230 may include the following two steps.

[0064] In the first step, the user's basic characteristics are fused with multiple user interest characteristics to obtain multiple user fused interest characteristics.

[0065] Since multiple user interest features correspond to multiple interests, the recall of materials for each interest considers user attributes, fusing basic user characteristics with each user interest feature separately. Therefore, the resulting fused user interest feature simultaneously represents both user attributes and interests.

[0066] In one embodiment, the fusion of user basic features and user interest features is achieved through a fusion layer. The fusion layer is used to concatenate user basic features and user interest features to obtain user fused interest features.

[0067] In the second step, each user fusion interest feature from the multiple user fusion interest features is input into multiple neural networks to obtain the target user's multi-interest features.

[0068] Here, multiple neural networks can be the same or different types. Each user's fused interest features are processed by multiple neural networks, resulting in multiple features that can represent the user's attributes and interests from different depths and / or perspectives.

[0069] In one embodiment, the number of neural networks in S230 can be two. For example, the two neural networks are a deep neural network (DNN) and a self-attention network.

[0070] Understandably, the type and number of neural networks can be set according to specific needs. Obviously, the more neural networks there are, the more features can be obtained by fusing interest features for a user, and the more comprehensive and accurate the representation of user attributes and interests will be. However, the computational load will also be greater.

[0071] In practical applications, the processing of user-fused interest features by multiple neural networks can be parallel or sequential. For example, to shorten processing time, user-fused interest features can be processed synchronously in multiple neural networks. Alternatively, to reduce computational resource consumption, user-fused interest features can be processed sequentially by multiple neural networks.

[0072] Furthermore, in one embodiment, after the user's fused interest features are processed by multiple neural networks, the resulting multiple features can be processed by a fusion layer and a fully connected layer to obtain a user interest representation corresponding to the fused interest features. Specifically, the multiple features are first input into the fusion layer for fusion; then they are input into the fully connected layer for processing to achieve further feature extraction. Here, the fully connected layer can be implemented by a dense layer.

[0073] Thus, by fusing all user interest features, we can obtain multi-interest features for each user. It should be noted that the multi-interest features obtained in S230 can be in vector form.

[0074] In S240, material recall is performed for target users based on users' multiple interest characteristics.

[0075] In practical applications, a material library can include a large number of materials and corresponding content tags. In this embodiment, the materials in the material library can also be referred to as materials to be recalled. The material library can be stored on a server. For example, the material library can be stored in the database area of ​​the server. Based on the materials and content tags in the material library, the basic characteristics of each material can be obtained. Therefore, in this embodiment, the basic characteristics of the material can also be referred to as tag features. In one embodiment, the basic characteristics of the material can be obtained by encoding and pooling all the content tags of each material. The basic characteristics of the material can also be stored in the material library and associated with the corresponding material. It is understood that the basic characteristics of the materials in the material library can be updated. In one embodiment, the basic characteristics of the material can be updated in real time. In one embodiment, the basic characteristics of the material can be updated periodically. For example, the material characteristics can be updated daily, weekly, or monthly. For example, when a new material is added to the material library, the basic characteristics of the material are immediately generated based on the new material and its content tags. As another example, when a material is deleted from the material library, the basic characteristics of the material corresponding to that material are immediately deleted. For example, when the content label of a material in the material library changes, new basic characteristics of the material are immediately generated based on the changed label.

[0076] Furthermore, based on the basic material characteristics of each material in the material library, the material characteristics of that material can be obtained through a neural network. It is understood that the material characteristics can also be updated. For example, as the materials in the material library change, or as the material labels change, the basic material characteristics change accordingly, and the corresponding material characteristics change accordingly as well. In one embodiment, the material characteristics can be updated in real time. In another embodiment, the material characteristics can be updated periodically.

[0077] Therefore, in S240, the user's multi-interest features can be compared with the material features of the materials in the material library, and the material corresponding to the material feature with the highest matching degree can be determined as the candidate material based on the comparison result. These candidate materials are used for material recall.

[0078] In one embodiment, the user's multi-interest features may include multiple user interest representations corresponding to multiple interests respectively. Since user fused interest features are obtained for each of the multiple interests in S230, each user fused interest feature can obtain a corresponding user interest representation after processing by the subsequent fusion layer and fully connected layer. For multiple interest features, multiple user interest representations can be obtained. After the processing in S230, a user interest representation can represent a user attribute and an interest related to the target user.

[0079] In one embodiment, S240 may include the following two steps.

[0080] In the first step, for each user interest representation among multiple user interest representations, the similarity between each user interest representation and the material characteristics of multiple materials to be recalled is determined, and the first candidate material corresponding to each user interest representation is determined based on the similarity.

