Method for personalized news recommendation representing user interest in time series packets

Abstract

The application discloses a personalized news recommendation method of time sequence grouping user interest representation and belongs to the technical field of news recommendation. In view of the problem that the existing news recommendation system ignores the role of the user's phased interest change in user interest modeling, first, a group feature encoder is designed by using the title, category and subcategory of news to depict the news features in the group; second, a group feature encoder is designed to learn the candidate news representation perceived between groups, the clicked news representation perceived between groups and the group feature representation perceived between groups; then, the user interest representation and the candidate news representation are obtained based on a gating network; finally, the inner product of the user interest representation and the candidate news representation is obtained to obtain the interest score. The application fuses the title, category and subcategory of news to learn the interest of the clicked news of the user, calculates the similarity between the user interest and the candidate news, improves the accuracy of news recommendation and realizes technical support in the application aspect of the personalized news recommendation system.

Description

Technical Field

[0001] This invention belongs to the field of news recommendation technology, specifically relating to a personalized news recommendation method based on time-series grouped user interest representation. Background Technology

[0002] In recent years, with the popularization of online news platforms and the convenience of online services, more and more users have shifted from traditional print newspaper reading habits to digital news reading. However, faced with a massive amount of news information every day, users often find it difficult to quickly find content that interests them. To solve the problem of information overload, news recommendation systems have emerged. These systems provide personalized services by accurately modeling user interests, thereby enhancing the user's reading experience.

[0003] Compared to traditional product recommendation systems, news recommendation systems face several unique challenges. First, news articles are highly time-sensitive. Traditional recommendation methods (such as collaborative filtering based on user or item IDs) encounter severe cold-start problems in news recommendations. Semantic modeling based on news text content can effectively capture the timeliness of news, thus providing accurate recommendations. Second, a user's news browsing history over a short period reflects their dynamic changes in interest. For example, a user might only focus on entertainment news for a period, making entertainment news their primary interest during that time. When a user has recent car-buying needs and reads related news, automotive news becomes one of their primary interests. Capturing these shifts in user interest over time is crucial for accurate personalized news recommendations.

[0004] The text content of news, such as titles, summaries, and categories, contains rich semantic information. This invention improves the quality of news recommendations by semantically modeling the text content of historical clicks and candidate news by users in time-series grouping, and mining the potential interests and preferences of users perceived by candidate news. Summary of the Invention

[0005] To address the problem that existing news recommendation systems neglect the role of users' periodic interest changes in user interest modeling, this invention provides a personalized news recommendation method based on time-series grouped user interest representation.

[0006] To achieve the above objectives, the present invention employs the following technical solutions:

[0007] Step 1: Based on the sliding time window, the news clicked by the user in history is grouped in time sequence, and the candidate news is fused in different groups. Within the same group, an in-group feature encoder is designed to model the content of news title, news category, and news subcategory, and to learn the feature representation of clicked news within the group, the feature representation of candidate news within the group, and the feature representation of group within the group.

[0008] Furthermore, the specific steps of step 1 are as follows:

[0009] Step 1.1: Divide the user's historical clicked news into G groups {S1,…,S} according to time sequence. g ,…,S G}, S G Let G be the set of historical clicked news items. Candidate news items (candi) are added sequentially to each group. The set of groups includes the historical clicked news items and candidate news items for each group, denoted as S1∪candi,…,S. g ∪candi,…,S G ∪candi;

[0010] Step 1.2: Within the g-th group, use the BERT pre-trained language model as a text encoder to process the candidate news and historical clicked news title word sequences within the group. As input, the feature representation of the words in the output titles is as follows:

[0011]

[0012] In equation (1), candi represents candidate news items within group g, hc represents all historically clicked news items within group g, N represents the number of historically clicked news items in each group, l represents the number of pre-trained words in the news title, and N l It is the l-th pre-training word of the Nth news item; i represents the i-th news item, i l This represents the l-th pre-trained word in the i-th news item;

[0013] Step 1.3: Using a convolutional neural network, analyze the j-th word w of the i-th news item in the historical clicked news hc within group g. j Feature representation Further encoding is performed, and the word feature representation of cross-semantics is denoted as:

[0014]

[0015] In equation (2), W' is the learnable parameter of the CNN filter, b' is the bias, and * is the convolution operator. It is the feature representation of the i-th news item from the word at position (jf) to (j+f);

[0016] Step 1.4: Based on the different importance levels of each word in the news headline, calculate the j-th word w of the i-th news item in the historically clicked news hc within group g. j Contribution score

[0017]

[0018] In equation (3), Vt and b t Here, q represents the weight parameter, q represents the attention query vector, and T represents the transpose operation.

