Personalized news recommendation method for time sequence grouping user interest representation

Abstract

The invention discloses a personalized news recommendation method for time sequence grouping user interest representation, and belongs to the technical field of news recommendation. Aiming at the problem that an existing news recommendation system neglects the effect of user stage interest change in user interest modeling, firstly, an intra-group feature encoder is designed by utilizing a title, a category and a subcategory of news to describe news features in groups; secondly, designing an inter-group feature encoder, and learning candidate news representation of inter-group perception, click news representation of inter-group perception and group feature representation of inter-group perception; then, obtaining a user interest representation and a candidate news representation based on a gated network; and finally, performing inner product on the user interest representation and the candidate news representation to obtain an interest score. According to the method, the news title, category and subcategory multi-feature learning user news clicking interest is fused, the similarity between the user interest and the candidate news is calculated, the news recommendation accuracy is improved, and technical support is achieved in the aspect of personalized news recommendation system application.

Description

Technical Field

[0001] The present invention belongs to the technical field of news recommendation, and in particular relates to a personalized news recommendation method based on time-sequential grouping user interest representation. Background Art

[0002] In recent years, with the popularity of online news platforms and the convenience of online services, more and more users have turned from traditional paper newspaper reading habits to digital news reading. However, faced with a large amount of news information every day, users often find it difficult to quickly find content that interests them. In order to solve the problem of information overload, news recommendation systems have emerged. News recommendation systems provide users with personalized services and enhance their reading experience by accurately modeling their interests.

[0003] Compared with traditional product recommendation systems, news recommendation systems face some unique challenges. First, news articles are highly time-sensitive. Traditional recommendation methods (such as collaborative filtering that relies on user or item IDs) encounter serious cold start problems in news recommendation. Semantic modeling based on news text content can effectively capture the timeliness of news and provide accurate recommendations. Secondly, a user's news browsing history in a short period of time reflects the user's dynamic interest changes. For example, a user may only pay attention to entertainment news for a period of time, then during this period of time, entertainment news is the user's main interest. When a user has a recent need to buy a car and reads related news, car news becomes one of the user's main interests. How to capture the user's periodic interest changes is very necessary for accurate personalized news recommendations.

[0004] The text content of news, such as title, summary, category, etc., contains rich semantic information. The present invention performs semantic modeling on the text content of historical clicked news and candidate news of users grouped in time series, mines the potential interest preferences of users perceived by candidate news, and improves the quality of news recommendation. Summary of the invention

[0005] Aiming at the problem that the existing news recommendation system ignores the role of user's periodic interest changes in user interest modeling, the present invention provides a personalized news recommendation method based on time-series grouping user interest representation.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] Step 1: Sequentially group the user's historical clicked news based on a sliding time window, merge candidate news in different groups, design a feature encoder within the same group, model the news title, news category, and news subcategory content, and 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:

[0009] Step 1.1: divide the user's historical click news into G groups {S1,…,S g ,…,S G},S G is the historical click news set of group G, and the candidate news candi is added to each group in turn. The group set includes the historical click news and candidate news of each group, which is expressed as S1∪candi,…,S g ∪candi,…,S G ∪candi;

[0010] Step 1.2: In the g-th group, use the Bert pre-trained language model as the text encoder to convert the candidate news and historical click news title word sequences in the group As input, the feature representation of the words between output titles is:

[0011]

[0012] In formula (1), candi is the candidate news in group g, hc is all the historical clicked news in group g, N is the number of historical clicked news in each group, l is the number of pre-trained words in the news title, and N l is the lth pre-trained word of the Nth news; i represents the i-th news, i l Represents the lth pre-trained word of the i-th news;

[0013] Step 1.3: Use the convolutional neural network to train the jth word w of the i-th news in the historical click news hc in group g j The feature representation Further encoding is performed, and the word features of cross-semantics are represented as:

[0014]

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

[0016] Step 1.4, according to the different importance of each word in the news title, calculate the jth word w of the i-th news of the historical clicked news hc in group g j Contribution score

[0017]

[0018] In formula (3), Vt and b t is the weight parameter, q is the attention query vector; T is the transposition operation;