[0081] It is understood that by calculating the similarity between each user interest representation and the material features of multiple materials, at least one material with the highest similarity to that user interest representation can be identified. The identified at least one material is then designated as the first candidate material. In one embodiment, the multiple materials used to calculate the similarity can be some or all of the materials in a material library.

[0082] In one embodiment, the similarity between user interest representations and material features is represented by the inner product between user interest representations and material features.

[0083] It is understandable that the number of first-choice materials identified can be the same for each user interest representation. In other cases, the number of first-choice materials identified can be different for each user interest representation.

[0084] In the second step, for multiple user interest representations, based on multiple first candidate materials corresponding to the multiple user interest representations, second candidate materials are determined as the materials to be recalled.

[0085] After determining the first candidate material for each user interest representation in the first step, the first candidate materials for all user interest representations are integrated to obtain the second candidate material, which is then used for material recall.

[0086] In one embodiment, a second candidate material can be obtained based on a first candidate material by employing a suitable strategy. For example, since there may be overlap in the first candidate materials corresponding to different user interest representations, the strategy may include counting the frequency of occurrence of each material in the first candidate materials. Based on the statistical results, the material with the most occurrences can be selected as the second candidate material. Another example is that the first candidate materials corresponding to different user interest representations can be scored and ranked. Based on the ranking results, the material with the highest score can be selected as the second candidate material. Of course, other strategies can also be used to obtain the second candidate material, and this embodiment does not specifically limit these methods.

[0087] Understandably, in some cases, the determined second candidate materials may not include the first candidate materials corresponding to a certain user interest representation. For example, if all materials in the first candidate materials corresponding to a certain user interest representation have low frequency or low scores, then none of those first candidate materials may be selected as second candidate materials. In other cases, when determining the second candidate materials, it can be ensured that at least one material in the first candidate materials corresponding to each user interest representation is included.

[0088] In practical applications, the quantity of materials in the material library can be on the order of hundreds of thousands, millions, or tens of millions. The quantity of the first candidate materials determined for each user's interest in the first step can be on the order of hundreds, thousands, or tens of thousands. The quantity of the second candidate materials determined in the second step can be on the order of tens, hundreds, or thousands. This application embodiment does not specifically limit this.

[0089] It should be noted that S210, S220, S230 and S240 mentioned above can be implemented on a server.

[0090] Figure 3 This is a schematic diagram illustrating the specific implementation process of the material recall method provided in this application. See also... Figure 3 The specific implementation process shown is an exemplary description of the material recall method according to the embodiments of this application. It is understood that this specific implementation process can be an online service process.

[0091] like Figure 3 As shown, the process of this specific embodiment mainly includes the following three parts: 310, 320 and 330.

[0092] In 310, the material characteristics are obtained.

[0093] First, when a new material is added to the material library, its content tag can be determined. Based on the determined content tag, the material's basic characteristics can be obtained through processes such as encoding and pooling. These basic characteristics can be stored in the material library along with the material itself. Alternatively, the basic characteristics can be stored independently of the material library.

[0094] Next, based on the multi-interest recommendation model, the material characteristics are predicted. Specifically, the obtained basic material characteristics are input into the material-side neural network of the multi-interest recommendation model to obtain the material characteristics.

[0095] It is understood that the process of obtaining the basic characteristics and material features of new materials can be performed in advance. For example, whenever a new material is added to the material library, the corresponding basic characteristics and material features are generated and stored. This process can be independent of material recall. In this case, S310 can be used as a pre-step of the material recall method. Of course, in some cases, the generation process of the basic characteristics and / or material features can also be performed after receiving a user request; this embodiment of the application does not specifically limit this.

[0096] In 320, user multi-interest characteristics were obtained.

[0097] First, the terminal application on the terminal device, either under the action of the target user or by sending a user request independently, requests a material recall. The user request can be sent from the terminal device to the server. The server can then be used to provide online material recall services.

[0098] In response to a user request, the server can obtain the target user's basic user characteristics and user behavior characteristics. For example, the server can receive basic user characteristics and user behavior characteristics uploaded by the terminal device. Alternatively, the server can obtain basic user characteristics and user behavior characteristics locally. Basic user characteristics include the target user's gender, age, and frequently used applications. For example, basic user characteristics might indicate that the target user's basic information includes: female gender, 26 years old age, and frequently used applications 1 and 2. User behavior characteristics include the materials the target user has interacted with. For example, user behavior characteristics might indicate that the target user has interacted with materials categorized as entertainment, astrology, or social topics.

[0099] Subsequently, user basic features and user behavior features are input into the multi-interest recommendation model. Through the multi-interest extraction network, fusion layer, multiple neural networks, and fusion and fully connected layers within the multi-interest recommendation model, user multi-interest features are obtained based on the user basic features and user behavior features. These user multi-interest features include user interest representations corresponding to n interests, i.e., user interest representation 1, user interest representation 2, ..., user interest representation n. For example, user multi-interest features may include three user interest representations: user interest representation 1 corresponds to entertainment, user interest representation 2 corresponds to astrology, and user interest representation 3 corresponds to society.

