A method for recommending news without bias by fusing crowd modeling

Abstract

The application discloses a kind of fusion from crowd modeling debiasing news recommendation method, its steps include:1, the factor that decides user click behavior in news recommendation is analyzed, and causal diagram is constructed based on observed phenomenon;2, based on the causal diagram established by construction, establish the debiasing news recommendation model of fusion news popularity and the modeling of the crowd characteristic of user;3, utilize news data set to construct more challenging debiasing news recommendation data set, and the model established is trained using negative sampling method;4, realize the click rate estimation task of user to news using the debiasing news recommendation model established goodly.The application can pay more attention to the complex factor behind user click behavior, so as to realize higher quality debiasing news recommendation.

Description

Technical Field

[0001] This invention relates to the field of news recommendation, specifically a novel debiased news recommendation method that integrates conformity modeling. Background Technology

[0002] Personalized news recommendation technology helps users find news of interest from a vast sea of ​​information, which is of paramount importance for improving the service level of news websites and the user reading experience. The key lies in learning accurate news and user representations. A large body of research focuses on mining user preferences from their news browsing history to achieve accurate user preference-based news recommendations.

[0003] Traditional news recommendation methods rely on manual feature engineering to learn news and user representations. However, these methods require a large amount of prior knowledge and do not take into account the context and sequence information of the news. In recent years, deep learning methods have been applied to news recommendation. Techniques such as convolutional neural networks, recurrent neural networks, and attention mechanisms are widely used to extract rich semantic representations from news and model user preferences from users' historical click records. These methods have improved the accuracy of news recommendation to some extent.

[0004] Existing news recommendation methods still suffer from over-reliance on biased information. To alleviate this dependence, various debiasing methods have been proposed, such as decoupled representation learning and adversarial learning. Integrating causal inference methods to achieve unbiased recommendations is a promising research direction. By utilizing causal inference techniques to distinguish spurious associations in observed data, the model can focus more on the true reasons reflecting user preferences, thus achieving unbiased recommendations.

[0005] While these methods have achieved significant results, users' news clicking behavior is the result of a series of complex cognitive behaviors interacting with each other. Existing unbiased methods only focus on single factors (such as news popularity) and fail to accurately analyze the complex cognitive factors hidden behind clicking behavior. In user preference modeling and unbiased news recommendation, users' own cognitive characteristics are also an important influencing factor that requires sufficient attention. Furthermore, how to accurately model user preferences and analyze user cognitive characteristics based on historical click data remains a challenging problem. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a bias-free news recommendation method that integrates conformity modeling, aiming to pay more attention to the complex factors behind user click behavior and thus achieve higher-quality bias-free news recommendations.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0008] The present invention discloses a bias-free news recommendation method that integrates conformity modeling, characterized by the following steps:

[0009] Step 1: Obtain the user's historical interaction news dataset Q, and denote the historical click news sequence of any user u as D. H The candidate news sequence is denoted as D. C The candidate news sequence D C Let x be any one of the candidate news items. c Let the candidate news x c The click tag is y c y c ∈{0,1}, where y c =1 indicates that user u clicked on candidate news x c y c =0 indicates that user u did not click on the candidate news x c The positive sample set consists of all the candidate news items clicked by user u. The negative sample set consists of all candidates that user u has not clicked.

[0010] Let D∈{D H D C} represents the historical news click sequence D for user u. H and candidate news sequence D C Given any news article D; preprocess news article D to obtain the word sequence W of the title of D. D Subject ID number T D Subtopic ID number V D ;

[0011] Step 2: Analyze user u's click behavior to construct a causal graph based on the matching degree between user preferences and news, the popularity of news, and the user's herd mentality.

[0012] Step 3: Construct a news recommendation model, including: a user preference and news matching module, a fusion of herd mentality model, MLP network, and a user click behavior prediction module;

[0013] Step 3.1: The user preference and news matching module is used to obtain the matching result m between user preferences and news. uc ;

[0014] Step 3.2: The fusion conformity model is used to obtain the historical click news sequence D of user u. H Popularity results of all news articles H and candidate news sequence D C Candidate news x cPopularity results c And preliminary results on user u's conformity traits

[0015] Step 3.3, Preliminary results of the conformity trait of user u. The input is processed in the MLP network to obtain the weights of user u.