[0019] Step 1.5, calculate the jth word w j Normalized attention weights for all words in a news title for:

[0020]

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

[0022] Step 1.6, according to the normalized attention weight, the cross-semantic word feature representations of all words in the news title of the i-th news in the historical click news hc in group g are fused to define the title feature representation for:

[0023]

[0024] Step 1.7, one-hot encode the news category c of the i-th news of the historical click news hc in group g, and use the query table to initialize it as And the category feature representation of the i-th news is obtained through the multi-layer perceptron for:

[0025]

[0026] In formula (6), V c and b c is the parameter in the category feature multilayer perceptron;

[0027] Step 1.8, one-hot encode the news subcategory sc of the i-th news of the historical clicked news hc in group g, and use the query table to represent it as And obtain the subcategory feature representation of the i-th news through the multi-layer perceptron for:

[0028]

[0029] In formula (7), V s and b s is the parameter in the subcategory feature multilayer perceptron;

[0030] Step 1.9, concatenate the title feature representation, category feature representation, and subcategory feature representation of the i-th news in the historical click news hc in group g, and obtain the feature representation of the i-th click news in the group through multi-layer perceptron cross fusion for:

[0031]

[0032] In formula (8), || is the concatenation operation, W n is the learnable parameter of the fully connected layer of news features, b n is the bias of the fully connected layer;

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

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

[0035]

[0036] In formula (9), V i,c and b i,c The learnable parameters contributed by the category, q i,c The attention query vector contributed to the category;

[0037] Step 1.12: Assume that each news in group g belongs to a category. Considering the categories corresponding to all news in group g, the normalized attention weight of category c of the i-th news is Denoted as:

[0038]

[0039] In formula (10), N+1 is the number of clicked news and candidate news in group g, p represents the pth news, is the feature of category c of the pth news in group g Contribution score;

[0040] Step 1.13, based on the normalized attention weight, the category feature representations of all news in group g are fused, and the category feature representation within group g is for:

[0041]

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

[0043] Step 1.14: The paragraph number feature [CLS] of the text content of group g is encoded using the Bert pre-trained language model and used as the intra-group paragraph number feature representation of group g.

[0044] Step 1.15: Represent the intra-group category features And the paragraph number feature within the group Splicing, feature fusion is performed through a multi-layer perceptron to obtain the intra-group feature representation s g for:

[0045]

[0046] In formula (12), || is the concatenation operation, W s is the learnable parameter of the fully connected layer, b s is the bias of the fully connected layer.

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

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

[0049] Step 2.1: Input the intra-group feature representations of all click news and candidate news in G groups into BiLSTM for crossover. For the i-th intra-group click news feature representation of group g, Defining cross-semantic feature representation of click news for:

[0050]

[0051] In formula (13), r g(candi) is the feature representation of candidate news within the group of candidate news candi;

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

[0053] Step 3: The cross-semantic click news feature representation, the cross-semantic candidate news feature representation, and the intra-group feature representation are fused through the attention mechanism to model the inter-group-aware click news feature representation, the inter-group-aware candidate news feature representation, and the inter-group-aware group feature representation;

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

[0055] Step 3.1: Considering that all clicked news of a user have different contributions to the user's interest expression, the contribution score of the i-th clicked news of the g-th group of user u is for:

[0056]

[0057] In formula (14), V u and v u is the weight parameter, q u is the attention query vector;

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

[0059]

[0060] In formula (15), G is the number of groups, N is the number of clicked news in a group, x is the counting loop variable of the group, and y is the counting loop variable of the news in the group. is the contribution score of the yth clicked news in the xth group of user u;

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

[0062]

[0063] In formula (16), represents the normalized attention weight of the yth clicked news of the xth group of user u among all the clicked news of the user, The click news feature representation of the yth click news in group x after performing cross-group cross-semantics;

[0064] Step 3.4: Repeat steps 3.1-3.3 to fuse the cross-semantic candidate news feature representations of G different groups through the attention mechanism to obtain the inter-group perceived candidate news feature representation. The intra-group feature representations of G different groups are fused through the attention mechanism to obtain the inter-group perceived group feature representation