[0100] In step 330, candidate materials are identified.

[0101] Here, the similarity between each user interest representation in the user's multi-interest features and the material features of multiple materials in the material library is calculated. Based on this similarity, at least one material with the highest similarity to the user interest representation can be identified as the first candidate material. For example, for each interest in entertainment, astrology, and society, the M most similar materials can be obtained based on the similarity. The value of M can be set according to actual requirements. For example, M can be one thousand, ten thousand, or other values.

[0102] It's understandable that both user interest representations and material features can be represented as vectors. Therefore, the similarity described above can be represented by the inner product of user interest representations and material features. In this case, the larger the value of the inner product, the higher the similarity; the smaller the value of the inner product, the lower the similarity.

[0103] After obtaining the first candidate materials corresponding to all user interest representations, all first candidate materials are integrated to determine N materials as second candidate materials for subsequent material recall. The value of N can be set according to actual requirements. For example, N can be one hundred, one thousand, or other values. Understandably, N is usually less than M.

[0104] Certain strategies can be employed to derive second candidate materials from the first candidate materials. For example, all first candidate materials can be statistically analyzed by their frequency of occurrence, ranked according to these frequencies, and the materials with the highest frequencies can be selected as second candidate materials. Alternatively, all first candidate materials can be scored according to certain rules, ranked according to their scores, and the materials with the highest scores can be selected as second candidate materials.

[0105] For example, for each of the three interests—entertainment, astrology, and society—ten thousand first-line candidate materials are identified. Then, these first-line candidate materials are ranked according to their frequency of occurrence, and the one thousand materials with the highest frequency are selected as second-line candidate materials.

[0106] At this point, Figure 3 The illustrated material recall method process has ended. Afterwards, the server can return the information of the identified second candidate material to the terminal device for material recall.

[0107] In addition, this application provides a neural network training method. This neural network training method is used to train the multi-interest recommendation model in this application. This neural network training method can be implemented using a combination of a terminal device and a server, or it can be implemented using only the server, or it can be implemented using only the terminal device.

[0108] Figure 4 This is a flowchart illustrating the neural network training method provided in an embodiment of this application. Figure 4 As shown, the neural network training method of this application embodiment includes S410 to S440.

[0109] In S410, the basic user characteristics and user behavior characteristics of the sample users, as well as the material characteristics of the sample materials, are obtained.

[0110] Here, the basic user characteristics of the sample users are used to characterize the attributes of the sample users, and the user behavior characteristics of the sample users are used to characterize the materials interacted with by the target users.

[0111] In one embodiment, user behavior characteristics may include the tag characteristics of the materials interacting with by the sample user, and the basic characteristics of the materials may include the tag characteristics of the sample materials.

[0112] Understandably, sample users are the users selected for training the multi-interest recommendation model. For example, target users could be users of the Weibo app. Furthermore, in practical applications, training typically requires a large number of sample users.

[0113] The acquisition of basic user characteristics and user behavior characteristics of sample users can be referenced. Figure 2 The acquisition of the target user's basic characteristics and user behavior characteristics in S210 of the material summoning method shown will not be elaborated here.

[0114] Furthermore, sample materials are materials selected for training the multi-interest recommendation model. For example, sample materials can be one or more materials selected from a material library. Alternatively, sample materials can be new materials provided for training purposes. Sample materials can be blog posts or other content. Sample materials have one or more levels of content tags. Based on the content tags, the basic characteristics of the sample materials can be obtained. In one embodiment, the basic characteristics of the sample materials can be obtained by encoding and pooling all the content tags of each sample material.

[0115] In one embodiment, the sample materials may include: materials interacted with by the sample user, and materials not interacted with by the sample user. In this embodiment, the materials interacted with by the sample user are different from the sample materials, and they may not necessarily have an inclusion relationship. That is, the materials interacted with by the sample user and the sample materials may be partially the same or completely different. For example, when the sample material is new, the materials interacted with by the sample user and the sample materials are completely different.

[0116] Understandably, the basic material characteristics of the sample users are input into a neural network for feature extraction to obtain the material characteristics of the sample users. In one embodiment, this neural network can be a deep neural network.

[0117] In S420, user behavior features are input into a multi-interest extraction network to extract multiple user interest features of the sample users.

[0118] Here, multiple user interest features are used to represent the multiple interests of the sample users.