[0016] The historical click news sequence D H Popularity results of all news articles in China H The data is input into an MLP network for processing to obtain historical click news (D). H weight w a ;

[0017] Step 3.4: Use equation (1) to obtain the final conformity trait result c of user u. u :

[0018]

[0019] In equation (1), Softplus(·) represents the activation function;

[0020] Step 3.5: The user click behavior prediction module predicts user u's click behavior on candidate news x using equation (2). c Click probability score results

[0021]

[0022] Step 4: Utilize the negative log-likelihood loss function Iteratively train the news recommendation model;

[0023] Step 4.1: Use the negative sampling method on the positive sample set. Sampling a positive sample Meanwhile in the negative sample set Randomly sample K negative samples Thus, positive samples are obtained using equation (3). Click probability score and K negative samples Click probability score in, Let j represent the j-th negative sample, i.e., the j-th candidate news item that was not clicked. Represents the j-th negative sample Click probability score;

[0024] Positive samples are obtained using equation (3). Posterior click probability score p ui :

[0025]

[0026] Step 4.2: Construct the negative log-likelihood loss function using equation (4).

[0027]

[0028] In equation (4), p uD This represents the posterior click probability score of user u for news item D;

[0029] Step 4.3: Based on user u's historical click news sequence D H and candidate news sequence D C The news recommendation model is trained, and the AdamW optimizer is used to optimize the negative log-likelihood loss function. The optimal news recommendation model is obtained by minimizing the solution, thereby optimizing all parameters of the model and predicting the click-through rate of users on news.

[0030] Step 5: In the reasoning stage, the news popularity result s c Set as a constant value s * c u Set each to a fixed value c * Thus, the optimal news recommendation model predicts user u's preference for candidate news x using equation (5). c Click probability score

[0031]

[0032] The characteristic of the bias-free news recommendation method that integrates conformity modeling described in this invention is that the user preference and news matching degree module in step 3.1 includes: news encoder Enc1, user encoder Enc2, and click predictor;

[0033] Step 3.1.1: Extract the word sequence W from the title of news article D. D Subject ID number T D and the ID number V of the subtopic D The news is input into the news encoder Enc1 for processing and outputs a news representation r, thereby obtaining the historical click news sequence D of user u. H All news representations r H and candidate news sequence D C Candidate News x c The characterization of r c ;

[0034] Step 3.1.2, the historical click news sequence D of user u.H All news representations r H The input is processed by the user encoder Enc2, and the user representation u is output.

[0035] Step 3.1.3, the candidate news x c The characterization of r d After performing an inner product operation on the user representation u input into the click predictor, the user u's click candidate news x is obtained. c The predicted probability is used as the matching result between user preferences and news. uc .

[0036] The fusion conformity model in step 3.2 includes: a news popularity module and a user conformity trait module TAUC; wherein, the conformity trait module TAUC is composed of coarse-grained topic units and fine-grained sub-topic units;

[0037] Step 3.2.1: Observe and statistically analyze news D in the user's historical interaction news dataset Q to obtain the click-through rate (ctr) of news D in the user's historical interaction news dataset Q. D and the probability of occurrence p D ;

[0038] Step 3.2.2: Calculate the click-through rate (CTR) of news item D. D and the probability of occurrence p D After being concatenated, the data is input into the news popularity module for processing to obtain the popularity result s of news D. D Thus, the historical click news sequence D of user u is obtained. H Popularity results of all news articles H and candidate news sequence D C Candidate news x c Popularity results c ;

[0039] Step 3.2.3: Analyze the historical click news sequence D of user u. H By observing and statistically analyzing, we can obtain the topic distribution of the news articles that user u clicked historically. Sub-topic distribution

[0040] Step 3.2.4: Observe and statistically analyze news topics and subtopics in the user's historical interaction news dataset Q to obtain the topic distribution of historical news data. Sub-topic distribution

[0041] Step 3.2.5: The coarse-grained topic unit is based on the topic distribution of the news articles clicked by user u in history. Thematic distribution of historical news data Calculate the KL divergence to obtain the topic-based user u conformity trait results.

[0042] Step 3.2.6: The fine-grained sub-topic units are distributed according to the sub-categories of the news articles clicked by user u in history. Subclass distribution of news historical data Calculate the KL divergence to obtain the conformity trait results of user u based on subtopics.

[0043] Step 3.2.7: Analyze the topic-based user u conformity trait results. Results based on the conformity trait of user u (subclass-based) The concatenated data serves as the initial result of user u's conformity traits.

[0044] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the method, and the processor is configured to execute the program stored in the memory.

