User hierarchical interest-driven diversified news recommendation method and system

Abstract

The invention discloses a user hierarchical interest-driven diversified news recommendation method and system. The method comprises the following steps: preprocessing a news data set; learning content representation from the context of a single news by using a multi-head self-attention network, and modeling a relationship between news by using a cross attention network; converting the popularity of the historical click news into a popularity embedding vector, and dynamically adjusting the importance of the historical click news content embedding representation in combination with the historical click news content embedding representation and the click rate; extracting user interest representations from the sub-category level, the category level and the global level, and calculating user hierarchical interest scores; calculating popularity scores of the candidate news; and adaptively aggregating the hierarchical interest score of the user and the popularity score of the candidate news through a dynamic gating mechanism to obtain a recommendation score of the candidate news, and performing diversified news recommendation based on the recommendation score. According to the invention, while personalized accurate recommendation is maintained, the recommendation diversity is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of news recommendation, and in particular to a diversified news recommendation method and system driven by user hierarchical interests. Background Art

[0002] In the internet age, online news platforms (such as Microsoft News, Google News, and Flipboard) have attracted hundreds of millions of users worldwide. With their convenience and timeliness, digital news services are gradually replacing traditional newspapers and becoming the primary means for users to obtain news information. However, with the massive amount of news generated and published globally every day, it is nearly impossible for users to browse all available news to filter out content of interest. This information overload not only increases users' cognitive burden but also impairs their reading experience. To address this issue, personalized news recommendation technology has emerged. By providing users with customized news content, news recommendation systems can effectively alleviate information overload and significantly enhance their reading experience. However, as users' demands for news content quality and diversity continue to increase, news recommendation systems face greater challenges. While existing methods incorporate auxiliary information such as popularity, user characteristics, or news attributes into user and news modeling, they often fail to adaptively aggregate users' hierarchical interests with the popularity of candidate news. Therefore, to achieve more diverse and accurate news recommendations, it is necessary to be driven by users' hierarchical interests and adaptively aggregate the popularity information of candidate news to improve the accuracy and diversity of recommendations, thereby better meeting users' personalized reading needs. Summary of the Invention

[0003] Aiming at the problem that directly applying clustering algorithms to large-scale data sets in the prior art may lead to high computational complexity and long time consumption, the present invention proposes a diversified news recommendation method and system driven by user hierarchical interests.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] On one hand, the present invention proposes a diversified news recommendation method driven by user hierarchical interests, comprising:

[0006] Step 1: Preprocess the news dataset to obtain the corresponding text vectors and entity vectors;

[0007] Step 2: The text vector and entity vector are used as input to the news content encoder. A multi-head self-attention network is used to learn content representation from the context of a single news item. A cross-attention network is used to model the relationship between news items to obtain an embedding representation of historical click-through news content.

[0008] Step 3: Unify and quantify the popularity of the i-th clicked news by the timeliness-aware news popularity predictor, and convert it into a popularity embedding vector through popularity embedding. Combined with the historical click news content embedding representation and click-through rate, the importance of the historical click news content embedding representation is dynamically adjusted to obtain the final historical click news content embedding representation.

[0009] Step 4: Based on the final historical click news content embedding representation and the user's news browsing behavior, a hierarchical interest modeling framework is constructed to extract user interest representations at the subcategory level, category level, and global level, and calculate the user's hierarchical interest score;

[0010] Step 5: Based on the timeliness, content characteristics, and click-through rate of the news, a timeliness-aware news popularity predictor is used to calculate the popularity score of the candidate news.

[0011] Step 6: Adaptively aggregate the user's hierarchical interest scores and the popularity scores of candidate news through a dynamic gating mechanism to obtain the recommendation scores of the candidate news, and perform diversified news recommendations based on the recommendation scores.

[0012] Furthermore, before step 3, the method further includes:

[0013] A timeliness-aware news popularity predictor is used to calculate the popularity scores of historical click news.

[0014] Furthermore, the step 2 includes:

[0015] Step 2.1: Use a multi-head self-attention network to learn word representations from the context of text and entities respectively and entity representation

[0016] Step 2.2: Use a multi-head cross attention network to interactively model word representation and entity representation to obtain word representation and entity representation

[0017] Step 2.3: Get a unified representation of the word based on the word representation and entity representation obtained in steps 2.1 and 2.2 and a unified representation of entities

[0018] Step 2.4: Combine the unified representation of words and entities and use the attention network to learn the news title representation and news summary representation based on words and entities respectively;

[0019] Step 2.5: Combine the unified representation of words and entities and use the attention network to learn a unified representation of news headlines by aggregating the word-based and entity-based news headline representations;

[0020] Step 2.6: Utilize the attention network to learn a unified representation of news summaries by aggregating word-based and entity-based news summary representations;

[0021] Step 2.7: Obtain the historical click news content embedding representation through the unified representation of news title and news summary.

[0022] Furthermore, in step 2.4, word-based news headline representation is learned in the following manner:

[0023]

[0024] Where,

[0025]

[0026] where n w is a word-based news headline representation, is the attention weight of the i-th word, V w and b w is a trainable parameter, q w represents the attention query vector and m represents the number of words.

[0027] Furthermore, in step 3, the importance of the embedded representation of the historical click news content is dynamically adjusted in the following manner:

[0028]

[0029] where α i and n i denote the attention weight and news content embedding representation of the i-th click news, respectively, p i represents the popularity embedding vector of the i-th clicked news, q and W u are trainable parameters used to generate the attention query vector and the projection of the learned representation, respectively.

[0030] Furthermore, in step 4, the user hierarchical interest score is calculated in the following manner:

[0031] f s =u s ·n c

[0032] f t =u t ·n c

[0033] f g =u g ·n c

[0034] where f s 、f t 、f g Represent the user interest scores at the subcategory level, category level, and global level, u s 、u t 、u g Represents the candidate perceived user interests at the subcategory level, category level, and global level, n c Represents candidate news content embedding representations.

