A Personalized News Recommendation Model Method Based on Multi-Level User Interests

By using a personalized news recommendation model based on multi-level user interests, combined with graph neural networks and self-attention loss function, the problem of inaccurate user interest extraction in existing technologies is solved, and more efficient personalized news recommendation is achieved.

CN117874334BActive Publication Date: 2026-11-17PEKING UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311767601.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-21
Publication Date
2026-11-17
Estimated Expiration
2043-12-21

AI Technical Summary

Technical Problem

Existing news recommendation technologies cannot accurately extract users' long-term and fine-grained interests, resulting in poor recommendation performance and failing to effectively utilize the potential interests of users who have not clicked on news articles.

Method used

A personalized news recommendation model based on multi-level user interests is adopted, combined with graph neural networks. By modeling at the news level and entity level, and using self-attention and category-guided interest-news matching loss functions, the model captures users' long-term and fine-grained interests, and trains the recommendation model by noise contrast estimation loss.

Benefits of technology

It improves the accuracy of news recommendations, enabling a more comprehensive understanding of user interests and achieving personalized recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117874334B_ABST
    Figure CN117874334B_ABST
Patent Text Reader

Abstract

This invention discloses a personalized news recommendation model method based on multi-level user interests, comprising: extracting and encoding basic user and news information to obtain feature representations of the user's historical clicked news and candidate news; news-level modeling to mine and encode long-term and potential user interests; designing an entity-level modeling module to extract and encode fine-grained user interests; fusing the two levels of interests into an interest vector, and matching it with the user's candidate news; and training the model using training data samples to obtain a personalized news recommendation model. This invention's method can more accurately extract users' long-term and fine-grained interests, thereby effectively improving the accuracy of personalized news recommendations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of information extraction and recommendation technology based on deep learning, specifically to a personalized news recommendation model method based on multi-level user interests and graph neural networks. Background Technology

[0002] With the continuous increase in online news, users face the challenge of information overload, which brings enormous pressure and challenges to news platforms. To address this, major news platforms have generally adopted personalized news recommendation methods. These methods utilize user information modeling to provide customized news recommendations that match individual user preferences. Specifically, they extract a set of interest information about news items a user is interested in based on their click history, then integrate this information into a single-category interest vector to match or select candidate news items for recommendation.

[0003] While existing methods have effectively improved users' news reading experience, they still have some shortcomings: 1) The source of user interest information is singular, as mentioned earlier, mainly from users' browsing history, representing only a single type of interest information at the news level; 2) They fail to consider users' long-term, potential interests. Because current methods are limited to or focus on news that users have clicked on and often assign high matching scores, they neglect news that users are interested in but have not actually clicked on or viewed. In real news reading scenarios, it is common for users not to click on news they are interested in, such as repetitive content or ambiguous headlines; 3) They lack a precise and granular description of user interests. Existing methods have not yet been able to analyze user behavior more deeply, especially the intrinsic relationships between entities involved in the behavior, thus lacking the refinement and utilization of the details of user interests. In conclusion, news recommendation technology still has considerable room for exploration and improvement in the extraction, description, and utilization of user interests. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a personalized news recommendation model based on multi-level user interests. This method, based on multi-level user interest modeling and graph neural network design, can more accurately extract users' long-term and fine-grained interests, thereby effectively improving the accuracy of personalized news recommendations.

[0005] The main steps of the method of the present invention (MIPNR for short) are as follows:

[0006] Step 1: Extract and encode basic user and news information.

[0007] The input to this invention is a set of users and news related to those users, where the news includes the user's historical clicked news and candidate news. Candidate news includes clicked news and unclicked news, which are used as positive and negative samples for training, respectively. MINPR performs news-level and entity-level modeling of users based on their historical clicked news.

[0008] Design a news encoder; input the embedded text information of each news article into the news encoder to obtain the feature representation of historical clicked news articles and the feature representation of candidate news articles;

[0009] Step 2: Through news-level modeling, mine and encode long-term and potential user interests;

[0010] To take into account users' long-term interests, MINPR improves the matching score of unclicked news that users are interested in by introducing category-guided interest-news matching (CGIN matching), thereby preventing users from becoming too reliant on the news they are currently clicking.