[0045] This invention discloses a computer-readable storage medium on which a computer program is stored, characterized in that the computer program, when executed by a processor, performs the steps of the method. Compared with the prior art, the beneficial effects of this invention are as follows:

[0046] 1. This invention proposes that user click behavior is jointly determined by the matching degree between user preferences and news, the popularity of news, and the user's herd mentality. Based on observed phenomena, a causal graph is constructed to achieve accurate modeling and analysis of user preferences in unbiased news recommendation, thereby improving the accuracy of news recommendation.

[0047] 2. This invention proposes a novel causal debiased news recommendation method that integrates conformity modeling. It considers both news popularity and users' conformity traits during the news recommendation process, enabling more accurate conformity modeling. In the model inference stage, fixed values ​​are set for news popularity and users' conformity traits to ensure that predicted user click behavior aligns with user preferences, thereby achieving higher-quality debiased news recommendations.

[0048] 3. This invention improves the robustness of the model in an OOD environment by constructing an unbiased challenge test set through random sampling of data samples in the test set with a certain probability. Attached Figure Description

[0049] Figure 1 This is a structural diagram of the method for integrating the herd mentality modeling into the bias-free news recommendation method of the present invention;

[0050] Figure 2 This is a causal graph constructed based on the analysis results of user click behavior in this invention;

[0051] Figure 3 This is a diagram illustrating the effects of different components of the TAUC module representing the conformity traits of users in this invention.

[0052] Figure 4 This is the intervention causal graph of the method of the present invention during the model reasoning stage. Detailed Implementation

[0053] In this embodiment, a bias-free news recommendation method that integrates conformity modeling is used, such as... Figure 1 As shown, the method is performed according to the following steps:

[0054] Step 1: Obtain the user's historical interaction news dataset Q, and denote the historical click news sequence of any user u as D. H The candidate news sequence is denoted as D. C Candidate news sequence D C Let x be any one of the candidate news items. c , making candidate news x c The click tag is y c y c ∈{0,1}, where y c =1 indicates that user u clicked on candidate news x c y c =0 indicates that user u did not click on the candidate news x c The positive sample set consists of all the candidate news items clicked by user u. The negative sample set consists of all candidates that user u has not clicked. Let D∈{D H D C} represents the historical news click sequence D for user u. H and candidate news sequence D C Given any news article D; preprocess news article D to obtain the word sequence W of the title of D. D Subject ID number T D Subtopic ID number V D ;

[0055] Let D∈{D H D C} represents the historical news click sequence D for user u. H and candidate news sequence D C Given any news article D; preprocess news article D to obtain the word sequence W of the title of D. D Subject ID number T D Subtopic ID number VD ;

[0056] In a specific example, this invention uses the Small dataset from the publicly available news dataset MIND for training and testing. The dataset contains 49,146 users and 65,239 news articles. Each user has an impression log, which includes the user ID, time, the user's historical clicked news items, the news items contained in the impression, and the user's click behavior on the news items (0 for non-click, 1 for click). Each news item includes a news ID, topic, subtopic, title, and summary. Using GloVe pre-training on the words in the news D titles, a word vector is obtained for each word, forming a word sequence W. D The dataset Q contains 16 topics and 153 subtopics. The topics and subtopics are numbered sequentially to obtain the topic ID T of news item D. D Subtopic ID number V D ;

[0057] Construction of the challenge test set: 70% of the interaction data is randomly selected from the general test set to form a new biased dataset, where the probability of each sample being sampled is the reciprocal of the frequency of the news occurrence.

[0058] Step 2: Analyze user u's click behavior to construct a causal graph based on the match between user preferences and news, news popularity, and the user's herd mentality. For example... Figure 2 As shown, (U,I)→M→Y represents the degree of matching between users' true preferences and news features, which is what the model hopes to learn from the data; U→C and I→S represent explicit modeling of users' conformity traits and news popularity attributes, respectively; (C,S)→R→Y represents the impact of the interaction between users' conformity and news popularity on the final prediction results of user click behavior.

[0059] Step 3: Construct a bias-free news recommendation model that integrates conformity modeling, including a user preference and news matching module, a conformity model, an MLP network, and a user click behavior prediction module.

[0060] The user preference and news matching module utilizes existing news recommendation models NRMS and NAML, and consists of three parts: news encoder Enc1, user encoder Enc2, and click predictor; the fusion conformity model includes: news popularity module and user conformity trait module TAUC; among which, the conformity trait module TAUC consists of coarse-grained topic units and fine-grained sub-topic units.