[0019] Step 1.5, calculate the w of the j-th word. j Normalized attention weights of all words in a news headline for:

[0020]

[0021] In equation (4), For the j'th word of the i-th news item in group g j' The contribution score; j' represents the j'-th word; exp is the exponential function;

[0022] Step 1.6: Based on the normalized attention weights, fuse the cross-semantic word feature representations of all words in the title of the i-th news item in the historical clicked news hc within group g, and define the title feature representation. for:

[0023]

[0024] Step 1.7: Perform one-hot encoding on the news category c of the i-th news item in the historical clicked news hc within group g, and initialize its representation using a lookup table. The category feature representation of the i-th news item is obtained through a multilayer perceptron. for:

[0025]

[0026] In equation (6), V c and b c These are parameters in a multilayer perceptron with categorical features;

[0027] Step 1.8: Perform one-hot encoding on the news subcategory sc of the i-th news item in the historical clicked news hc within group g, and represent it using a lookup table. The sub-category feature representation of the i-th news item is obtained through a multilayer perceptron. for:

[0028]

[0029] In equation (7), V s and b s These are parameters in a multilayer perceptron for sub-category features;

[0030] Step 1.9: Concatenate the title feature representation, category feature representation, and subcategory feature representation of the i-th historical clicked news item hc within group g, and then perform cross-fusion using a multilayer perceptron to obtain the feature representation of the i-th clicked news item within the group. for:

[0031]

[0032] In equation (8), || represents the splicing operation, W n b is the learnable parameter of the fully connected layer for news features. n For biasing the fully connected layer;

[0033] Step 1.10, repeat steps 1.3-1.9, model the candidate news features within group g, represented as r. g(candi) ;

[0034] Step 1.11: News category is a feature of grouping. Based on the different importance of news categories in group g, the feature of category c of the i-th news item is... Contribution score for:

[0035]

[0036] In equation (9), V i,c and b i,c Learnable parameters q that contribute to the category i,c Attention query vectors contributed to the category;

[0037] Step 1.12: Assume each news item in group g belongs to a category. Considering all the categories corresponding to the news items in group g, normalize the attention weight of category c of the i-th news item. Recorded as:

[0038]

[0039] In equation (10), N+1 is the number of clicked news items and candidate news items within group g, and p represents the p-th news item. Features of category c of the p-th news item within group g Contribution score;

[0040] Step 1.13: Based on normalized attention weights, fuse the category feature representations of all news items within group g, and the in-group category feature representation of group g. for:

[0041]

[0042] In equation (11), N+1 is the number of clicked news items and candidate news items within the group;

[0043] Step 1.14: Encode the paragraph number features [CLS] of the text content of group g using a BERT pre-trained language model, and use this encoding to represent the paragraph number features within group g.

[0044] Step 1.15, Represent the within-group categorical features and the characteristic of paragraph numbering within the group The features are concatenated and fused using a multilayer perceptron to obtain the within-group feature representation s. g for:

[0045]

[0046] In equation (12), || represents the splicing operation, W s b are the learnable parameters of the fully connected layer. s This is the bias for the fully connected layer.

[0047] Step 2: Design an inter-group feature encoder to capture the semantic relevance of news between different groups, and model the click news feature representation and candidate news feature representation of cross semantics.

[0048] Furthermore, the specific steps of step 2 are as follows:

[0049] Step 2.1: Input the intra-group feature representations of all clicked news items and candidate news items within the G groups into a BiLSTM for cross-interaction. For the i-th intra-group clicked news item feature representation of group g... Define click news feature representation with cross semantics for:

[0050]

[0051] In equation (13), r g(candi) It represents the in-group candidate news features of the candidate news candi;

[0052] Step 2.2, repeat step 2.1, to obtain the candidate news features represented by cross-semantics as r g,candi .

[0053] Step 3: The click news feature representation, candidate news feature representation, and intra-group feature representation of cross-semantics are fused through an attention mechanism to model the click news feature representation, candidate news feature representation, and group feature representation of inter-group perception.