[0011] 21) Self-attention is used as the news-level user modeling module. The feature representation of historical clicked news is used as the input of the news-level user modeling module, and multiple news-level interest vectors representing user interests are output, which represent news-level interests.

[0012] 22) For all users in the training dataset, obtain news-level interests and training labels; calculate matching scores using the feature representations of candidate news articles;

[0013] 23) Design objectives and define the loss function for the news-level user modeling module, obtain supervised labels; enhance the matching score for candidate news of interest;

[0014] The present invention defines the category-based matching loss of CGIN-Matching as follows:

[0015]

[0016] in, For category-based matching loss; In the supervision label When the value is equal to 1, it takes the value 1, which is the kth candidate news item. User u i The j-th news-level interest vector Positive samples; This represents the training set data; M is the number of historical news clicks for each user; K is the number of candidate news items for each user; s i ′ jk s represents the matching score between the j-th historical clicked news item and the k-th candidate news item for user u_i in the training set;i ′ jl This represents the matching score between the j-th historical clicked news item and the l-th candidate news item for user u_i in the training set.

[0017] Step 3: Design an entity-level modeling module and extract and encode fine-grained user interests.

[0018] For entity-level user modeling, to extract fine-grained user interests, MIPNR introduces a Local News Entity Graph (LNEG) to capture subtle relationships between news entities in implicit groups within a user's reading history. This includes:

[0019] 31) Use the large language model (ChatGPT) to extract entities from the local news entity graph; divide it into R entity groups, each entity group representing an entity-level interest;

[0020] 32) Construct a Local News Entity Graph (LNEG) to transform historical clicked news into graph nodes in the entity graph; group nodes with the same interest into the same entity group, and ensure that all entity nodes within each entity group are interconnected. Then, process entity features through entity-level attention encoding to obtain entity-level interests.

[0021] 33) Entity-level attention modeling employs a two-layer self-attention network model ψ; by obtaining the embedding of each entity, the overall entity embedding is obtained; the entity embedding is input into model ψ to obtain entity features; and a local attention mechanism is used to better capture the local relationships of entities.

[0022] Set the first layer to entity-level local attention and the second layer to global attention;

[0023] Self-attention mechanism (global attention) is defined as:

[0024]

[0025] Where Q, K, and V are the query, key, and value embedding, respectively; d is the length of the vector; and T represents the transpose.

[0026] Local attention is defined as follows:

[0027]

[0028] Entity Feature H e index Define it as I = [1, 1, ..., R, R], where I i Indicates that the i-th entity is in the i-th position. i Within a group of interests. This invention utilizes entity features H ePooling is used to obtain the entity-level interest of the r-th group. Entity-level interest matrix E e ∈R R×d Defined as [;] indicates a join operation.

[0029] Step 4: Training the overall personalized recommendation model

[0030] Finally, this invention fuses news-level and entity-level interests into an interest vector, and trains the model by matching it with the user's candidate news.

[0031] 41) Integrate interests to obtain multi-level integrated interests;

[0032] 42) During the training phase, the ranking model (i.e., the recommendation model constructed in this invention) is trained using NCE loss (Noise Contrastive Estimation Loss) to obtain the final interest vector (MFI vector) E∈R for a user after fusion. d ;

[0033] For the training dataset For all users, obtain the fused interest vector.

[0034] 43) Calculate the matching score using interest vectors and candidate news embeddings;

[0035] The candidate news items include one clicked news item and K-1 non-clicked news items, which together predict the click score (match score) of the clicked news item and the click score of the K-1 non-clicked news items;

[0036] 44) Jointly predict the matching score of a candidate positive sample news item The matching score between the news items and the K-1 candidate negative sample news items [s] i2 ,s i3 ,…,s iK The ranking model of this invention is trained using NCE loss:

[0037]

[0038] in, It is the main matching loss function; It is the training set; User u i The matching score of positive samples (i.e., clicked news items) in the candidate news; s ij User u i The matching score of the j-th negative sample (i.e., the unclicked news) in the candidate news.

[0039] 45) Final loss function Represented as:

[0040]

[0041] Where α is a hyperparameter; It is a category-based matching loss function. It is the main matching loss function. It is the final loss function.