[0061] Step 3.1: Extract the word sequence W from the headline of news article D. D Subject ID number T D and the ID number V of the subtopicD The news is processed by the news encoder Enc1 and outputs a news representation r, thus obtaining the historical click news sequence D of user u. H All news representations r H and candidate news sequence D C Candidate News x c The characterization of r c ;

[0062] In specific implementations, the news encoder in NRMS includes a word embedding layer, a word-level multi-head self-attention layer, and a word-level attention layer; the user encoder includes a news-level multi-head self-attention network layer and an additive news attention layer; the news encoder in NAML consists of three parts: a title encoder, a type encoder, and attention pooling. The title encoder includes a word embedding layer, a convolutional neural network layer, and a word-level attention layer.

[0063] Step 3.2, User u's historical click news sequence D H All news representations r H The input is processed in the user encoder Enc2, and the user representation u is output.

[0064] In specific implementations, the user encoder in NRMS includes a news-level multi-head self-attention network layer and an additive news attention layer; the user encoder in NAML consists of a news attention network.

[0065] Step 3.3: Select candidate news items x c The characterization of r d After performing an inner product operation with the user representation u in the click predictor, we obtain the candidate news item x that user u clicked. c The predicted probability is used as the matching result between user preferences and news. uc ;

[0066] Step 3.4: Observe and statistically analyze news item D in the user's historical interaction news dataset Q to obtain the click-through rate (ctr) of news item D in the user's historical interaction news dataset Q. D and the probability of occurrence p D ;

[0067] In practice

[0068]

[0069] Click rate of news D (ctr) D and the probability of occurrence p D After concatenation, the data is input into the news popularity module for processing, yielding the popularity result s of news D. D Thus, the historical click news sequence D of user u is obtained.H Popularity results of all news articles H and candidate news sequence D C Candidate news x c Popularity results c ;

[0070] The news popularity module inputs the click-through rate (ctr) of news item D. D and the probability of occurrence p D The two results are combined to obtain the popularity result s of news D. D Thus, the historical click news sequence D of user u is obtained. H Popularity results of all news articles H and candidate news sequence D C any news item x c Popularity results c ;

[0071] Step 3.5: Analyze the historical click sequence D of user u. H By observing and statistically analyzing, we can obtain the topic distribution of the news articles that user u clicked historically. Sub-topic distribution

[0072] By observing and statistically analyzing news topics and subtopics in the user's historical interaction news dataset Q, the topic distribution of historical news data is obtained. Sub-topic distribution

[0073] In practice, the topic distribution of news clicked in a user's history is determined by the sequence of news clicked in the user's history, D. H The probability composition of each type of topic, i.e. The thematic distribution of historical news data is composed of the probability of each thematic category appearing in the historical news data, i.e.

[0074] The topic distribution of user's historical click news is based on the user's historical click news sequence D. H The probability composition of each sub-topic category, i.e. The thematic distribution of historical news data is constituted by the probability of each sub-theme appearing in the historical news data, i.e.

[0075] Step 3.5.1: Coarse-grained topic units are based on the topic distribution of news clicks by user u in history. Thematic distribution of historical news data Calculate the KL divergence to obtain the topic-based user u conformity trait results. The operation of this step is not exactly the same for different backbone models, NRMS and NAML. In specific implementations, NMRS directly uses the subclass-based user conformity trait results. As NAML uses the result of coarse-grained and fine-grained splicing. like Figure 3 As shown;

[0076] Step 3.5.2: Fine-grained sub-topic units are distributed based on the sub-categories of news clicked by user u in the past. Subclass distribution of news historical data Calculate the KL divergence to obtain the conformity trait results of user u based on subtopics.

[0077] The results of topic-based user conformity traits Results based on the conformity trait of user u (subclass-based) The concatenated data serves as the initial result of user u's conformity traits.

[0078] The initial conformity trait results of user u The input is processed in an MLP network to obtain the weights of user u.