[0065] Step 4: A gating mechanism is used to aggregate the important features of the inter-group perceived click news feature representation and the inter-group perceived group feature representation to model the user interest representation, and a gating mechanism is used to aggregate the important features of the inter-group perceived candidate news feature representation and the inter-group perceived group feature representation to model the candidate news representation;

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

[0067] Step 4.1, select the important inter-group-aware group feature representation and the important inter-group-aware click news feature representation through the gating mechanism, which are denoted as:

[0068]

[0069] In formula (17), m is the gate vector of the inter-group-aware group feature representation and the inter-group-aware click news feature representation; W m is the learnable parameter matrix; b m represents the parameter vector;

[0070] Step 4.2: Aggregate inter-group-aware group feature representation and inter-group perceived click news representation Modeling user interest representation u for:

[0071]

[0072] Step 4.3, repeat steps 4.1-4.2, and aggregate the group feature representation of inter-group awareness through the gating mechanism and inter-group-aware candidate news feature representation Get candidate news representation

[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:

[0075] Step 5.1: The predicted score of user u for candidate news candi is represented by user interest r u and candidate news The interest score is obtained through the inner product, expressed as:

[0076]

[0077] Step 5.2: Select the news clicked by the user as the positive sample, and randomly select K news that have not been clicked in the same session as the negative sample. Calculate the interest scores of the positive and negative samples as and And define the NCE loss function as:

[0078]

[0079] In formula (20), P is the set of positive samples, is the prediction score of the jth negative sample corresponding to the i-th positive sample.

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

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

[0082] For unclicked candidate news, the user's interest score for the candidate news is calculated, and the top k news with higher interest scores are recommended to the user.

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

[0084] The method provided in the present invention is different from the existing methods in that it designs a personalized news recommendation method for time-series grouping user interest representation, establishes an interest prediction model based on user and news semantic embedding, and proposes a personalized news recommendation strategy for time-series grouping user interest representation. The present invention obtains a group feature coding representation by fusing the historical click news representation, candidate news representation, and category feature representation within the group with the group tag feature representation of the group, thereby reflecting the user's staged interest preference, and learns the association of the user's staged interest changes from the inter-group coding of different groups, and then calculates the similarity between the user and the candidate news, thereby improving the accuracy of news recommendation, and can provide major applications in personalized news recommendation systems. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0086] Figure 2 It is a schematic diagram of the architecture of step 1 of the present invention. DETAILED DESCRIPTION

[0087] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.

[0088] The personalized news recommendation method of the present invention based on the time-sequential grouping user interest representation is implemented by a computer program. The specific implementation method of the technical solution proposed by the present invention will be described in detail below according to the process.

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

[0090] Step 1: Sequentially group the user's historical clicked news based on a sliding time window, merge candidate news in different groups, design a feature encoder within the same group, model the news title, news category, and news subcategory content, and 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, if Figure 2 As shown, the specific steps of step 1 are:

[0092] Step 1.1: divide the user's historical click news into G groups {S1,…,S g ,…,S G},S G is the historical click news set of group G, and the candidate news candi is added to each group in turn. The group set includes the historical click news and candidate news of each group, which is expressed as S1∪candi,…,S g ∪candi,…,S G ∪candi;

[0093] Step 1.2: In the g-th group, use the Bert pre-trained language model as the text encoder to convert the candidate news and historical click news title word sequences in the group As input, the feature representation of the words between output titles is:

[0094]

[0095] In formula (1), candi is the candidate news in group g, hc is all the historical clicked news in group g, N is the number of historical clicked news in each group, l is the number of pre-trained words in the news title, and N l is the lth pre-trained word of the Nth news; i represents the i-th news, i l Represents the lth pre-trained word of the i-th news;

[0096] Step 1.3: Use the convolutional neural network to train the jth word w of the i-th news in the historical click news hc in group g j The feature representation Further encoding is performed, and the word features of cross-semantics are represented as:

[0097]

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

[0099] Step 1.4, according to the different importance of each word in the news title, calculate the jth word w of the i-th news of the historical clicked news hc in group g j Contribution score

[0100]

[0101] In formula (3), V t and b t is the weight parameter, q is the attention query vector; T is the transposition operation;