[0042] Through steps 41) to 44), a personalized news recommendation model is trained based on the training data samples. Based on the user's profile, news that the user likes is recommended. Specifically, feature extraction is performed in step 1), and news-level and entity-level interests are extracted in steps 2) and 3), respectively. These are then fused to obtain an interest vector, which represents the user to achieve personalized recommendation model modeling. The interest vector is matched with the user's candidate news to select the news that the user is most likely to be interested in and recommend it to the user. The trained personalized news recommendation model is obtained by simultaneously training using equations (3) and (7).

[0043] Step 5: During the testing phase, this invention obtains a test dataset. MFI vectors of all users This invention only calculates the matching score s = E te (H c ) T And select the news with the highest score for recommendation.

[0044] Through the above steps, personalized news recommendations based on multi-level user interests can be achieved.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This invention proposes a personalized news recommendation model based on multi-level user interests. By modeling multi-level user interests and designing personalized news recommendations using graph neural networks, this method can more accurately extract users' long-term and fine-grained interests, thereby effectively improving the accuracy of personalized news recommendations. The technical advantages of this invention include:

[0047] (i) Improve recommendation accuracy: Through deep learning and graph neural network technologies, it is possible to extract and understand users' interests more accurately, thereby improving the accuracy of news recommendations.

[0048] (ii) Comprehensive mining of user interests: This invention can not only extract users' long-term interests, but also capture users' short-term interests and fine-grained interests, making recommendations more personalized. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart illustrating the overall process of the news recommendation method in this embodiment of the invention.

[0051] Figure 2 This is a flowchart of the personalized news recommendation modeling method based on users and graph neural networks in an embodiment of the present invention.

[0052] Figure 3 This is a schematic diagram of the CGIN matching and main matching models in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the CGIN matching tag design in an embodiment of the present invention; Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the invention is not limited in any way.

[0055] This invention proposes a personalized news recommendation model based on multi-level user interests. The personalized news recommendation method is designed based on multi-level user interest modeling and graph neural network design. This method can more accurately extract users' long-term and fine-grained interests, thereby effectively improving the accuracy of personalized news recommendations.

[0056] This invention can be used for any news recommendation that requires user modeling. The overall flowchart is as follows: Figure 1 As shown. The specific implementation steps of the method of the present invention are as follows.

[0057] Step 1: Extraction and encoding of basic user and news information: The news associated with each user includes historical clicked news and candidate news; design a news encoder; input the embedded text information of each news into the news encoder to obtain the feature representation of historical clicked news and the feature representation of candidate news;

[0058] This invention, MIPNR, is a personalized news recommendation system based on user modeling. The user modeling process is described below. Figure 2 In this invention, each user u is associated with a specific news item, including a set of historically clicked news articles. And a set of candidate news Where M and K represent the number of historically clicked news items and the number of candidate news items, respectively. In candidate news item N... c middle, These are the clicked news articles; the others are the unclicked ones. Each news article... and All of them are related to its text information and Related, such as the title. Each text message or It can contain up to L n One word.

[0059] This invention employs a two-layer self-attention model (Reference: Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30) as the news encoder φ n For news feature extraction, the self-attention mechanism is defined as follows:

[0060]

[0061] Where Q1, K1, and V1 are the feature vectors of the query, key, and value, respectively; d is the dimension of Q1, K1, and V1. This is the transpose operation; Softmax is the activation function.

[0062] This invention will separate each news item and The text information (i.e., the news headline) and Represented as feature embedding and Subsequently, the present invention embeds features and Input to news encoder φ n In the middle, average pooling is performed to obtain the feature representation H of historical clicked news. n ∈R M×d The feature representation of candidate news H c ∈R K×d .

[0063] Step 2: Long-term and potential user interest mining and encoding: Designing objectives and loss functions, and obtaining supervision labels;

[0064] To reflect users' long-term and potential interests, we propose a loss function called Category Guided Interest-News Matching (CGIN-Matching), such as... Figure 3As shown, the goal is to improve the matching score of candidate news items that users might click on in the future. In this method, we determine the matching score by analyzing historically clicked news items. and candidate news Whether they belong to the same category attribute determines whether they match, thus obtaining the supervisory label for CGIN-Matching. However, it is worth noting that in current datasets, such as the Microsoft News Dataset (MIND) (Reference: Wu F, Qiao Y, Chen JH, et al. Mind: A large-scale dataset for news recommendation [C] / / Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020:3597-3606), although categorical attributes are present, their distribution is very sparse, which may lead to the corresponding supervisory labels being equally sparse.