[0054] Furthermore, the specific steps of step 3 are as follows:

[0055] Step 3.1, considering that all the news articles clicked by a user contribute differently to the user's interest, the contribution score of the i-th news article clicked by user u in the g-th group is calculated. for:

[0056]

[0057] In equation (14), V u and v u q is the weight parameter. u For attention query vectors;

[0058] Step 3.2, the normalized attention weight of the i-th clicked news item in user u's g-th group among all clicked news items is:

[0059]

[0060] In equation (15), G is the number of groups, N is the number of clicked news items within a group, x represents the counting loop variable for the group, and y represents the counting loop variable for the news items within the group. The contribution score for the y-th clicked news item in the x-th group of user u;

[0061] Step 3.3: Based on the attention mechanism, the click news feature representations of the cross-semantic relationships of G different groups are fused, and the inter-group perceived click news feature representation is defined as:

[0062]

[0063] In equation (16), This represents the normalized attention weight of the y-th clicked news item in user u's x-th group among all the news items clicked by the user. The clicked news feature representation of the y-th clicked news item in group x is obtained by performing cross-group semantics.

[0064] Step 3.4: Repeat steps 3.1-3.3 to fuse the candidate news feature representations of cross-semantics from the G different groups using an attention mechanism to obtain the candidate news feature representations with inter-group perception. The intra-group feature representations of G different groups are fused using an attention mechanism to obtain inter-group perceived group feature representations.

[0065] Step 4: Use a gating mechanism to aggregate the important features of the click-based news feature representation and the group feature representation between groups to model the user interest representation. Then, use a gating mechanism to aggregate the candidate news feature representation and the important features of the group feature representation between groups to model the candidate news representation.

[0066] Furthermore, the specific steps of step 4 are as follows:

[0067] Step 4.1: Select important inter-group perception group feature representations and important inter-group perception click news feature representations through a gating mechanism, denoted as:

[0068]

[0069] In equation (17), m is the gate vector for the inter-group perceived group feature representation and the inter-group perceived click news feature representation; W m It is a learnable parameter matrix; b m Represents a parameter vector;

[0070] Step 4.2, Aggregate the group feature representations of inter-group perception. Inter-group perceived click news representation Modeling user interest representation r u for:

[0071]

[0072] Step 4.3: Repeat steps 4.1-4.2 to aggregate the group feature representations of inter-group perception through a gating mechanism. Candidate news feature representation based on inter-group perception The candidate news indicated

[0073] Step 5: Based on the user interest representation and the candidate news representation, calculate the user's interest score for the candidate news, optimize the interest score based on the negative sampling loss function, and update the model parameters;

[0074] Furthermore, the specific steps of step 5 are as follows:

[0075] Step 5.1, user u's prediction score for candidate news candi is represented by user interest r. u And candidate news indicates The interest score is obtained through the inner product and is represented as follows:

[0076]

[0077] Step 5.2: Select news articles clicked by users to form a positive sample; randomly select K news articles from the same session that were not clicked to form a negative sample; calculate the interest scores of the positive and negative samples. and The NCE loss function is defined as follows:

[0078]

[0079] In equation (20), P is the set of positive samples. It is the predicted score of the j-th negative sample corresponding to the i-th positive sample.

[0080] Step 6: Based on the optimized user interest representation and candidate news representation, calculate the user's interest score for the unclicked candidate news, sort them according to the score, and generate a recommendation list;

[0081] Furthermore, the specific steps of step 6 are as follows:

[0082] For candidate news items that have not been clicked, calculate the user's interest score for the candidate news items, and recommend the top k news items with the highest interest scores to the user.

[0083] Compared with the prior art, the present invention has the following advantages:

[0084] The significant difference between the method provided in this invention and existing methods lies in the design of a personalized news recommendation method based on time-series grouped user interest representations. This invention establishes an interest prediction model based on user and news semantic embeddings and proposes a personalized news recommendation strategy based on time-series grouped user interest representations. This invention obtains intra-group feature encoding representations by fusing intra-group historical click news representations, candidate news representations, intra-group category feature representations, and group label feature representations of different subgroups. This reflects the user's stage-specific interest preferences. Furthermore, it learns the correlation of user stage-specific interest changes from inter-group encodings of different subgroups, and then calculates the similarity between users and candidate news, improving the accuracy of news recommendations. This invention can provide a major application in personalized news recommendation systems. Attached Figure Description

[0085] Figure 1 This is a schematic diagram of the overall model architecture of the present invention.