[0102] Step 1.5, calculate the jth word w j Normalized attention weights for all words in a news title for:

[0103]

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

[0105] Step 1.6, according to the normalized attention weight, the cross-semantic word feature representations of all words in the news title of the i-th news in the historical click news hc in group g are fused to define the title feature representation for:

[0106]

[0107] Step 1.7, one-hot encode the news category c of the i-th news of the historical click news hc in group g, and use the query table to initialize it as And the category feature representation of the i-th news is obtained through the multi-layer perceptron for:

[0108]

[0109] In formula (6), V c and b c is the parameter in the category feature multilayer perceptron;

[0110] Step 1.8, one-hot encode the news subcategory sc of the i-th news of the historical clicked news hc in group g, and use the query table to represent it as And obtain the subcategory feature representation of the i-th news through the multi-layer perceptron for:

[0111]

[0112] In formula (7), V s and b s is the parameter in the subcategory feature multilayer perceptron;

[0113] Step 1.9, concatenate the title feature representation, category feature representation, and subcategory feature representation of the i-th news in the historical click news hc in group g, and obtain the feature representation of the i-th click news in the group through multi-layer perceptron cross fusion for:

[0114]

[0115] In formula (8), || is the concatenation operation, W n is the learnable parameter of the fully connected layer of news features, b n is the bias of the fully connected layer;

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

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

[0118]

[0119] In formula (9), V i,c and b i,c The learnable parameters contributed by the category, q i,c The attention query vector contributed to the category;

[0120] Step 1.12: Assume that each news in group g belongs to a category. Considering the categories corresponding to all news in group g, the normalized attention weight of category c of the i-th news is Denoted as:

[0121]

[0122] In formula (10), N+1 is the number of clicked news and candidate news in group g, p represents the pth news, is the feature of category c of the pth news in group g Contribution score;

[0123] Step 1.13, based on the normalized attention weight, the category feature representations of all news in group g are fused, and the category feature representation within group g is for:

[0124]

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

[0126] Step 1.14: The paragraph number feature [CLS] of the text content of group g is encoded using the Bert pre-trained language model and used as the intra-group paragraph number feature representation of group g.

[0127] Step 1.15: Represent the intra-group category features And the paragraph number feature within the group Splicing, feature fusion is performed through a multi-layer perceptron to obtain the intra-group feature representation s g for:

[0128]

[0129] In formula (12), || is the concatenation operation, W s is the learnable parameter of the fully connected layer, b s is the bias of the fully connected layer.

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

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

[0132] Step 2.1: Input the intra-group feature representations of all click news and candidate news in G groups into BiLSTM for crossover. For the i-th intra-group click news feature representation of group g, Defining cross-semantic feature representation of click news for:

[0133]

[0134] In formula (13), r g(candi) is the feature representation of candidate news within the group of candidate news candi;

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

[0136] Step 3: The cross-semantic click news feature representation, the cross-semantic candidate news feature representation, and the intra-group feature representation are fused through the attention mechanism to model the inter-group-aware click news feature representation, the inter-group-aware candidate news feature representation, and the inter-group-aware group feature representation;

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

[0138] Step 3.1: Considering that all clicked news of a user have different contributions to the user's interest expression, the contribution score of the i-th clicked news of the g-th group of user u is for:

[0139]

[0140] In formula (14), V u and v u is the weight parameter, q u is the attention query vector;

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

[0142]

[0143] In formula (15), G is the number of groups, N is the number of clicked news in a group, x is the counting loop variable of the group, and y is the counting loop variable of the news in the group. is the contribution score of the yth clicked news in the xth group of user u;

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

[0145]

[0146] In formula (16), represents the normalized attention weight of the yth clicked news of the xth group of user u among all the clicked news of the user, The click news feature representation of the yth click news in group x after performing cross-group cross-semantics;

[0147] Step 3.4: Repeat steps 3.1-3.3 to fuse the cross-semantic candidate news feature representations of G different groups through the attention mechanism to obtain the inter-group perceived candidate news feature representation. The intra-group feature representations of G different groups are fused through the attention mechanism to obtain the inter-group perceived group feature representation