[0065] To address this problem, this invention fully utilizes the powerful text information extraction capabilities of Large Language Models (LLMs). For example... Figure 4 As shown, based on a preset set of news categories (such as entertainment, family, and fifty other categories), this invention uses a large language model (ChatGPT) to analyze historical clicked news. Select k most suitable category nouns Candidate News Select k most suitable category nouns This invention determines and Whether there is an intersection between them, obtain the supervision label y∈R M×K That is, if and There is an intersection between them, then y ij =1, otherwise y ij =0. For example... Figure 3 The image shown is a list of historical news clicks. Selected category noun Includes category nouns (tourism, law), and is a candidate news item. and Selected category noun and These include the category nouns "tourism" and "(tourism, law)". and and There is overlap, therefore, based on historical news clicks... News-level interests extracted and and Matching, i.e., y 11 =1 and y 12 =1.

[0066] After obtaining the supervised labels from CGIN-Matching, this invention employs commonly used self-attention as the news-level user modeling module Φ to verify its effectiveness in various personalized news recommendation scenarios. This invention uses historical clicked news N... h Feature representation H n As input, and through E n =Φ(H n Output multiple news-level interest vectors E representing user interests. n ∈R M×d .

[0067] For the training dataset This invention has garnered news-level interest from all users. and training labels like Figure 2 As shown, this invention uses the feature representation H of candidate news. c ∈R K×d Calculate CGIN-Matching score

[0068]

[0069] Wherein, the three-dimensional matrix s ′ The element s i ′ jk Indicates user u i The jth historical click news Extracted news-level interests With the kth candidate news The matching score between them.

[0070] After obtaining the matching score between news-level interests and candidate news, in order to enhance the matching score for candidate news of interest, this invention defines the CGIN-Matching loss as follows:

[0071]

[0072] in, In the supervision label When the value is equal to 1, it takes the value 1, which means the k-th candidate news. User u i The j-th news-level interest vector Positive samples.

[0073] Step 3: Extraction and Encoding of Fine-Grained User Interests

[0074] To accurately depict fine-grained user interests and capture subtle relationships between news entities in a user's reading history, we introduce a Local News Entity Graph (LNEG). Entities are phrases. For example, a user's clicked news history might contain three entity groups, each encompassing different interests. LNEG transforms entities in the clicked news history into graph nodes, grouping nodes with the same interest (e.g., family life) into the same entity group, with all entity nodes within each group interconnected. Then, MIPNR processes entity features through entity-level attention encoding to derive entity-level interests.

[0075] Specifically, L is extracted using ChatGPT. e individual entities Entities such as "A father" and "his6children" are categorized into R entity groups based on whether they belong to the same interest. Each entity group represents an entity-level interest, such as family life. Then, this invention introduces a Local News Entity Graph (LNEG) to transform entities from historically clicked news into graph nodes, such as... Figure 4 As shown, this invention represents each entity (circle) as a node, where nodes with the same interest are grouped into an entity group (square). All entity nodes within each entity group are interconnected. To enhance the association between different interests, this invention randomly adds edges between different nodes in different entity groups with probability p. For entity-level attention, this invention defines the adjacency matrix of LNEG. If node i and node j are connected, then A ij The value of A is 1; otherwise, A ij The value is 0.

[0076] Entity-level attention is also a two-layer self-attention network ψ. This invention obtains the embedding of each entity to obtain the overall entity embedding. Then, the present invention embeds the entity into h e Input into model ψ, via H e =ψ(h e ) Obtain entity features Since entities within the same group reflect the same entity-level interests, they have local relationships. This invention uses a local attention mechanism to better capture the local relationships between different entities in the historical clicked news of the same user, thereby improving the model's performance. Specifically, this invention sets the first layer as entity-level local attention and the second layer as global attention. Global attention is defined in equation (1), and local attention is defined as follows:

[0077]

[0078] Entity Feature H e index Define it as I = [1, 1, ..., R, R], where I i Indicates that the i-th entity is in the i-th position. i Within a group of interests. This invention utilizes entity features H e Pooling is used to obtain the entity-level interest of the r-th group. Entity-level interest matrix E e ∈R R×d Defined as [;] indicates a join operation.