[0086] Figure 2 This is a schematic diagram of the architecture of step 1 of the present invention. Detailed Implementation

[0087] To gain a deeper understanding of this invention, we will provide a comprehensive and detailed description. However, this invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a full understanding of the disclosure of this invention.

[0088] The personalized news recommendation method based on time-series grouped user interest representation described in this invention is implemented through a computer program. The specific implementation methods of the technical solution proposed in this invention will be detailed below according to the flowchart.

[0089] like Figure 1 As shown, the personalized news recommendation method based on time-series grouped user interest representation of the present invention includes the following steps:

[0090] Step 1: Based on the sliding time window, the news clicked by the user in history is grouped in time sequence, and the candidate news is fused in different groups. Within the same group, an in-group feature encoder is designed to model the content of news title, news category, and news subcategory, and to learn the feature representation of clicked news within the group, the feature representation of candidate news within the group, and the feature representation of group within the group.

[0091] Furthermore, such as Figure 2 As shown, the specific steps of step 1 are as follows:

[0092] Step 1.1: Divide the user's historical clicked news into G groups {S1,…,S} according to time sequence. g ,…,S G}, S G Let G be the set of historical clicked news items. Candidate news items (candi) are added sequentially to each group. The set of groups includes the historical clicked news items and candidate news items for each group, denoted as S1∪candi,…,S. g ∪candi,…,S G ∪candi;

[0093] Step 1.2: Within the g-th group, use the BERT pre-trained language model as a text encoder to process the candidate news and historical clicked news title word sequences within the group. As input, the feature representation of the words in the output titles is as follows:

[0094]

[0095] In equation (1), candi represents candidate news items within group g, hc represents all historically clicked news items within group g, N represents the number of historically clicked news items in each group, l represents the number of pre-trained words in the news title, and N l It is the l-th pre-training word of the Nth news item; i represents the i-th news item, i l This represents the l-th pre-trained word in the i-th news item;

[0096] Step 1.3: Using a convolutional neural network, analyze the j-th word w of the i-th news item in the historical clicked news hc within group g. j Feature representation Further encoding is performed, and the word feature representation of cross-semantics is denoted as:

[0097]

[0098] In equation (2), W' is the learnable parameter of the CNN filter, b' is the bias, and * is the convolution operator. It is the feature representation of the i-th news item from the word at position (jf) to (j+f);

[0099] Step 1.4: Based on the different importance levels of each word in the news headline, calculate the j-th word w of the i-th news item in the historically clicked news hc within group g. j Contribution score

[0100]

[0101] In equation (3), V t and b t Here, q represents the weight parameter, q represents the attention query vector, and T represents the transpose operation.

[0102] Step 1.5, calculate the w of the j-th word. j Normalized attention weights of all words in a news headline for:

[0103]

[0104] In equation (4), For the j'th word of the i-th news item in group g j' The contribution score; j' represents the j'-th word; exp is the exponential function;

[0105] Step 1.6: Based on the normalized attention weights, fuse the cross-semantic word feature representations of all words in the title of the i-th news item in the historical clicked news hc within group g, and define the title feature representation. for:

[0106]

[0107] Step 1.7: Perform one-hot encoding on the news category c of the i-th news item in the historical clicked news hc within group g, and initialize its representation using a lookup table. The category feature representation of the i-th news item is obtained through a multilayer perceptron. for:

[0108]

[0109] In equation (6), V c and b c These are parameters in a multilayer perceptron with categorical features;

[0110] Step 1.8: Perform one-hot encoding on the news subcategory sc of the i-th news item in the historical clicked news hc within group g, and represent it using a lookup table. The sub-category feature representation of the i-th news item is obtained through a multilayer perceptron. for:

[0111]

[0112] In equation (7), V s and b s These are parameters in a multilayer perceptron for sub-category features;

[0113] Step 1.9: Concatenate the title feature representation, category feature representation, and subcategory feature representation of the i-th historical clicked news item hc within group g, and then perform cross-fusion using a multilayer perceptron to obtain the feature representation of the i-th clicked news item within the group. for:

