New media intelligent propagation and distribution method and system based on AI

Abstract

The invention discloses an AI-based new media intelligent propagation and distribution method and system, and relates to the technical field of news recommendation. The method comprises the steps of determining dominant interests of a target user by obtaining historical browsing records of the target user; neighbor users are determined according to the historical browsing records, and possible hidden interests of the target user are determined according to the browsing records of the neighbor users; interest characteristics of the user are described by combining dominant interests and implicit interests of the user, so that related media information is matched. The possible interest range of the user is expanded by introducing behavior data of the neighbor user, and dynamic interest modeling and content feature matching are combined, so that recommended content is more accurate and diversified, common limitations in a personalized recommendation system are avoided, and recommendation accuracy and user experience are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of information recommendation, and in particular to a method and system for new media intelligent dissemination and distribution based on AI. Background Art

[0002] With the popularization of the Internet and the surge in the amount of information, personalized recommendation systems have emerged to help users filter out content related to their interests from massive amounts of information. This change has greatly improved the efficiency of users' information acquisition, but it has also brought some side effects. The core idea of ​​the personalized recommendation system is to provide customized content recommendations to users based on their historical behaviors, preferences, and interaction records. However, this approach will gradually narrow the diversity and breadth of user-perceived content. Users may only be exposed to content that they already know and are interested in, gradually forming an information cocoon effect, limiting their exposure to information in different fields, which will result in feedback data that cannot effectively express the dynamic changes in user interests, limiting the recommendation system's mining of user interests, resulting in one-sided user interest modeling, making it difficult to fully capture user interests, and resulting in insufficient accuracy in pushed content, which in turn affects the user stickiness and activity of the platform. Summary of the invention

[0003] The purpose of the present invention is to solve the problem of insufficient accuracy of pushed content mentioned in the above background technology, and to propose a method and system for new media intelligent dissemination and distribution based on AI.

[0004] The first aspect of the present invention provides a method for intelligent communication and distribution of new media based on AI, the method comprising: Obtaining historical browsing records of the target user, and determining a set of neighbor users based on the historical browsing records; a neighbor user refers to a user whose browsing records have a non-empty intersection with the target user; Obtain the text content and category label of the target candidate news; Calling a pre-trained artificial intelligence model to calculate the attractiveness score of the target candidate news to the target user; the artificial intelligence model includes an interest representation module, a content representation module and a score calculation module; wherein: the interest representation module is used to determine the interest feature vector of the target user based on the historical browsing records of the target user and its neighboring users; the content representation module is used to determine the content feature vector of the candidate news based on the text content and category label of the candidate news; the score calculation module is used to calculate the attractiveness score based on the interest feature vector and the content feature vector; According to the attractiveness scores of multiple candidate news, a news recommendation list for the target user is generated.

[0005] Preferably, the content representation module includes a text content embedding module, a category label embedding module and a fusion module; wherein: The text content embedding module is used to vectorize the text content of the news to obtain a first feature; use a convolutional neural network to extract the first feature, capture the context information of the text, and obtain a second feature; use an attention network to perform weighted calculation on the second feature to obtain an embedding vector of the text content; The category label embedding module is used to vectorize the category label of the news to obtain a third feature; perform a nonlinear transformation operation on the third feature to obtain an embedding vector of the category label; The fusion module is used to fuse the embedding vectors of text content and category labels using the attention mechanism to obtain the content feature vector of the target candidate news.

[0006] Preferably, the interest representation module includes an explicit interest perception module, an implicit interest perception module and an integration module; The explicit interest perception module is used to extract text content features and category label features from the target user's historical browsing records using an attention mechanism, and fuse the two types of features to obtain the explicit interest features of the target user; The implicit interest perception module is used to integrate the explicit interest features of multiple neighbor users to obtain the implicit interest features of the target user; The integration module is used to concatenate and linearly transform the explicit interest features and implicit interest features of the target user to obtain the interest feature vector of the target user.

[0007] Preferably, the explicit interest perception module includes a content interest perception module, a tag interest perception module and an aggregation module; wherein: The content interest perception module is used to encode the text content of multiple historical news in the browsing history to obtain multiple text vectors, and perform weighted summation on the multiple text vectors to obtain content interest representation; The tag interest perception module is used to encode the category tags of multiple historical news in the browsing history to obtain multiple tag vectors; jointly calculate each tag vector and the content interest representation to obtain the attention weight of each tag vector; and perform weighted fusion on multiple tag vectors according to the attention weight of each tag vector to obtain the tag interest representation; The aggregation module is used to concatenate and linearly transform the content interest representation and the tag interest representation to obtain explicit interest features.