[0079] Step 4: Training the overall model

[0080] Since the interest vectors at different levels do not reside in the same semantic space, this invention requires interest fusion. Specifically, this invention connects two levels of interests along the 0th dimension and performs pooling on the 0th dimension of the interests to obtain multi-level fused interests:

[0081] E′=[E n E e (5)

[0082]

[0083] Where W∈R d×d It is a weight matrix, b∈R d It is the bias vector. E n It's news-level interest, E e It represents entity-level interests, where R is the number of entity groups, M is the number of news articles the user has historically clicked, and E′ is the number of news items. i and E′ j Let E' represent the i-th or j-th vector, where E' is a two-dimensional matrix.

[0084] During the training phase, this invention obtains the MFI vector E∈R of a certain user through equation (6). d Then, for the training dataset For all users, this invention obtains the fused interest vector. Next, this invention utilizes these vectors E tr And candidate news embedded H c ∈R K×d To calculate the matching score s = E tr (H c ) T This invention jointly predicts the matching score of a candidate positive sample news item. The matching score between the news items and the K-1 candidate negative sample news items [s]i2 ,s i3 ,…,s iK This invention applies NCE loss to train its ranking model (recommendation model):

[0085]

[0086] in, It is the main matching loss function; It is the training set; User u i The matching score of positive samples (i.e., clicked news items) in the candidate news; s ij User u i The matching score of the j-th negative sample (i.e., the unclicked news) in the candidate news;

[0087] Combining the CGIN-Matching loss in equation (3), the final loss function of this invention is:

[0088]

[0089] Where α is a hyperparameter. It is a category-based matching loss function. It is the main matching loss function. It is the final loss function.

[0090] By learning the entire network including the aforementioned modules and loss function, this invention can obtain a personalized news recommendation model based on training data samples. During the testing phase, this invention obtains a test dataset. MFI vectors of all users This invention only calculates the matching score s = E te (H c ) T without considering s ′ And select the news with the highest score for recommendation.