[0114]

[0115] In equation (8), || represents the splicing operation, W n b is the learnable parameter of the fully connected layer for news features. n For biasing the fully connected layer;

[0116] Step 1.10, repeat steps 1.3-1.9, model the candidate news features within group g, represented as r. g(candi) ;

[0117] Step 1.11: News category is a feature of grouping. Based on the different importance of news categories in group g, the feature of category c of the i-th news item is... Contribution score for:

[0118]

[0119] In equation (9), V i,c and b i,c Learnable parameters q that contribute to the category i,c Attention query vectors contributed to the category;

[0120] Step 1.12: Assume each news item in group g belongs to a category. Considering all the categories corresponding to the news items in group g, normalize the attention weight of category c of the i-th news item. Recorded as:

[0121]

[0122] In equation (10), N+1 is the number of clicked news items and candidate news items within group g, and p represents the p-th news item. Features of category c of the p-th news item within group g Contribution score;

[0123] Step 1.13: Based on normalized attention weights, fuse the category feature representations of all news items within group g, and the in-group category feature representation of group g. for:

[0124]

[0125] In equation (11), N+1 is the number of clicked news items and candidate news items within the group;

[0126] Step 1.14: Encode the paragraph number features [CLS] of the text content of group g using a BERT pre-trained language model, and use this encoding to represent the paragraph number features within group g.

[0127] Step 1.15, Represent the within-group categorical features and the characteristic of paragraph numbering within the group The features are concatenated and fused using a multilayer perceptron to obtain the within-group feature representation s. g for:

[0128]

[0129] In equation (12), || represents the splicing operation, W s b are the learnable parameters of the fully connected layer. s This is the bias for the fully connected layer.

[0130] Step 2: Design an inter-group feature encoder to capture the semantic relevance of news between different groups, and model the click news feature representation and candidate news feature representation of cross semantics.

[0131] Furthermore, the specific steps of step 2 are as follows:

[0132] Step 2.1: Input the intra-group feature representations of all clicked news items and candidate news items within the G groups into a BiLSTM for cross-interaction. For the i-th intra-group clicked news item feature representation of group g... Define click news feature representation with cross semantics for:

[0133]

[0134] In equation (13), r g(candi) It represents the in-group candidate news features of the candidate news candi;

[0135] Step 2.2, repeat step 2.1, to obtain the candidate news features represented by cross-semantics as r g,candi .

[0136] Step 3: The click news feature representation, candidate news feature representation, and intra-group feature representation of cross-semantics are fused through an attention mechanism to model the click news feature representation, candidate news feature representation, and group feature representation of inter-group perception.

[0137] Furthermore, the specific steps of step 3 are as follows:

[0138] Step 3.1, considering that all the news articles clicked by a user contribute differently to the user's interest, the contribution score of the i-th news article clicked by user u in the g-th group is calculated. for:

[0139]

[0140] In equation (14), V u and v u q is the weight parameter. u For attention query vectors;

[0141] Step 3.2, the normalized attention weight of the i-th clicked news item in user u's g-th group among all clicked news items is:

[0142]

[0143] In equation (15), G is the number of groups, N is the number of clicked news items within a group, x represents the counting loop variable for the group, and y represents the counting loop variable for the news items within the group. The contribution score for the y-th clicked news item in the x-th group of user u;

[0144] Step 3.3: Based on the attention mechanism, the click news feature representations of the cross-semantic relationships of G different groups are fused, and the inter-group perceived click news feature representation is defined as:

[0145]

[0146] In equation (16), This represents the normalized attention weight of the y-th clicked news item in user u's x-th group among all the news items clicked by the user. The clicked news feature representation of the y-th clicked news item in group x is obtained by performing cross-group semantics.

[0147] Step 3.4: Repeat steps 3.1-3.3 to fuse the candidate news feature representations of cross-semantics from the G different groups using an attention mechanism to obtain the candidate news feature representations with inter-group perception. The intra-group feature representations of G different groups are fused using an attention mechanism to obtain inter-group perceived group feature representations.

[0148] Step 4: Use a gating mechanism to aggregate the important features of the click-based news feature representation and the group feature representation between groups to model the user interest representation. Then, use a gating mechanism to aggregate the candidate news feature representation and the important features of the group feature representation between groups to model the candidate news representation.