[0148] Step 4: A gating mechanism is used to aggregate the important features of the inter-group perceived click news feature representation and the inter-group perceived group feature representation to model the user interest representation, and a gating mechanism is used to aggregate the important features of the inter-group perceived candidate news feature representation and the inter-group perceived group feature representation to model the candidate news representation;

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

[0150] Step 4.1, select the important inter-group-aware group feature representation and the important inter-group-aware click news feature representation through the gating mechanism, which are denoted as:

[0151]

[0152] In formula (17), m is the gate vector of the inter-group-aware group feature representation and the inter-group-aware click news feature representation; W m is the learnable parameter matrix; b m represents the parameter vector;

[0153] Step 4.2: Aggregate inter-group-aware group feature representation and inter-group perceived click news representation Modeling user interest representation u for:

[0154]

[0155] Step 4.3, repeat steps 4.1-4.2, and aggregate the group feature representation of inter-group awareness through the gating mechanism and inter-group-aware candidate news feature representation Get candidate news representation

[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:

[0158] Step 5.1: The predicted score of user u for candidate news candi is represented by user interest r u and candidate news The interest score is obtained through the inner product, expressed as:

[0159]

[0160] Step 5.2: Select the news clicked by the user as the positive sample, and randomly select K news that have not been clicked in the same session as the negative sample. Calculate the interest scores of the positive and negative samples as and And define the NCE loss function as:

[0161]

[0162] In formula (20), P is the set of positive samples, is the prediction score of the jth negative sample corresponding to the i-th positive sample.

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

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

[0165] For unclicked candidate news, the user's interest score for the candidate news is calculated, and the top k news with higher interest scores are recommended to the user.

[0166] In order to verify the effectiveness of the method, experiments were conducted on the MIND dataset (MIND (msnews.github.io)), which contains user-related information and news content information. User information covers the user's browsing and clicking news, click time, and user log data. User log data contains the user's candidate news and interest tags for the candidate news, where 1 indicates that the user is interested and 0 indicates that the user is not interested. The 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

[0168]

[0169] We use three news features, namely news titles, categories and subcategories. The evaluation indicators are AUC, MRR, and nDCG@5. In order to verify the effectiveness and advancement of the technical solution proposed in this invention, several existing recommendation prediction model methods are selected for comparison: DKN, NPA, NRMS, FIM and UNbert. The experimental results are shown in Table 2:

[0170] Table 2 Experimental results

[0171]

[0172] It can be seen from the results in Table 2 that the technical solution of the present invention can obtain detection results with better accuracy and reliability than existing methods when predicting user interest in news.

[0173] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A personalized news recommendation method based on time-series grouping user interest representation, characterized in that: The following steps are involved: Step 1: Sequentially group the user's historical clicked news based on a sliding time window, merge candidate news in different groups, design a feature encoder within the same group, model the news title, news category, and news subcategory content, and 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 2: Design an inter-group feature encoder to capture the semantic correlation of news between different groups, and model the cross-semantic click news feature representation and the cross-semantic candidate news feature representation; Step 3: The cross-semantic click news feature representation, the cross-semantic candidate news feature representation, and the intra-group feature representation are fused through the attention mechanism to model the inter-group-aware click news feature representation, the inter-group-aware candidate news feature representation, and the inter-group-aware group feature representation; Step 4: A gating mechanism is used to aggregate the important features of the inter-group perceived click news feature representation and the inter-group perceived group feature representation to model the user interest representation, and a gating mechanism is used to aggregate the important features of the inter-group perceived candidate news feature representation and the inter-group perceived group feature representation 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: Calculate the user's interest score for the unclicked candidate news based on the optimized user interest representation and candidate news representation, sort them according to the score, and generate a recommendation list.