[0091] It should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. A personalized news recommendation model method based on multi-level user interests, characterized in that, Includes the following steps: Step 1: Extract and encode user information and news information to obtain feature representations of the user's historical clicked news and feature representations of candidate news; Step Two: News-Level Modeling: Improving the match score of unclicked news items of interest to users through Category-Guided Interest-News Matching (CGIN), and mining and encoding long-term and potential user interests; including: 21) Design a news-level user modeling module based on the self-attention mechanism, take the feature representation of historical clicked news as the input of the news-level user modeling module, and output multiple news-level interest vectors representing user interests; 22) For all users in the training dataset, obtain news-level interests and training labels; calculate the matching score using the feature representation of candidate news. 23) Design objectives and define the loss function for the news-level user modeling module, obtain supervised labels; enhance the matching score for candidate news of interest; 231) Use a large language model to select k most suitable category nouns for historical click news, and select k most suitable category nouns for candidate news; by judging whether there is an intersection between two categories of nouns, if there is an intersection, the label value is set to 1, otherwise it is 0, thus obtaining the supervised label; 232) The CGIN-Matching matching score is calculated using the feature representation of the candidate news; 233) Define the category-based matching loss for CGIN-Matching as follows: in, For category-based matching loss; In the supervision label When the value is equal to 1, it takes the value 1, which is the kth candidate news item. User u i The j-th news-level interest vector Positive samples; This represents the training set data; M is the number of historical news clicks for each user; K is the number of candidate news items for each user; s i ′ jk User u in the training set i The matching score between the j-th historical click news item and the k-th candidate news item; s i ′ jl User u in the training set i The matching score between the j-th historical click news item and the l-th candidate news item; Step 3: Design an entity-level modeling module to extract and encode fine-grained user interests; including: 31) Use a large language model to extract entities from the local news entity graph; divide it into R entity groups, each entity group representing an entity-level interest; 32) Construct a Local News Entity Graph (LNEG) to convert historical clicked news into graph nodes in the entity graph; group nodes with the same interests into the same entity group, and connect all entity nodes in each entity group; then process entity features through entity-level attention encoding to obtain entity-level interests. 33) Entity-level attention modeling adopts a two-layer self-attention network model; by obtaining the embedding of each entity, the overall entity embedding is obtained; the entity embedding is input into the entity-level modeling model to obtain entity features; a local attention mechanism is used to capture the local relationships of entities; the first layer of self-attention is set as entity-level local attention, and the second layer of self-attention is set as global attention. Global attention is defined as: Where Q, K, and V are the query, key, and value embedding, respectively; d is the length of the vector; and T represents the transpose. Local attention is defined as: Define the index I of the entity features as: I = [1,1,…,R,R]; obtain the entity-level interest and entity-level interest matrix of the r-th group by pooling the entity features; Step 4: Merge the two levels of interest into an interest vector, and train a personalized news recommendation model based on the training data samples by matching it with the user's candidate news items; including: 41) Perform interest fusion to obtain multi-level fused interests; specifically, connect two levels of interests along the 0th dimension and perform pooling in the 0th dimension of the interests to obtain multi-level fused interests, represented as: AND ′ =[And n ;AND e ] Where W is the weight matrix and b is the bias vector; E n It's news-level interest, E e This refers to entity-level interests, where R is the number of entity groups and M is the number of news articles the user has historically clicked; E i ′ and E j ′ E ′ The i-th and j-th vectors, E ′ It is a two-dimensional matrix; 42) During the training phase, the Noise Contrast Estimation (NCE) loss function is applied to train the model and obtain the fused interest vectors of users; for all users in the training dataset, the fused interest vectors are obtained. 43) Calculate the matching score using interest vectors and candidate news embeddings, and jointly predict the click score of 1 positive news item and the click score of K-1 negative news items; the click score is the matching score. 44) Jointly predict the matching score of a candidate positive sample news item The matching score between the news items and the K-1 candidate negative sample news items [s] i2 ,s i3 ,…,s iK The model is trained using the noise contrast estimation NCE loss function, expressed as follows: in, It is the main matching loss function; User u i The matching score of positive samples in the candidate news; It is the training set; s ij User u i The matching score of the j-th negative sample in the candidate news; 45) Final loss function Represented as: Where α is a hyperparameter; Step 5: During the testing phase, obtain the MFI vectors of all users in the test dataset; calculate the matching score, and select the news with the highest score for recommendation; thus realizing personalized news recommendation based on multi-level user interests.

2. The personalized news recommendation model method based on multi-level user interests as described in claim 1, characterized in that, Step one specifically includes: Retrieve user-related news; user-related news includes the user's historical clicked news and candidate news. Based on the user's historical news clicks, the user is modeled at both the news level and the entity level. A two-layer self-attention model is used as a news encoder to extract news features; The embedded text information of the news is input into the news encoder to obtain the feature representation of historical clicked news and the feature representation of candidate news.

3. The personalized news recommendation model method based on multi-level user interests as described in claim 1, characterized in that, The large language model specifically uses ChatGPT.

4. The personalized news recommendation model method based on multi-level user interests as described in claim 3, characterized in that, In step 231), obtaining the supervision labels specifically involves: Use ChatGPT to view historical news. Select k most suitable category nouns Candidate News Select k most suitable category nouns By determining and Whether there is an intersection between them, obtain the supervision label y∈R M×K That is, if and There is an intersection between them, then y ij =1, otherwise y ij =0.

5. The personalized news recommendation model method based on multi-level user interests as described in claim 4, characterized in that, step... In 232), specifically, the feature representation H of candidate news is used. c ∈R K×d Calculate CGIN-Matching score Represented as: in, For news-level interest; s i ′ jk Indicates user u i The jth historical click news Extracted news-level interests With the kth candidate news The matching score between them.

Citation Information

Patent Citations

  • Interest activation news recommendation method and system based on multistage matching

    CN114201683A

  • Intelligent news recommendation method and system based on multi-interest characteristics of user

    CN114896510A