[0079] Click on the historical news sequence D H Popularity results of all news articles in China H The data is input into an MLP network for processing to obtain the weights w of historically clicked news articles. a ;

[0080] Using equation (1), we obtain the final conformity trait result c of user u. u :

[0081]

[0082] In equation (1), Softplus(·) represents the activation function;

[0083] Step 3.6: The user click behavior prediction module predicts user u's click behavior on candidate news x using formula (2). c Click probability score results

[0084]

[0085] In equation (2), s c c represents the effect of news popularity (corresponding to I→S in the causal graph). u This represents the effect of user u's conformity (corresponding to U→C in the causal graph);

[0086] Step 4: Utilize the negative log-likelihood loss function Iteratively train the news recommendation model;

[0087] Step 4.1: Use negative sampling on the positive sample set. Sampling a positive sample Meanwhile in the negative sample set Randomly sample K negative samples Thus, positive samples are obtained using equation (3). Click probability score and K negative samples Click probability score in, Let j represent the j-th negative sample, i.e., the j-th candidate news item that was not clicked. Represents the j-th negative sample Click probability score;

[0088] Positive samples are obtained using equation (3). Posterior click probability score p ui :

[0089]

[0090] In practice, the negative sample sampling ratio K = 4. When K is too small, the number of negative samples used for training is very small, and the useful information provided by the negative samples cannot be fully utilized. When K is too large, the model has difficulty correctly identifying positive samples.

[0091] Step 4.2: Analyze the candidate news sequence D. C The negative log-likelihood loss function is constructed using equation (4) for all samples.

[0092]

[0093] In equation (4), p uD This represents the posterior click probability score of user u for news item D.

[0094] Step 4.3: Based on user u's historical click news sequence D H and candidate news sequence D C The news recommendation model was trained, and the AdamW optimizer was used to optimize the negative log-likelihood loss function. The optimal news recommendation model is obtained by minimizing the solution, thereby optimizing all parameters of the model and predicting the click-through rate of users on news.

[0095] Step 5: In the reasoning stage, the news popularity result s c Set as a constant value s *c u Set each to a fixed value c * ,like Figure 4 As shown, the aim is to ensure that the click-through rate predicted by the model is consistent with the user's preferences, thereby achieving a bias-free effect. The optimal news recommendation model then predicts the user u's preference for candidate news x using equation (5). c Click probability score

[0096]

[0097] In specific implementation, the s set in step 5 * and c * The value is not fixed, but it will not be set to 0 at the same time. It is necessary to ensure the interpretability of the final prediction result.

[0098] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0099] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0100] The optimal model obtained in step 4.3 was tested using both a general test set and a constructed challenge test set. The AUC (area under the receiver operating characteristic curve), MRR (mean reciprocal ranking), and NDCG (normalized depreciation cumulative gain) on the test set were used as evaluation indicators. The larger the AUC, the better the prediction effect of the model. The higher the MRR and NDCG, the higher the recommendation quality of the model. The experimental results are shown in Tables 1 and 2.

[0101] Table 1 compares the performance of various models based on the MIND Small dataset with NRMS as the backbone model.

[0102]

[0103] Table 2 compares the performance of various models based on the MIND_Small dataset with NAML as the backbone.

[0104]

[0105]

[0106] As shown in Tables 1 and 2, this invention improves the accuracy of prediction and the quality of recommendation on different evaluation metrics and test sets for different backbone models NRMS and NAML. This demonstrates that by deeply modeling the news popularity, user herd mentality, and the degree of matching between user preferences and news that influence user click behavior, biased modeling of user preferences is achieved, thereby improving the quality of biased news recommendations.

[0107] Table 3. The impact of different modules on user click behavior

[0108]

[0109] Table 3 shows that, without considering the user's conformity trait (w / oc) module u ) or the news popularity module (w / os) d In the case of [missing information], the performance of the present invention on the constructed challenge test set decreased to varying degrees, indicating the importance of these two modules in user preference modeling. Among them, the user conformity trait module has a greater impact on the model performance, indicating that the user conformity modeling method TAUC proposed in this invention achieves more accurate user conformity modeling.