[0008] Preferably, the implicit interest perception module includes a neighbor interest acquisition module and a joint calculation module; wherein: The neighbor interest acquisition module is used to acquire explicit interest features of multiple neighbor users, which are recorded as neighbor interest features; The joint calculation module is used to jointly calculate each neighbor interest feature and the explicit interest feature of the target user to obtain the attention weight of each neighbor interest feature; according to the attention weight of each neighbor interest feature, multiple neighbor interest features are weightedly integrated to obtain the implicit interest feature of the target user.

[0009] A second aspect of the present invention provides a system for intelligent communication and distribution of new media based on AI, the system comprising: A user data acquisition module is used to acquire the historical browsing records of the target user and determine a neighbor user set based on the historical browsing records; a neighbor user refers to a user who has a non-empty intersection with the target user in browsing records; A news data acquisition module is used to obtain the text content and category labels of target candidate news; The intelligent matching module is used to call the pre-trained artificial intelligence model to calculate the attractiveness score of the target candidate news to the target user; the artificial intelligence model includes an interest representation module, a content representation module and a score calculation module; wherein: the interest representation module is used to determine the interest feature vector of the target user based on the historical browsing records of the target user and its neighboring users; the content representation module is used to determine the content feature vector of the candidate news based on the text content and category label of the candidate news; the score calculation module is used to calculate the attractiveness score based on the interest feature vector and the content feature vector; The list generation module is used to generate a news recommendation list for the target user based on the attractiveness scores of multiple candidate news.

[0010] Preferably, the content representation module includes a text content embedding module, a category label embedding module and a fusion module; wherein: The text content embedding module is used to vectorize the text content of the news to obtain a first feature; use a convolutional neural network to extract the first feature, capture the context information of the text, and obtain a second feature; use an attention network to perform weighted calculation on the second feature to obtain an embedding vector of the text content; The category label embedding module is used to vectorize the category label of the news to obtain a third feature; perform a nonlinear transformation operation on the third feature to obtain an embedding vector of the category label; The fusion module is used to fuse the embedding vectors of text content and category labels using the attention mechanism to obtain the content feature vector of the target candidate news.

[0011] Preferably, the interest representation module includes an explicit interest perception module, an implicit interest perception module and an integration module; The explicit interest perception module is used to extract text content features and category label features from the target user's historical browsing records using an attention mechanism, and fuse the two types of features to obtain the explicit interest features of the target user; The implicit interest perception module is used to integrate the explicit interest features of multiple neighbor users to obtain the implicit interest features of the target user; The integration module is used to concatenate and linearly transform the explicit interest features and implicit interest features of the target user to obtain the interest feature vector of the target user.

[0012] Preferably, the explicit interest perception module includes a content interest perception module, a tag interest perception module and an aggregation module; wherein: The content interest perception module is used to encode the text content of multiple historical news in the browsing history to obtain multiple text vectors, and perform weighted summation on the multiple text vectors to obtain content interest representation; The tag interest perception module is used to encode the category tags of multiple historical news in the browsing history to obtain multiple tag vectors; jointly calculate each tag vector and the content interest representation to obtain the attention weight of each tag vector; and perform weighted fusion on multiple tag vectors according to the attention weight of each tag vector to obtain the tag interest representation; The aggregation module is used to concatenate and linearly transform the content interest representation and the tag interest representation to obtain explicit interest features.

[0013] Preferably, the implicit interest perception module includes a neighbor interest acquisition module and a joint calculation module; wherein: The neighbor interest acquisition module is used to acquire explicit interest features of multiple neighbor users, which are recorded as neighbor interest features; The joint calculation module is used to jointly calculate each neighbor interest feature and the explicit interest feature of the target user to obtain the attention weight of each neighbor interest feature; according to the attention weight of each neighbor interest feature, multiple neighbor interest features are weightedly integrated to obtain the implicit interest feature of the target user.