[0119] Specifically, for all user behavior features obtained in S410, the multi-interest network clusters these user behavior features to obtain multiple user interest features. These multiple user interest features are used to represent the diverse interests of the sample users. For example, for materials interacting with by the sample users, the multiple user interest features obtained through the multi-interest extraction network can correspond to different interests, such as social, music, and entertainment.

[0120] The S420 multi-interest extraction network can identify multiple interests related to materials interacting with sample users and aggregate information contained in user behavior features according to different interests.

[0121] It should be noted that the user interest features obtained in S420 can be in vector form.

[0122] In S430, basic user features and multiple user interest features are input into multiple neural networks to obtain the multi-interest features of the sample users.

[0123] Here, based on the basic user features obtained in S410 and the multiple user interest features obtained in S420, multiple user interest features can be obtained through multiple neural networks.

[0124] In one embodiment, S430 may include the following two steps.

[0125] In the first step, the user's basic characteristics are fused with multiple user interest characteristics to obtain multiple user fused interest characteristics.

[0126] Since multiple user interest features correspond to multiple interests, user attributes are considered in order to achieve material recall for each interest. Therefore, basic user features are fused with each user interest feature separately. The resulting fused user interest feature simultaneously represents both user attributes and interests.

[0127] In one embodiment, the fusion of user basic features and user interest features is achieved through a fusion layer. The fusion layer is used to concatenate user basic features and user interest features to obtain user fused interest features.

[0128] In the second step, each user fusion interest feature from the multiple user fusion interest features is input into multiple neural networks to obtain the user multi-interest features of the sample users.

[0129] Here, multiple neural networks can be the same or different types. Each user's fused interest features are processed by multiple neural networks, resulting in multiple features that can represent the user's attributes and interests from different depths and / or perspectives.

[0130] In one embodiment, the number of neural networks in S430 can be two. For example, the two neural networks are a deep neural network (DNN) and a self-attention network.

[0131] Understandably, the type and number of neural networks can be set according to specific needs. Obviously, the more neural networks there are, the more features can be obtained by fusing interest features for a user, and the more comprehensive and accurate the representation of user attributes and interests will be. However, the computational load will also be greater.

[0132] In practical applications, the processing of user-fused interest features by multiple neural networks can be parallel or sequential. For example, to shorten processing time, user-fused interest features can be processed synchronously in multiple neural networks. Alternatively, to reduce computational resource consumption, user-fused interest features can be processed sequentially by multiple neural networks.

[0133] Furthermore, in one embodiment, after the user's fused interest features are processed by multiple neural networks, the resulting multiple features can be processed by a fusion layer and a fully connected layer to obtain a user interest representation corresponding to the fused interest features. Specifically, the multiple features are first input into the fusion layer for fusion; then they are input into the fully connected layer for processing to achieve further feature extraction. Here, the fully connected layer can be implemented by a dense layer.

[0134] Thus, by fusing all user interest features, we can obtain multi-interest features for each user. It should be noted that the multi-interest features obtained in S430 can be in vector form.

[0135] In S440, the multi-interest extraction network and / or multiple neural networks are adjusted based on the multi-interest features of sample users and the material features of sample materials.

[0136] Here, we have obtained the multi-interest features of the sample users and the material features of the sample materials. Then, based on these features, we can adjust the multi-interest extraction network and / or multiple neural networks in the multi-interest recommendation system.

[0137] In one embodiment, S440 may include the following three steps.

[0138] In the first step, the binary cross-entropy is determined based on the multi-interest characteristics of the sample users and the material characteristics of the sample materials.

[0139] In this step, based on the multi-interest characteristics of sample users and the material characteristics of sample materials, the loss is calculated using a weighted attention network with labels.

[0140] Specifically, each user interest representation in the multi-interest features of the sample users is designated as feature K (key) and feature V (value), and the material feature is designated as feature Q (query). Attention weights are determined based on features V and Q, and feature K is weighted and averaged using these weights. Then, for all user interest representations in the multi-interest features, a weighted user feature is obtained by averaging them. The inner product between the weighted user feature and the material feature is calculated as the similarity between the two features. Finally, based on the similarity, it is determined whether the sample user interacted with the sample material.

[0141] In one embodiment, the binary cross-entropy (BCE) function can be used as the loss function.

[0142] In the second step, if the binary cross-entropy does not meet the convergence condition, the parameters of the multi-interest extraction network and / or multiple neural networks are adjusted.

[0143] Specifically, the convergence condition determines whether to adjust the parameters of the multi-interest extraction network and / or multiple neural networks. That is, if the binary cross-entropy does not meet the convergence condition (e.g., the binary cross-entropy is greater than or equal to a threshold), it is determined that the multi-interest extraction network and / or multiple neural networks have not been sufficiently trained, and the parameters of the multi-interest extraction network and / or multiple neural networks are adjusted; if the binary cross-entropy meets the convergence condition (e.g., the binary cross-entropy is less than a threshold), it is determined that the multi-interest extraction network and / or multiple neural networks have been sufficiently trained, and the parameters of the multi-interest extraction network and / or multiple neural networks are no longer adjusted.