Claims

1. A bias-free news recommendation method that integrates conformity modeling, characterized by: Follow these steps: Step 1: Obtain the user's historical interaction news dataset and any one of the users The historical click news sequence is recorded as Candidate news sequences are denoted as The candidate news sequence Any one of the candidate news items is , making the candidate news The clicked tags are , {0,1}, where, =1 indicates user Clicked on candidate news , =0 indicates user Unclicked candidate news By user All clicked candidate news items constitute the positive sample set. By user All unclicked candidates constitute the negative sample set. ; make Indicates user Historical Click News Sequence and candidate news sequence Any news item in the text; regarding the news Preprocessing is performed to obtain The word sequence of the title Subject ID number Subtopic ID number ; Step 2: Analyze users The click behavior is used to construct a causal graph based on the matching degree between user preferences and news, the popularity of news, and the user's herd mentality. Step 3: Construct a news recommendation model, including: a user preference and news matching module, a fusion of herd mentality model, MLP network, and a user click behavior prediction module; Step 3.1: The user preference and news matching module is used to obtain the matching results between user preferences and news. ; Step 3.2: The fusion conformity model is used to obtain user... Historical Click News Sequence Popularity results of all news and candidate news sequence Candidate news in Popularity results and users Preliminary results of conformity traits ; Step 3.3, the user Preliminary results of conformity traits The input is processed in the MLP network to obtain the user's information. weight ; The historical click news sequence Popularity results of all news in China The data is input into an MLP network for processing to obtain historical click news. weight ; Step 3.4: Use equation (1) to obtain the user's information. The final result of conformity traits : (1) In equation (1), Indicates the activation function; Step 3.5: The user click behavior prediction module predicts the user's click behavior using formula (2). News on the candidates Click probability score results : (2) Step 4: Utilize the negative log-likelihood loss function Iteratively train the news recommendation model; Step 4.1: Use the negative sampling method on the positive sample set. Sampling a positive sample Meanwhile, in the negative sample set Random sampling negative samples Thus, positive samples can be obtained using equation (3). Click probability score and negative samples Click probability score ;in, Indicates the first The nth negative sample, i.e., the nth negative sample Then the candidate news that has not been clicked, Indicates the first negative samples Click probability score; Positive samples are obtained using equation (3). Posterior click probability score : (3) Step 4.2: Construct the negative log-likelihood loss function using equation (4). : (4) In equation (4), Indicates user News The posterior click probability score; Step 4.3, based on user Historical Click News Sequence and candidate news sequence The news recommendation model is trained, and the AdamW optimizer is used to optimize the negative log-likelihood loss function. The optimal news recommendation model is obtained by minimizing the solution, thereby optimizing all parameters of the model and predicting the click-through rate of users on news. Step 5: In the reasoning stage, analyze the news popularity results. Set as a constant value herd mentality among users Set to constant values ​​respectively Thus, the optimal news recommendation model predicts user behavior through equation (5). News on the candidates Click probability score : (5)。 2. The method for bias-free news recommendation based on conformity modeling according to claim 1, characterized in that, The user preference and news matching module in step 3.1 includes: a news encoder. User encoder Click predictor; Step 3.1.1: Transfer the news The word sequence of the title Subject ID number and the ID number of the child topic Input the news encoder The data is processed and a news representation is output. Thus gaining users Historical Click News Sequence All news representations and candidate news sequence News from the candidate The representation ; Step 3.1.2, the user Historical Click News Sequence All news representations Input the user encoder The process is performed and the user representation is output. ; Step 3.1.3, the candidate news The representation and user representation After performing the inner product operation on the input click predictor, the user's... Click on candidate news The predicted probability is used as the matching result between user preferences and news. .

3. The method for bias-free news recommendation based on conformity modeling according to claim 2, characterized in that, The fusion conformity model in step 3.2 includes: a news popularity module and a user conformity trait module TAUC; wherein, the conformity trait module TAUC is composed of coarse-grained topic units and fine-grained sub-topic units; Step 3.2.1, in the user's historical interaction news dataset China News Observation and statistics were conducted to obtain news. In user history interaction news dataset Click-through rate and probability of occurrence ; Step 3.2.2: Transfer the news Click-through rate and probability of occurrence After being spliced ​​together, the news is input into the news popularity module for processing to obtain the news. Popularity results ; thereby gaining users Historical Click News Sequence Popularity results of all news and candidate news sequence Candidate news in Popularity results ; Step 3.2.3, for the user Historical Click News Sequence Observation and statistics were conducted to obtain user data. Historical Click News Topic Distribution Sub-topic distribution ; Step 3.2.4: In the user's historical interaction news dataset The study observes and statistically analyzes news topics and subtopics to obtain the topic distribution of historical news data. Sub-topic distribution ; Step 3.2.5: The coarse-grained topic unit is based on the user... Historical Click News Topic Distribution Thematic distribution of historical news data Calculate the KL divergence to obtain topic-based user data. Conformity trait results ; Step 3.2.6: The fine-grained sub-topic unit is based on the user... Distribution of historical click news by subcategories Subclass distribution of news historical data Calculate the KL divergence to obtain user data based on subtopics. Conformity trait results ; Step 3.2.7: The topic-based user... Conformity trait results and subclass-based users Conformity trait results After splicing, it can be used as a user Preliminary results of conformity traits .

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the method of any one of claims 1-3, the processor being configured to execute the program stored in the memory.

5. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of any of the methods described in claims 1-3.

Images