[0149] Furthermore, the specific steps of step 4 are as follows:

[0150] Step 4.1: Select important inter-group perception group feature representations and important inter-group perception click news feature representations through a gating mechanism, denoted as:

[0151]

[0152] In equation (17), m is the gate vector for the inter-group perceived group feature representation and the inter-group perceived click news feature representation; W m It is a learnable parameter matrix; b m Represents a parameter vector;

[0153] Step 4.2, Aggregate the group feature representations of inter-group perception. Inter-group perceived click news representation Modeling user interest representation r u for:

[0154]

[0155] Step 4.3: Repeat steps 4.1-4.2 to aggregate the group feature representations of inter-group perception through a gating mechanism. Candidate news feature representation based on inter-group perception The candidate news indicated

[0156] Step 5: Based on the user interest representation and the candidate news representation, calculate the user's interest score for the candidate news, optimize the interest score based on the negative sampling loss function, and update the model parameters;

[0157] Furthermore, the specific steps of step 5 are as follows:

[0158] Step 5.1, user u's prediction score for candidate news candi is represented by user interest r. u And candidate news indicates The interest score is obtained through the inner product and is represented as follows:

[0159]

[0160] Step 5.2: Select news articles clicked by users to form a positive sample; randomly select K news articles from the same session that were not clicked to form a negative sample; calculate the interest scores of the positive and negative samples. and The NCE loss function is defined as follows:

[0161]

[0162] In equation (20), P is the set of positive samples. It is the predicted score of the j-th negative sample corresponding to the i-th positive sample.

[0163] Step 6: Based on the optimized user interest representation and candidate news representation, calculate the user's interest score for the unclicked candidate news, sort them according to the score, and generate a recommendation list;

[0164] Furthermore, the specific steps of step 6 are as follows:

[0165] For candidate news items that have not been clicked, calculate the user's interest score for the candidate news items, and recommend the top k news items with the highest interest scores to the user.

[0166] To verify the effectiveness of the method, experiments were conducted on the MIND dataset (MIND(msnews.github.io)). This dataset contains user-related information and news content information. User information includes the user's browsing and clicking data, click time, and user log data. User log data includes the user's candidate news items and their interest tags for those items, where 1 indicates the user is interested and 0 indicates the user is not interested. News information includes the news ID, category, subcategory, title, summary, entities contained in the title, and entities contained in the summary. The dataset information is shown in Table 1.

[0167] Table 1 Dataset Information

[0168]

[0169] We used three news features: news title, category, and subcategory. The evaluation metrics were AUC, MRR, and nDCG@5. To verify the effectiveness and advancement of the proposed technical solution, several existing recommendation prediction models were compared: DKN, NPA, NRMS, FIM, and UNbert. The experimental results are shown in Table 2.

[0170] Table 2 Experimental Results

[0171]

[0172] As can be seen from the results in Table 2, the technical solution of the present invention can achieve better accuracy and reliability than existing methods when predicting users' interest in news.

[0173] Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications are readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of this invention are protected.

Claims

1. A personalized news recommendation method based on time-series grouped user interest representation, characterized in that, Includes the following steps: Step 1: Based on the sliding time window, the news clicked by the user in history is grouped in time sequence, and the candidate news is fused in different groups. Within the same group, an in-group feature encoder is designed to model the content of news title, news category, and news subcategory, and to learn the feature representation of clicked news within the group, the feature representation of candidate news within the group, and the feature representation of group within the group. Step 1 includes: Step 1.1: Divide the user's historical clicked news into time series. Groups , Grouping The historical news collection will include candidate news items. Add them sequentially to each group. The group set includes the historical clicked news and candidate news for each group, represented as follows: ; Step 2: Design an inter-group feature encoder to capture the semantic relevance of news between different groups, and model the click news feature representation and candidate news feature representation of cross semantics. Step 2 includes: Step 2.1, The intra-group feature representations of all clicked news items and candidate news items within each group are input into a BiLSTM and cross-validated. The Clicking on news features within a group indicates Define the click news feature representation with cross-semantic meaning. for: （13） In equation (13), It is a candidate news item. The characteristics of candidate news within the group are represented; Step 2.2, repeat step 2.1, to obtain the candidate news feature representation of cross-semantics as follows: ; Step 3: The click news feature representation, candidate news feature representation, and intra-group feature representation of cross-semantics are fused through an attention mechanism to model the click news feature representation, candidate news feature representation, and group feature representation of inter-group perception. Step 4: Use a gating mechanism to aggregate the important features of the click-based news feature representation and the group feature representation between groups to model the user interest representation. Then, use a gating mechanism to aggregate the candidate news feature representation and the important features of the group feature representation between groups to model the candidate news representation. Step 5: Based on the user interest representation and the candidate news representation, calculate the user's interest score for the candidate news, optimize the interest score based on the negative sampling loss function, and update the model parameters; Step 6: Based on the optimized user interest representation and candidate news representation, calculate the user's interest score for the unclicked candidate news, sort them according to the score, and generate a recommendation list.