2. The personalized news recommendation method based on time-series grouping user interest representation according to claim 1 is characterized in that: In step 1, the user's historical clicked news is grouped in time series based on a sliding time window, and candidate news is merged in different groups. A feature encoder is designed within the same group, news titles, news categories, and news subcategory content are modeled, and feature representations of clicked news within the group, feature representations of candidate news within the group, and feature representations of group within the group are learned. The specific steps are: Step 1.1: divide the user's historical click news into G groups {S1,…,S g ,…,S G },S G is the historical click news set of group G, and the candidate news candi is added to each group in turn. The group set includes the historical click news and candidate news of each group, which is expressed as S1∪candi,…,S g ∪candi,…,S G ∪candi; Step 1.2: In the g-th group, use the Bert pre-trained language model as the text encoder to convert the candidate news and historical click news title word sequences in the group As input, the feature representation of the words between output titles is: In formula (1), candi is the candidate news in group g, hc is all the historical clicked news in group g, N is the number of historical clicked news in each group, l is the number of pre-trained words in the news title, and N l is the lth pre-trained word of the Nth news; i represents the i-th news, i l Represents the lth pre-trained word of the i-th news; Step 1.3: Use the convolutional neural network to train the jth word w of the i-th news in the historical click news hc in group g j The feature representation Further encoding is performed, and the word features of cross-semantics are represented as: In formula (2), W' is the learnable parameter of CNN filter, b' is the bias, * is the convolution operator, is the feature representation of the words from position (jf) to (j+f) of the i-th news; Step 1.4, according to the different importance of each word in the news title, calculate the jth word w of the i-th news hc in the historical click news in group g j Contribution score In formula (3), V t and b t is the weight parameter, q is the attention query vector; T is the transposition operation; Step 1.5, calculate the jth word w j Normalized attention weights for all words in a news title for: In formula (4), is the j'th word w of the i-th news in group g j' The contribution score of ; j' represents the j'th word; exp is an exponential function; Step 1.6, according to the normalized attention weight, the cross-semantic word feature representations of all words in the news title of the i-th news in the historical click news hc in group g are fused to define the title feature representation for: Step 1.7, one-hot encode the news category c of the i-th news of the historical click news hc in group g, and use the query table to initialize it as And the category feature representation of the i-th news is obtained through the multi-layer perceptron for: In formula (6), V c and b c is the parameter in the category feature multilayer perceptron; Step 1.8, one-hot encode the news subcategory sc of the i-th news of the historical click news hc in group g, and use the query table to represent it as And the subcategory feature representation of the i-th news is obtained through the multi-layer perceptron for: In formula (7), V s and b s is the parameter in the subcategory feature multilayer perceptron; Step 1.9, concatenate the title feature representation, category feature representation, and subcategory feature representation of the i-th news in the historical click news hc in group g, and obtain the feature representation of the i-th click news in the group through multi-layer perceptron cross fusion for: In formula (8), || is the concatenation operation, W n is the learnable parameter of the fully connected layer of news features, b n is the bias of the fully connected layer; Step 1.10, repeat steps 1.3-1.9, and model the candidate news features within group g as r g(candi) ; Step 1.11, news category is a feature of the group. According to the different importance of news categories in group g, the feature of category c of the i-th news Contribution score for: In formula (9), V i,c and b i,c The learnable parameters contributed by the category, q i,c The attention query vector contributed to the category; Step 1.12: Assume that each news in group g belongs to a category. Considering the categories corresponding to all news in group g, the normalized attention weight of category c of the i-th news is Denoted as: In formula (10), N+1 is the number of clicked news and candidate news in group g, p represents the pth news, is the feature of category c of the pth news in group g Contribution score; Step 1.13, based on the normalized attention weight, the category feature representations of all news in group g are fused, and the category feature representation within group g is for: In formula (11), N+1 is the number of clicked news and candidate news in the group; Step 1.14: The paragraph number feature [CLS] of the text content of group g is encoded using the Bert pre-trained language model and used as the intra-group paragraph number feature representation of group g. Step 1.15: Represent the intra-group category features And the paragraph number feature within the group Splicing, feature fusion is performed through a multi-layer perceptron to obtain the intra-group feature representation s g for: In formula (12), || is the concatenation operation, W s is the learnable parameter of the fully connected layer, b s is the bias of the fully connected layer.