[0144] It is understood that, in addition to the multi-interest extraction network and multiple neural networks, other parts of the multi-interest recommendation system can also be adjusted, and the embodiments of this application do not specifically limit this.

[0145] It should be noted that the above S410, S420, S430, S440 and S450 can be implemented on terminal devices and / or servers.

[0146] Figure 5 This is a schematic diagram illustrating the specific implementation process of the neural network training method provided in this application. See also... Figure 5 The specific implementation process shown is an exemplary description of the neural network training method of this application embodiment. It is understood that this specific implementation process can be online training or offline training.

[0147] like Figure 5 As shown, the process of this specific embodiment mainly includes the following two parts: 510 and 520.

[0148] In step 510, obtain the training samples.

[0149] The training samples include sample users and sample materials. More specifically, the training samples include user information for sample users and material information for sample materials.

[0150] The number of sample users can be multiple. User information includes attribute information and interaction sequences. To obtain user information for the sample users, the sample users can be identified first, and then their attribute information and interaction sequences can be collected through data statistics or data mining. Attribute information includes the sample users' demographic attributes and interests. For example, demographic attributes include gender, age, and frequently used applications. Interests include the sample users' interest tags. Thus, a large amount of user information can be obtained from a large number of sample users. Interaction data can be data on materials interacted with by the sample users over a past period. For example, interaction data can include materials interacted with by the sample users in the past week. Based on the sample users' interaction data, an interaction sequence consisting of all content tags for each interacted material can be obtained.

[0151] The number of sample materials can be multiple. Sample materials can be multiple materials obtained from the material library, or they can be new materials. For all sample materials, material information is obtained. Material information includes the content tags of the sample materials.

[0152] It's understandable that the material interacted with by the user is at least one of the sample materials. Therefore, based on whether it was clicked by the sample user, the sample materials can be divided into positive samples and negative samples. As the names suggest, positive samples are those interacted with by the sample user, and negative samples are those not interacted with by the sample user. In some cases, negative samples can be considered as sample materials interacted with by other users. Clearly, the sample user interaction data is obtained based on the positive samples.

[0153] Thus, by combining the above user information and material information, training samples can be obtained.

[0154] In 520, model training is performed.

[0155] The obtained training samples are input into a multi-interest recommendation model to obtain the user multi-interest features of the sample users and the material features of the sample materials. Then, a label-weighted attention network is used to calculate the similarity between the user multi-interest vector and the sample materials, thereby determining the interaction between the sample users and the sample materials.

[0156] Subsequently, a binary cross-entropy function was used as the loss function, and the multi-interest recommendation model was trained based on a large number of sample users and sample materials. In particular, based on the calculated binary cross-entropy, the parameters of the user-side multi-interest extraction network and / or multiple neural networks in the multi-interest recommendation model were adjusted.

[0157] It should be noted that model validation is required after model training.

[0158] Model validation can be conducted based on a validation sample. The validation sample includes validation users and validation materials. It is understood that the validation users and validation materials can be obtained from the sample users and sample materials respectively, or they can be different from the sample users and sample materials.

[0159] The user's interaction with the verification material is known. Then, the user's basic characteristics and behavioral characteristics, along with the verification material's basic characteristics, are input into the multi-interest recommendation model to determine whether the user interacted with the verification material (or, in other words, whether it's a recalled material). By comparing the known interaction data with the interaction data obtained from the multi-interest recommendation model, it can be determined whether the multi-interest recommendation model has been adequately trained.

[0160] At this point, Figure 5 The process of the neural network training method shown is now complete.

[0161] Based on the same inventive concept, embodiments of this application provide a material recall device. Figure 6 This is a schematic diagram of the material recall device provided in an embodiment of this application. Figure 6As shown, the material recall device 600 includes a first acquisition module 610, a first extraction module 620, a first processing module 630, and a recall module 640. The first acquisition module 610 is configured to acquire the basic user characteristics and user behavior characteristics of the target user. The basic user characteristics are used to characterize the attributes of the target user. The user behavior characteristics are used to characterize the materials interacted with by the target user. The first extraction module 620 is configured to input the user behavior characteristics into a multi-interest extraction network for feature extraction, obtaining multiple user interest features of the user. These multiple user interest features represent multiple interests of the target user. The first processing module 630 is configured to input the basic user characteristics and the multiple user interest features into multiple neural networks, obtaining the target user's multi-interest features. The recall module 640 is configured to recall materials for the target user based on the multi-interest features.