2. The personalized news recommendation method based on time-series grouped user interest representation according to claim 1, characterized in that, In step 1, the user's historical clicked news is time-series grouped based on a sliding time window, and candidate news is fused between different groups. Within the same group, an intra-group feature encoder is designed to model the content of news titles, news categories, and news subcategories, learning intra-group clicked news feature representations, intra-group candidate news feature representations, and intra-group group feature representations. The specific steps are as follows: Step 1.2, in the Within each group, a BERT pre-trained language model is used as a text encoder to process the word sequences of candidate news articles and historical clicked news titles within the group. As input, the feature representation of the words in the output titles is as follows: （1） In formula (1) It is a group Candidate news within, It is a group All historical news clicks within, This is the number of historically clicked news items within each group. The number of pre-trained words for news headlines. It is the first The first news item One pre-trained word; Indicates the first This is a news item. Indicates the first The first news item One pre-trained word; Step 1.3, grouping using a convolutional neural network Internal History Click News The The first news item one word Feature representation Further encoding is performed, and the word feature representation of cross-semantics is denoted as: （2） In equation (2), These are the learnable parameters of the CNN filter. For bias, For convolution operators, It is the position of the i-th news item. arrive Word feature representation; Step 1.4: Calculate groupings based on the different importance levels of each word in the news headline. Internal History Click News The The first news item one word Contribution score : （3） In equation (3), and For weight parameters, For attention query vectors; This is a transpose operation; Step 1.5, calculate the first... one word Normalized attention weights of all words in a news headline for: （4） In equation (4), Grouping Inner The first news item one word Contribution score; Indicates the first One word; It is an exponential function; Step 1.6: Group according to normalized attention weights Internal History Click News The title feature representation is defined by fusing the cross-semantic word feature representations of all words in the title of the i-th news item. for: （5） Step 1.7, grouping Internal History Click News The news category of the i-th news item Perform one-hot encoding and initialize its representation using a lookup table. The category feature representation of the i-th news item is obtained through a multilayer perceptron. for: （6） In equation (6), and These are parameters in a multilayer perceptron with categorical features; Step 1.8, grouping Internal History Click News The news subcategory of the i-th news item Perform one-hot encoding and represent it using a lookup table. The sub-category feature representation of the i-th news item is obtained through a multilayer perceptron. for: （7） In equation (7), and These are parameters in a multilayer perceptron for sub-category features; Step 1.9, group Internal History Click News The title feature representation, category feature representation, and subcategory feature representation of the i-th news item are concatenated and then cross-fused using a multilayer perceptron to obtain the feature representation of the i-th clicked news item within the group. for: （8） In equation (8), For splicing operations, The learnable parameters of the fully connected layer for news features. For biasing the fully connected layer; Step 1.10, repeat steps 1.3-1.9, model and group. Within-group candidate news features are represented as ; Step 1.11, news category is a feature of grouping, based on grouping The different levels of importance of news categories in China, the first News categories Features Contribution score for: （9） In equation (9), and Learnable parameters that contribute to the category. Attention query vectors contributed to the category; Step 1.12, set up groups Each news item belongs to a category, taking into account grouping. All news items within the category will be ranked according to their respective categories. News categories Normalized attention weights Recorded as: （10） In equation (10), It is a group The number of clicked news items and candidate news items. Indicates the first This is a news item. Grouping Inner News categories Features Contribution score; Step 1.13: Grouping based on normalized attention weights The category characteristics of all news items are represented by the fusion, the first Within-group categorical features of a group for: （11） In equation (11), This represents the number of clicked news items and candidate news items within the group; Step 1.14, group The paragraph numbering features [CLS] of the text content, after being encoded using a BERT pre-trained language model, are used as grouping data. The characteristic of paragraph numbering within a group ; Step 1.15, Represent the within-group categorical features and the characteristic of paragraph numbering within a group The features are concatenated and fused using a multilayer perceptron to obtain within-group feature representations. for: （12） In equation (12), For splicing operations, These are the learnable parameters for the fully connected layer. This is the bias for the fully connected layer.