[0014] Beneficial effects of the present invention: The present invention proposes a method for intelligent communication and distribution of new media based on AI, which includes: obtaining historical browsing records of a target user, and determining a neighbor user set according to the historical browsing records; a neighbor user refers to a user having a non-empty intersection with the target user in browsing records; obtaining text content and category labels of target candidate news; calling a pre-trained artificial intelligence model to calculate the attractiveness score of the target candidate news to the target user; generating a news recommendation list for the target user according to the attractiveness scores of multiple candidate news; the artificial intelligence model includes an interest representation module, a content representation module and a score calculation module; wherein: the interest representation module is used to determine the interest feature vector of the target user according to the historical browsing records of the target user and its neighbor users; the content representation module is used to determine the content feature vector of the candidate news according to the text content and category labels of the candidate news; the score calculation module is used to calculate the attractiveness score according to the interest feature and the content feature.

[0015] By introducing the behavioral data of neighbor users to expand the user's possible interests, and combining dynamic interest modeling with content feature matching, the recommended content is made more accurate and diversified, avoiding the common limitations of personalized recommendation systems and improving the accuracy of recommendations and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the accompanying drawings.

[0017] Figure 1 A flowchart of a method for intelligent communication and distribution of new media based on AI is provided for an embodiment of the present invention; Figure 2 An architecture diagram of an artificial intelligence model is provided for an embodiment of the present invention; Figure 3 An architecture diagram of a system for intelligent communication and distribution of new media based on AI is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The embodiment of the present invention provides a method for intelligent communication and distribution of new media based on AI. Figure 1 , Figure 1 A flowchart of a method for intelligent communication and distribution of new media based on AI provided by an embodiment of the present invention. The method comprises the following steps: S101, obtaining historical browsing records of a target user, and determining a neighbor user set according to the historical browsing records.

[0020] S102, obtaining text content and category labels of target candidate news.

[0021] S103, calling a pre-trained artificial intelligence model to calculate the attractiveness score of the target candidate news to the target user.

[0022] S104, generating a news recommendation list for the target user according to the attractiveness scores of the plurality of candidate news.

[0023] Among them, neighbor users refer to users who have non-empty intersections with the target user in browsing records; the artificial intelligence model includes an interest representation module, a content representation module and a score calculation module; among them: the interest representation module is used to determine the interest feature vector of the target user based on the historical browsing records of the target user and its neighbor users; the content representation module is used to determine the content feature vector of the candidate news based on the text content and category labels of the candidate news; the score calculation module is used to calculate the attractiveness score based on the interest features and content features.

[0024] An AI-based new media intelligent dissemination and distribution method provided in an embodiment of the present invention expands the user's possible interest range by introducing the behavioral data of neighbor users, and combines dynamic interest modeling with content feature matching to make recommended content more accurate and diversified, avoiding common limitations in personalized recommendation systems, and improving the accuracy of recommendations and user experience.

[0025] In one implementation, different users can be divided into different clusters based on the interest tags set by the users and their geographic locations. For the cluster where the target user is located, a global bipartite graph is generated based on the browsing records of all users (users click on news). Then, the neighbor nodes of the user nodes are the news nodes, and the neighbor nodes of the news nodes are the user nodes. , and its browsing history , find each historical news through the adjacency matrix of the global bipartite graph A set of neighbor nodes, from which K users are selected to form the neighbor user set of target user u .

[0026] In one embodiment, see Figure 2 , Figure 2 An architectural diagram of an artificial intelligence model provided for an embodiment of the present invention.

[0027] like Figure 2 As shown, the content representation module includes a text content embedding module, a category label embedding module and a fusion module; wherein: The specific processing process of the text content embedding module is as follows: Step 1: vectorize the text content of the news to obtain the first feature. This step can use word embedding (such as Word2Vec, GloVe, BERT, etc.) to convert the text content into a word vector sequence to obtain the first feature . There are at most LW words.

[0028] Step 2: Use convolutional neural network to extract the first feature, capture the context information of the text, and obtain the second feature Specific: , in, is the output of the i-th convolution kernel, is the weight parameter of the i-th convolution kernel; So Concatenate the word vectors within the starting window; is the bias term of the i-th convolution kernel; ReLU is the activation function; there are LK convolution kernels.