3. The personalized news recommendation method based on time-series grouping user interest representation according to claim 1 is characterized in that: In step 2, between different groups, an inter-group feature encoder is designed to capture the news semantic correlation between cross-groups, and model the cross-semantic click news feature representation and the cross-semantic candidate news feature representation. The specific steps are: Step 2.1: Input the intra-group feature representations of all click news and candidate news in G groups into BiLSTM for crossover. For the i-th intra-group click news feature representation of group g, Defining cross-semantic feature representation of click news for: In formula (13), r g(candi) is the feature representation of candidate news within the group of candidate news candi; Step 2.2: Repeat step 2.1 to obtain the candidate news features of cross-semantics represented as r g,candi .

4. The personalized news recommendation method based on time-series grouping user interest representation according to claim 1 is characterized in that: In step 3, the cross-semantic click news feature representation, the cross-semantic candidate news feature representation, and the intra-group feature representation are fused through the attention mechanism to model the inter-group-aware click news feature representation, the inter-group-aware candidate news feature representation, and the inter-group-aware group feature representation. The specific steps are: Step 3.1: Considering that all clicked news of a user have different contributions to the user's interest expression, the contribution score of the i-th clicked news of the g-th group of user u is for: In formula (14), V u and v u is the weight parameter, q u is the attention query vector; Step 3.2, the normalized attention weight of the i-th clicked news of the g-th group of user u in all the clicked news of the user is: In formula (15), G is the number of groups, N is the number of clicked news in a group, x is the counting loop variable of the group, and y is the counting loop variable of the news in the group. is the contribution score of the yth clicked news in the xth group of user u; Step 3.3, based on the attention mechanism, the cross-semantic click news feature representations of G different groups are fused, and the inter-group perceived click news feature representation is defined as: In formula (16), represents the normalized attention weight of the yth clicked news of the xth group of user u among all the clicked news of the user, The click news feature representation of the yth click news in group x after performing cross-group cross-semantics; Step 3.4: Repeat steps 3.1-3.3 to fuse the cross-semantic candidate news feature representations of G different groups through the attention mechanism to obtain the inter-group perceived candidate news feature representation. The intra-group feature representations of G different groups are fused through the attention mechanism to obtain the inter-group perceived group feature representation 5. The personalized news recommendation method based on time-series grouping user interest representation according to claim 1 is characterized in that: In step 4, the gating mechanism is used to aggregate the important features of the click news feature representation perceived between groups and the group feature representation perceived between groups, and the user interest representation is modeled. The gating mechanism is also used to aggregate the important features of the candidate news feature representation perceived between groups and the group feature representation perceived between groups, and the candidate news representation is modeled. The specific steps are: Step 4.1, select the important inter-group-aware group feature representation and the important inter-group-aware click news feature representation through the gating mechanism, which are denoted as: In formula (17), m is the gate vector of the inter-group-aware group feature representation and the inter-group-aware click news feature representation; W m is the learnable parameter matrix; b m represents the parameter vector; Step 4.2: Aggregate inter-group-aware group feature representation and inter-group perceived click news representation Modeling user interest representation u for: Step 4.3, repeat steps 4.1-4.2, and aggregate the group feature representation of inter-group awareness through the gating mechanism and inter-group-aware candidate news feature representation Get candidate news representation 6. The personalized news recommendation method based on time-series grouping user interest representation according to claim 1, 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 is calculated, the interest score is optimized based on the negative sampling loss function, and the model parameters are updated. The specific steps are: Step 5.1: The predicted score of user u for candidate news candi is represented by user interest r u and candidate news The interest score is obtained through the inner product, expressed as: Step 5.2: Select the news clicked by the user as the positive sample, and randomly select K news that have not been clicked in the same session as the negative sample. Calculate the interest scores of the positive and negative samples as and And define the NCE loss function as: In formula (20), P is the set of positive samples, is the prediction score of the jth negative sample corresponding to the i-th positive sample.

7. The personalized news recommendation method based on time-series grouping user interest representation according to claim 1 is characterized in that: In step 6, the user's interest score for the unclicked candidate news is calculated based on the optimized user interest representation and candidate news representation, and the candidate news is sorted according to the score to generate a recommendation list. The specific steps are: For unclicked candidate news, the user's interest score for the candidate news is calculated, and the top k news with higher interest scores are recommended to the user.

Images