3. The personalized news recommendation method based on time-series grouped user interest representation according to claim 2, characterized in that, In step 3, the click news feature representation, candidate news feature representation, and within-group feature representation of cross-semantic semantics are fused using an attention mechanism to model the click news feature representation, candidate news feature representation, and group feature representation of inter-group perception. The specific steps are as follows: Step 3.1, considering that all the news articles clicked by a user contribute differently to the user's interest, the user... The Group 1 Contribution points per click of news for: （14） In equation (14), and For weight parameters, For attention query vectors; Step 3.2, User The Group 1 The normalized attention weight of each clicked news item among all clicked news items by the user is: （15） In equation (15), The number of groups. y represents the number of clicked news items within the group, where x is the loop counter for the group and y is the loop counter for the number of news items within the group. For users The Group 1 The contribution score for each click on a news article; Step 3.3, based on the attention mechanism, The click news feature representations of different groups with cross-semantic meaning are fused, and the inter-group perceived click news feature representation is defined as follows: （16） In equation (16), Indicates user The Group 1 The normalized attention weight of each clicked news item among all the news items clicked by the user. Grouping The Click news feature representation based on the cross-semantic representation of click news items after inter-group cross-referencing; Step 3.4, repeat steps 3.1-3.3, and then... The cross-semantic candidate news feature representations of different groups are fused through an attention mechanism to obtain inter-group perceptual candidate news feature representations. ;right The intra-group feature representations of different groups are fused through an attention mechanism to obtain the inter-group perceived group feature representations. .

4. The personalized news recommendation method based on time-series grouped user interest representation according to claim 3, characterized in that, In step 4, a gating mechanism is used to aggregate the key features of the click-based news feature representation and the group feature representation perceived by the inter-group perception, to model user interest representation. Furthermore, a gating mechanism is used to aggregate the candidate news feature representation and the key features of the group feature representation perceived by the inter-group perception, to model candidate news representation. The specific steps are as follows: Step 4.1: Select important inter-group perception group feature representations and important inter-group perception click news feature representations through a gating mechanism, denoted as: （17） In equation (17), The gate vectors for group feature representation and click news feature representation of inter-group perception; It is a learnable parameter matrix; Represents a parameter vector; Step 4.2, Aggregate the group feature representations of inter-group perception. Click news representation with inter-group perception Modeling user interest representation for: （18） Step 4.3: Repeat steps 4.1-4.2 to aggregate the group feature representations of inter-group perception through a gating mechanism. Candidate news feature representation based on inter-group perception News of the candidates indicates .

5. The personalized news recommendation method based on time-series grouped user interest representation according to claim 4, characterized in that, In step 5, based on the user interest representation and the candidate news representation, the user's interest score for the candidate news set is calculated. The interest score is then optimized based on the negative sampling loss function, and the model parameters are updated. The specific steps are as follows: Step 5.1, User News on the candidates The predicted score is represented by user interests. And candidate news indicates The interest score is obtained through the inner product and is represented as follows: （19） Step 5.2: Select news articles clicked by users to form a positive sample, and randomly select news articles that were not clicked in the same session. The news items form a negative sample, and the interest scores for the positive and negative samples are calculated. and The NCE loss function is defined as follows: （20） In equation (20), It is a set of positive samples. It is the first The corresponding positive sample of the th The predicted score of each negative sample.

6. The personalized news recommendation method based on time-series grouped user interest representation according to claim 5, characterized in that, In step 6, based on the optimized user interest representation and candidate news representation, the user's interest score for the unclicked candidate news is calculated, and the news is sorted according to the score to generate a recommendation list. The specific steps are as follows: For candidate news items that were not clicked, calculate the user's interest score for the candidate news items and prioritize those with higher interest scores. The news is recommended to the user.

Images