[0029] Step 3: Use the attention network to perform weighted calculation on the second feature to obtain the embedding vector of the text content. The specific process includes: Step 1: Calculate the attention weight of each element of the second feature ;in, , and is a learnable parameter; tanh is an activation function. Step 2: According to the attention weight of each element, use the softmax function to calculate the weight coefficient ;Step 3: weighted summation to obtain the embedding vector CE of the text content: .

[0030] The specific processing process of the category label embedding module is as follows: Step 1: vectorize the category labels of the news to obtain the third feature. In this step, word embedding can be used to vectorize the category labels to obtain the third feature. . There are at most LM tags.

[0031] Step 2: Perform a nonlinear transformation on the third feature to obtain the embedding vector TE of the category label: ; and are learnable parameters.

[0032] The specific processing process of the fusion module is: using the attention mechanism to fuse the embedding vectors of the text content and the category label to obtain the content feature vector of the target candidate news. Specifically, after merging the embedding vectors of the text content and the category label, a temporary feature is obtained; according to the above process of using the attention network to perform weighted calculation on the second feature, similarly, the temporary feature is used as the role of the second feature, and weighted calculation is performed, and the result is used as the content feature vector of the target candidate news.

[0033] In one implementation, through the synergy of the text content embedding module and the category label embedding module, the text semantic features and category information of the news are extracted simultaneously, so that the news content feature representation is more comprehensive and the accuracy of the recommendation is improved.

[0034] In one embodiment, the interest representation module includes an explicit interest perception module, an implicit interest perception module and an integration module.

[0035] The explicit interest perception module is used to extract text content features and category label features from the target user's historical browsing records using the attention mechanism, and fuse the two types of features to obtain the explicit interest features of the target user. Specifically, the explicit interest perception module includes a content interest perception module, a label interest perception module, and an aggregation module; among which: The specific processing process of the content interest perception module is: encoding the text content of multiple historical news in the browsing history to obtain multiple text vectors, and performing weighted summation on the multiple text vectors to obtain content interest representation.

[0036] The specific processing process of the tag interest perception module is as follows: Step 1: Encode the category labels of multiple historical news in the browsing records to obtain multiple label vectors.

[0037] Step 2: Combine each label vector and content interest representation to obtain the attention weight of each label vector ; is the attention weight of the i-th label vector; tanh is the activation function; , , , is a learnable parameter; It is a content interest expression; is the i-th label vector.

[0038] Step 3: According to the attention weight of each label vector, multiple label vectors are weighted and fused to obtain the label interest representation; specifically, the attention weight is normalized using the softmax function to obtain the weight coefficient of each label vector, and then the multiple label vectors are weighted and summed to obtain the label interest representation. .

[0039] The specific processing process of the aggregation module is: concatenate and linearly transform the content interest representation and the tag interest representation to obtain the explicit interest feature ; is a learnable parameter, and concat represents serial concatenation.

[0040] In one implementation, during the explicit interest modeling process, by jointly calculating category labels and content, the model not only considers the specific content that users read, but also captures their preferred news categories, thereby improving the generalization ability of interest modeling.

[0041] The implicit interest perception module is used to integrate the explicit interest features of multiple neighbor users to obtain the implicit interest features of the target user. Specifically, the implicit interest perception module includes a neighbor interest acquisition module and a joint calculation module; wherein: The neighbor interest acquisition module is used to obtain explicit interest features of multiple neighbor users, which are recorded as neighbor interest features.

[0042] The specific processing process of the joint calculation module is as follows: Step 1: Combine the interest features of each neighbor and the explicit interest features of the target user to obtain the attention weight of each neighbor interest feature. ; is the attention weight of the i-th neighbor user; , , , is a learnable parameter; It is the explicit interest characteristics of the target users; is the explicit interest feature of the i-th neighbor user.

[0043] Step 2: According to the attention weight of each neighbor interest feature, multiple neighbor interest features are weighted and fused to obtain the implicit interest features of the target user. Specifically, the attention weight is normalized using the softmax function to obtain the weight coefficient of each neighbor interest feature, and then the weighted sum of multiple neighbor interest features is performed to obtain the implicit interest features of the target user. .

[0044] The specific processing process of the integration module is: splicing and linearly transforming the explicit interest features and implicit interest features of the target user to obtain the interest feature vector of the target user ; are learnable parameters.