[0162] In some possible implementations, the first processing module 620 may be specifically configured to: fuse user basic features with multiple user interest features to obtain multiple user fused interest features; and input each of the multiple user fused interest features into the neural network to obtain the target user's multiple interest features.

[0163] In some possible implementations, the user's multi-interest features may include multiple user interest representations corresponding to multiple interests respectively. The recall module 640 may be configured to: for each user interest representation among the multiple user interest representations, determine the similarity between each user interest representation and the material features of multiple materials to be recalled, and determine a first candidate material corresponding to each user interest representation based on the similarity, wherein the material features of the multiple materials to be recalled are obtained based on the tag features of the multiple materials to be recalled; for the multiple user interest representations, determine a second candidate material as the material to be recalled based on the multiple first candidate materials corresponding to the multiple user interest representations.

[0164] In some possible implementations, user behavior characteristics may include the tag characteristics of the materials that the target user interacts with.

[0165] Based on the same inventive concept, embodiments of this application provide a neural network training device. Figure 7 This is a schematic diagram of the structure of the neural network training device provided in an embodiment of this application. Figure 7As shown, the neural network training device 700 includes a second acquisition module 710, a second extraction module 720, a second processing module 730, a third extraction module 740, and an adjustment module 750. The second acquisition module 710 is configured to acquire the basic user features and user behavior features of the sample users, as well as the material features of the sample materials. The basic user features of the sample users are used to characterize the attributes of the sample users. The user behavior features of the sample users are used to characterize the materials interacted with by the sample users. The second extraction module 720 is configured to input the user behavior features into a multi-interest extraction network for feature extraction, obtaining multiple user interest features of the sample users. The multiple user interest features are used to represent the multiple interests of the sample users. The second processing module 730 is configured to input the basic user features and the multiple user interest features into multiple neural networks, obtaining the multi-interest features of the sample users. The adjustment module 740 is configured to adjust the multi-interest extraction network and / or the multiple neural networks based on the multi-interest features of the sample users and the material features of the sample materials.

[0166] In some possible implementations, the sample material may include: material that the sample user interacts with, and material that the sample user does not interact with.

[0167] In some possible implementations, the adjustment module 750 may be configured to: determine the binary cross-entropy based on the user multi-interest characteristics of the sample users and the material characteristics of the sample materials; and adjust the parameters of the multi-interest extraction network and / or multiple neural networks if the binary cross-entropy does not meet the convergence condition.

[0168] In some possible implementations, user behavior features include the tag features of the materials interacting with by the sample user, and the material features of the sample materials are obtained based on the tag features of the sample materials.

[0169] The material recall method provided in this application can be implemented by various electronic devices. For example, the material recall method can be implemented by a server alone, or by a terminal device and a server working together. For example, the server executes the material recall method alone. Another example is that the terminal device sends a request message to the server, and the server executes the material recall method based on the received request message.

[0170] In practice, a server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, and big data and artificial intelligence platforms. Terminal devices can be smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, etc., but are not limited to these. Terminal devices and servers can be connected directly or indirectly through wired or wireless communication methods; this application does not impose specific limitations on this.

[0171] In some embodiments, the terminal device or server can implement the material recall method provided in this application by running a computer program. For example, the computer program can be a native program or software module in an operating system; for instance, it can be a local application (APP), i.e., a program that needs to be installed in the operating system to run, such as the Weibo APP; it can also be a Weibo mini-program, i.e., a Weibo program that only needs to be downloaded into a browser environment to run; or it can be a Weibo mini-program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module, or plugin, and this application does not specifically limit it.

[0172] Here, we will use a terminal device as an example for explanation. The terminal device displays the APP interface on the current screen. For example, the terminal device can also display recommended materials on the current screen. The server can receive user information of the target user sent by the terminal device. Then, the server extracts the user's basic characteristics and user behavior characteristics from the user information of the target user, inputs the user's basic characteristics and user behavior characteristics into the multi-interest recommendation model, and obtains the output user's multi-interest characteristics; finally, based on the user's multi-interest characteristics of the target user and the material characteristics of multiple materials, the server recalls the corresponding materials for the target user.

[0173] A server may include at least one processor, memory, at least one network interface, and a user interface. The various components within the server are coupled together via a bus system. Understandably, the bus system is used to enable communication and connection between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0174] A processor can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, a general-purpose processor can be a microprocessor or any conventional processor.

[0175] The user interface includes one or more output devices that enable the presentation of media content, including one or more speakers and / or one or more visual displays. The user interface also includes one or more input devices, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touchscreen display, camera, other input buttons and controls.

[0176] Memory can be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. Memory may optionally include one or more storage devices physically located away from the processor.

[0177] The memory includes volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM), and volatile memory can be random access memory (RAM). The memory described in the embodiments of this application is intended to include any suitable type of memory.