[0045] In one implementation, the interest features of neighbor users are used to supplement the interest modeling of the target user, making up for the problem of insufficient data or interest drift when relying only on explicit interests, comprehensively characterizing the user's personalized interests and improving the accuracy of interest representation.

[0046] In one embodiment, the score calculation module may calculate the dot product or cosine similarity based on the interest feature vector and the content feature vector as the attractiveness score.

[0047] The embodiment of the present invention provides a system for intelligent communication and distribution of new media based on AI. Figure 3 , Figure 3 The following is an architecture diagram of a system for AI-based new media intelligent communication and distribution provided in an embodiment of the present invention. The system includes: The user data acquisition module is used to obtain the historical browsing records of the target user and determine the neighbor user set based on the historical browsing records. The neighbor user refers to the user who has a non-empty intersection with the target user in browsing records.

[0048] The news data acquisition module is used to obtain the text content and category labels of the target candidate news.

[0049] The intelligent matching module is used to call the pre-trained artificial intelligence model to calculate the attractiveness score of the target candidate news to the target user.

[0050] The list generation module is used to generate a news recommendation list for the target user based on the attractiveness scores of multiple candidate news.

[0051] An AI-based new media intelligent communication and distribution system provided based on an embodiment of the present invention expands the user's possible interest range by introducing the behavioral data of neighbor users, and combines dynamic interest modeling with content feature matching to make recommended content more accurate and diversified, avoiding common limitations in personalized recommendation systems and improving the accuracy of recommendations and user experience.

[0052] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for intelligent communication and distribution of new media based on AI, characterized in that: The method comprises: Obtaining historical browsing records of the target user, and determining a set of neighbor users based on the historical browsing records; a neighbor user refers to a user whose browsing records have a non-empty intersection with the target user; Obtain the text content and category label of the target candidate news; Calling a pre-trained artificial intelligence model to calculate the attractiveness score of the target candidate news to the target user; the artificial intelligence model includes an interest representation module, a content representation module and a score calculation module; wherein: the interest representation module is used to determine the interest feature vector of the target user based on the historical browsing records of the target user and its neighboring users; the content representation module is used to determine the content feature vector of the candidate news based on the text content and category label of the candidate news; the score calculation module is used to calculate the attractiveness score based on the interest feature vector and the content feature vector; According to the attractiveness scores of multiple candidate news, a news recommendation list for the target user is generated.

2. According to claim 1, a method for new media intelligent communication and distribution based on AI is characterized in that: The content representation module includes a text content embedding module, a category label embedding module and a fusion module; wherein: The text content embedding module is used to vectorize the text content of the news to obtain a first feature; use a convolutional neural network to extract the first feature, capture the context information of the text, and obtain a second feature; use an attention network to perform weighted calculation on the second feature to obtain an embedding vector of the text content; The category label embedding module is used to vectorize the category label of the news to obtain a third feature; perform a nonlinear transformation operation on the third feature to obtain an embedding vector of the category label; The fusion module is used to fuse the embedding vectors of text content and category labels using the attention mechanism to obtain the content feature vector of the target candidate news.

3. The method of new media intelligent communication and distribution based on AI according to claim 1, characterized in that: The interest representation module includes an explicit interest perception module, an implicit interest perception module and an integration module; The explicit interest perception module is used to extract text content features and category label features from the target user's historical browsing records using an attention mechanism, and fuse the two types of features to obtain the explicit interest features of the target user; The implicit interest perception module is used to integrate the explicit interest features of multiple neighbor users to obtain the implicit interest features of the target user; The integration module is used to concatenate and linearly transform the explicit interest features and implicit interest features of the target user to obtain the interest feature vector of the target user.

4. The method of new media intelligent communication and distribution based on AI according to claim 3, characterized in that: The explicit interest perception module includes a content interest perception module, a tag interest perception module and an aggregation module; wherein: The content interest perception module is used to encode the text content of multiple historical news in the browsing history to obtain multiple text vectors, and perform weighted summation on the multiple text vectors to obtain content interest representation; The tag interest perception module is used to encode the category tags of multiple historical news in the browsing history to obtain multiple tag vectors; jointly calculate each tag vector and the content interest representation to obtain the attention weight of each tag vector; and perform weighted fusion on multiple tag vectors according to the attention weight of each tag vector to obtain the tag interest representation; The aggregation module is used to concatenate and linearly transform the content interest representation and the tag interest representation to obtain explicit interest features.