[0178] In some embodiments, the memory is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0179] An operating system is a system program that handles various basic system services and performs hardware-related tasks, such as the framework layer, core library layer, and driver layer, and is used to implement various basic business functions and handle hardware-based tasks.

[0180] A network communication module for reaching other computing devices via one or more (wired or wireless) network interfaces, such as Bluetooth, Wi-Fi, and Universal Serial Bus (USB).

[0181] The presentation module enables the presentation of live content (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices (e.g., a display screen, a speaker, etc.) associated with a user interface.

[0182] An input processing module is used to detect and translate one or more user inputs or interactions from one or more input devices.

[0183] In some embodiments, the material recall device provided in this application can be implemented in software. For example, the material recall device can be software in the form of programs and plug-ins, and includes a first acquisition module, a first extraction module, a first processing module, and a recall module. These modules are logically related, and therefore can be arbitrarily combined or further divided according to the functions they implement.

[0184] In other embodiments, the apparatus provided in this application can be implemented in hardware. As an example, the material recall apparatus provided in this application can be a processor in the form of a hardware decoding processor, which is programmed to execute the material recall method provided in this application. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components. This application does not specifically limit the specific implementation of such processors.

[0185] Furthermore, the neural network training method provided in this application embodiment can also be implemented by various electronic devices. For example, the material recall method can be implemented by the server alone, or by the terminal device and the server working together. For example, the server executes the neural network training method alone. Or, for another example, the terminal device sends a request message to the server, and the server executes the neural network training method according to the received request message.

[0186] Therefore, embodiments of this application provide an electronic device. The electronic device includes a memory and a processor. The memory is configured to store executable instructions. When the processor is configured to execute the executable instructions stored in the memory, it implements the material recall method or neural network training method of embodiments of this application.

[0187] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium. This computer-readable storage medium stores executable instructions. When executed by a processor, the executable instructions implement the material recall method or neural network training method of embodiments of this application.

[0188] Based on the same inventive concept, embodiments of this application provide a computer program product. This computer program product includes a computer program or instructions. When executed by a processor, the computer program or instructions implement the material recall method or neural network training method of embodiments of this application.

[0189] In summary, based on the technical solution of this application embodiment, user basic characteristics and multiple user interest characteristics are input into multiple neural networks. Each neural network processes the user basic characteristics and multiple user interest characteristics to obtain corresponding user multi-interest characteristics. In this way, the user multi-interest characteristics corresponding to an interest obtained through multiple neural networks can include user interest representations from different perspectives. This makes the representation of user interests by user multi-interest characteristics more comprehensive and accurate. Furthermore, based on these user multi-interest characteristics, materials that better match user interests can be identified for material recall, better capturing changes in user multi-interests, thereby improving the accuracy of material recall.

[0190] Furthermore, when acquiring user interaction materials, invalid interaction data, such as materials with excessive interactions, is filtered out from a large number of interactive materials. This "cleaning" of user interaction data avoids data bias caused by a few materials with overly frequent interactions, further ensuring the accuracy of material retrieval in multi-interest scenarios.

[0191] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A method for recalling materials, characterized in that, The method includes: Acquire the basic user characteristics and user behavior characteristics of the target user, wherein the basic user characteristics are used to characterize the attributes of the target user, and the user behavior characteristics are used to characterize the materials interacted with by the target user; The user behavior features are input into a multi-interest extraction network for feature extraction to obtain multiple user interest features of the target user, wherein the multiple user interest features are used to represent multiple interests of the target user; The user's basic features and the multiple user interest features are input into multiple neural networks to obtain the target user's multiple interest features. Based on the user's multi-interest characteristics, material recall is performed for the target user; The step of inputting the user's basic features and the multiple user interest features into multiple neural networks to obtain the target user's multiple interest features includes: The basic user features are fused with each of the multiple user interest features to obtain multiple user fused interest features; wherein, the user fused interest features simultaneously represent user attributes and interests; Each of the multiple user fusion interest features is input into the multiple neural networks, and each user fusion interest feature is processed by the multiple neural networks to obtain the target user's multiple interest features. The multiple interest features include multiple user interest representations corresponding to the multiple interests. Each user fusion interest feature is processed by the multiple neural networks, and the resulting multiple features represent the user's attributes and interests from different depths and / or angles. The number of features obtained from the user fusion interest features is positively correlated with the comprehensiveness and accuracy of the representation of user attributes and interests. The user interest representation is obtained based on the processing of the multiple features through a fusion layer and a fully connected layer.

2. The method according to claim 1, characterized in that, The step of recalling materials for the target user based on the user's multi-interest characteristics includes: For each of the plurality of user interest representations, the similarity between each user interest representation and the material features of the plurality of materials to be recalled is determined, and a first candidate material corresponding to each user interest representation is determined based on the similarity, wherein the material features of the plurality of materials to be recalled are obtained based on the tag features of the plurality of materials to be recalled; For the multiple user interest representations, based on the multiple first candidate materials corresponding to the multiple user interest representations, a second candidate material is determined as the material to be recalled.

3. The method according to claim 1 or 2, characterized in that, The user behavior characteristics include the tag characteristics of the materials interacted with by the target user.

4. A neural network training method, characterized in that, The method includes: The basic user characteristics and user behavior characteristics of the sample users, as well as the material characteristics of the sample materials, are obtained. The basic user characteristics of the sample users are used to characterize the attributes of the sample users, and the user behavior characteristics of the sample users are used to characterize the materials that the sample users interact with. The user behavior features are input into a multi-interest extraction network for feature extraction to obtain multiple user interest features of the sample user, wherein the multiple user interest features are used to represent multiple interests of the sample user. The user's basic features and the multiple user interest features are input into multiple neural networks to obtain the user's multiple interest features of the sample user. Based on the user multi-interest characteristics of the sample users and the material characteristics of the sample materials, the multi-interest extraction network and / or the multiple neural networks are adjusted; The step of inputting the user's basic features and the multiple user interest features into multiple neural networks to obtain the user's multiple interest features of the sample user includes: The basic user features are fused with each of the multiple user interest features to obtain multiple user fused interest features; wherein, the user fused interest features simultaneously represent user attributes and interests; Each of the multiple user fusion interest features is input into the multiple neural networks, and each user fusion interest feature is processed by the multiple neural networks to obtain the user multi-interest features of the sample user. The user multi-interest features include multiple user interest representations corresponding to the multiple interests. Each user fusion interest feature is processed by the multiple neural networks, and the resulting multiple features represent the user attributes and interests from different depths and / or angles. The number of features obtained from the user fusion interest features is positively correlated with the comprehensiveness and accuracy of the representation of user attributes and interests. The user interest representation is obtained based on the processing of the multiple features through a fusion layer and a fully connected layer.

5. The method according to claim 4, characterized in that, The sample materials include: materials that the sample users interacted with, and materials that the sample users did not interact with.

6. The method according to claim 5, characterized in that, The adjustment of the multi-interest extraction network and / or the multiple neural networks based on the user multi-interest features of the sample users and the material features of the sample materials includes: Based on the user multi-interest characteristics of the sample users and the material characteristics of the sample materials, the binary cross-entropy is determined. If the binary cross-entropy does not meet the convergence condition, the parameters of the multi-interest extraction network and / or the multiple neural networks are adjusted.

7. The method according to any one of claims 4 to 6, characterized in that, The user behavior features include the tag features of the materials interacting with by the sample users, and the material features of the sample materials are obtained based on the tag features of the sample materials.

8. A material recall device, characterized in that, The device includes: The first acquisition module is configured to acquire the basic user characteristics and user behavior characteristics of the target user, wherein the basic user characteristics are used to characterize the attributes of the target user, and the user behavior characteristics are used to characterize the materials interacted by the target user. The first extraction module is configured to input the user behavior features into a multi-interest extraction network for feature extraction to obtain multiple user interest features of the target user, wherein the multiple user interest features are used to represent multiple interests of the target user; The first processing module is configured to input the user's basic features and the multiple user interest features into multiple neural networks to obtain the target user's multiple interest features. The recall module is configured to recall materials for the target user based on the user's multi-interest characteristics; The first processing module is configured as follows: The basic user features are fused with each of the multiple user interest features to obtain multiple user fused interest features; wherein, the user fused interest features simultaneously represent user attributes and interests; Each of the multiple user fusion interest features is input into the multiple neural networks, and each user fusion interest feature is processed by the multiple neural networks to obtain the target user's multiple interest features. The multiple interest features include multiple user interest representations corresponding to the multiple interests. Each user fusion interest feature is processed by the multiple neural networks, and the resulting multiple features represent the user's attributes and interests from different depths and / or angles. The number of features obtained from the user fusion interest features is positively correlated with the comprehensiveness and accuracy of the representation of user attributes and interests. The user interest representation is obtained based on the processing of the multiple features through a fusion layer and a fully connected layer.

9. An electronic device, characterized in that, The electronic device includes: Memory, configured to store executable instructions; When the processor is configured to execute executable instructions stored in the memory, it implements the material recall method of any one of claims 1 to 3, or the neural network training method of any one of claims 4 to 7.

10. A computer-readable storage medium storing executable instructions, characterized in that, When the executable instructions are executed by the processor, they implement the material recall method of any one of claims 1 to 3, or the neural network training method of any one of claims 4 to 7.

11. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the material recall method according to any one of claims 1 to 3, or the neural network training method according to any one of claims 4 to 7.

Images