5. The method of new media intelligent communication and distribution based on AI according to claim 3 is characterized in that: The implicit interest perception module includes a neighbor interest acquisition module and a joint calculation module; wherein: The neighbor interest acquisition module is used to acquire explicit interest features of multiple neighbor users, which are recorded as neighbor interest features; The joint calculation module is used to jointly calculate each neighbor interest feature and the explicit interest feature of the target user to obtain the attention weight of each neighbor interest feature; according to the attention weight of each neighbor interest feature, multiple neighbor interest features are weightedly integrated to obtain the implicit interest feature of the target user.

6. A new media intelligent communication and distribution system based on AI, characterized in that: The system comprises: A user data acquisition module is used to acquire the historical browsing records of the target user and determine a neighbor user set based on the historical browsing records; a neighbor user refers to a user who has a non-empty intersection with the target user in browsing records; A news data acquisition module is used to obtain the text content and category labels of target candidate news; The intelligent matching module is used to call the pre-trained artificial intelligence model to calculate the attractiveness score of the target candidate news to the target user; the artificial intelligence model includes an interest representation module, a content representation module and a score calculation module; wherein: the interest representation module is used to determine the interest feature vector of the target user based on the historical browsing records of the target user and its neighbor users; the content representation module is used to determine the content feature vector of the candidate news based on the text content and category label of the candidate news; the score calculation module is used to calculate the attractiveness score based on the interest feature vector and the content feature vector; The list generation module is used to generate a news recommendation list for the target user based on the attractiveness scores of multiple candidate news.

7. The AI-based new media intelligent communication and distribution system according to claim 6 is characterized in that: The content representation module includes a text content embedding module, a category label embedding module and a fusion module; wherein: The text content embedding module is used to vectorize the text content of the news to obtain a first feature; use a convolutional neural network to extract the first feature, capture the context information of the text, and obtain a second feature; use an attention network to perform weighted calculation on the second feature to obtain an embedding vector of the text content; The category label embedding module is used to vectorize the category label of the news to obtain a third feature; perform a nonlinear transformation operation on the third feature to obtain an embedding vector of the category label; The fusion module is used to fuse the embedding vectors of text content and category labels using the attention mechanism to obtain the content feature vector of the target candidate news.

8. The AI-based new media intelligent communication and distribution system according to claim 6 is characterized in that: The interest representation module includes an explicit interest perception module, an implicit interest perception module and an integration module; The explicit interest perception module is used to extract text content features and category label features from the target user's historical browsing records using an attention mechanism, and fuse the two types of features to obtain the explicit interest features of the target user; The implicit interest perception module is used to integrate the explicit interest features of multiple neighbor users to obtain the implicit interest features of the target user; The integration module is used to concatenate and linearly transform the explicit interest features and implicit interest features of the target user to obtain the interest feature vector of the target user.

9. The AI-based new media intelligent communication and distribution system according to claim 8, characterized in that: The explicit interest perception module includes a content interest perception module, a tag interest perception module and an aggregation module; wherein: The content interest perception module is used to encode the text content of multiple historical news in the browsing history to obtain multiple text vectors, and perform weighted summation on the multiple text vectors to obtain content interest representation; The tag interest perception module is used to encode the category tags of multiple historical news in the browsing history to obtain multiple tag vectors; jointly calculate each tag vector and the content interest representation to obtain the attention weight of each tag vector; and perform weighted fusion on multiple tag vectors according to the attention weight of each tag vector to obtain the tag interest representation; The aggregation module is used to concatenate and linearly transform the content interest representation and the tag interest representation to obtain explicit interest features.

10. The AI-based new media intelligent communication and distribution system according to claim 8, characterized in that: The implicit interest perception module includes a neighbor interest acquisition module and a joint calculation module; wherein: The neighbor interest acquisition module is used to acquire explicit interest features of multiple neighbor users, which are recorded as neighbor interest features; The joint calculation module is used to jointly calculate each neighbor interest feature and the explicit interest feature of the target user to obtain the attention weight of each neighbor interest feature; according to the attention weight of each neighbor interest feature, multiple neighbor interest features are weightedly integrated to obtain the implicit interest feature of the target user.

Images