[0035] Furthermore, the u s 、u t 、u g Calculated as follows:

[0036] u=y u +p

[0037] In the formula

[0038]

[0039] Where p represents the context embedding of the corresponding level, which can be subcategory level, category level or global level, and u is u s 、u t or u g , α k represents the normalized weight used to weight the embedding of click news, n u,k represents the embedding of the k-th clicked news by user u, a u,k represents the attention score of the k-th clicked news, y u represents the candidate-aware user interest embedding at the corresponding level, and L represents the number of news clicked by the user.

[0040] Furthermore, the step 5 includes:

[0041] Step 5.1: Use the time embedding layer to convert the timeliness of the news into a timeliness embedding vector t. The timeliness of the news is the length of time between the news release time and the current prediction time.

[0042] Step 5.2: Use the Dense network to process t and predict the time score t′;

[0043] Step 5.3: Calculate the inverse of t′ and adjust its intensity through the parameter μ to obtain the popularity of news timeliness p t ;

[0044] Step 5.4: Embed candidate news content n through Dense network c Processing is performed to generate content-based news popularity score p c ;

[0045] Step 5.5: Use aggregators to integrate p t and p c Get the comprehensive news popularity p based on time-aware content:

[0046] p=θ·p c +(1-θ)·p t

[0047] Among them, θ is a dynamically adjusted weight used to balance p c and p t the impact of;

[0048] Step 5.6: Use the user interaction data in the last t hours to calculate the near-real-time click-through rate c t ;

[0049] Step 5.7: By calculating p and c t The popularity score s of the candidate news is obtained by taking the weighted sum of p .

[0050] Furthermore, in step 5.6, c is calculated as follows: t :

[0051]

[0052] Clicks i,t is the number of clicks on news i in time window t, Views i,t is the number of exposures of news i in time window t, ε is a positive value to avoid the denominator being zero, ω t is the time weight, T is the total number of time windows, U i Indicates the number of independent user clicks on news i.

[0053] Furthermore, in step 6, the recommendation score of the candidate news is obtained in the following manner:

[0054] s=σ(α1f s +α2f t +α3f g +α4s p )

[0055] Where,

[0056]

[0057] a i =V T tanh(W s l i +W h n c +b)

[0058]

[0059] Where s is the recommendation score of the candidate news, f s 、f t 、f g They represent the user interest scores at the subcategory level, category level, and global level, respectively. α1, α2, and α3 represent f s 、f t 、f g The weight of α4 is the news popularity weight, s p is the popularity score of the candidate news, σ is the Sigmoid function, γ is the control coefficient of popularity, l i represents the user's i-th level interest, i∈{1,2,3} represents the subcategory level, category level and global level respectively, n c represents the candidate news content embedding representation, W s 、W h and b are trainable parameters, and V is the attention weight vector.

[0060] Another aspect of the present invention provides a user-level interest-driven diversified news recommendation system, comprising:

[0061] The data preprocessing module is used to preprocess the news dataset to obtain the corresponding text vectors and entity vectors;

[0062] A news content encoding module is configured to use the text vector and entity vector as input to a news content encoder, utilize a multi-head self-attention network to learn content representation from the context of a single news item, utilize a cross-attention network to model the relationship between news items, and obtain an embedding representation of historical click news content;

[0063] The popularity bias mitigation module is used to unify and quantify the popularity of the i-th clicked news predicted by the timeliness-aware news popularity predictor. This module converts the predicted popularity into a popularity embedding vector through popularity embedding. The module then dynamically adjusts the importance of the historical clicked news content embedding representation by combining the historical clicked news content embedding representation and click-through rate to obtain the final historical clicked news content embedding representation.

[0064] The user hierarchical interest learning module is used to build a hierarchical interest modeling framework based on the final historical click news content embedding representation and the user's news browsing behavior. It extracts user interest representations at the subcategory level, category level, and global level, and calculates the user hierarchical interest score.

[0065] The candidate news popularity prediction module is used to calculate the popularity score of candidate news by integrating the timeliness, content characteristics and click-through rate of the news and using a timeliness-aware news popularity predictor;

[0066] The multi-dimensional attribute score adaptive aggregation module is used to adaptively aggregate user hierarchical interest scores and candidate news popularity scores through a dynamic gating mechanism to obtain the recommendation scores of candidate news and perform diversified news recommendations based on the recommendation scores.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] This paper proposes a diversified news recommendation method and system driven by hierarchical user interests, aiming to improve the accuracy and diversity of recommendations. The present invention introduces a popularity bias mitigation process (i.e., step 3) during the news encoding phase to effectively distinguish between the influence of news popularity and users' actual interests, thereby reducing the interference of popular news on user interest modeling and highlighting news content that better meets users' personalized needs. Furthermore, the present invention constructs a user interest tree through a candidate-aware subcategory-category-global hierarchical user interest learning method, accurately capturing users' interest preferences at different levels. Simultaneously, the present invention dynamically integrates the multidimensional attributes of candidate news (such as news timeliness, content characteristics, and click-through rate) to calculate a comprehensive popularity score for each candidate news, ensuring that recommendations balance personalization and news trends. To improve recommendation diversity while maintaining accurate recommendations and mitigate the cold start problem, the present invention proposes a dynamic gating mechanism. This network dynamically adjusts the weights of different levels of interest based on the distribution of user interests and, combined with the popularity information of candidate news, achieves adaptive optimization, ensuring that recommendation results are more aligned with user preferences while improving the breadth of news content coverage and fairness of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a basic flow chart of a user-level interest-driven diversified news recommendation method according to an embodiment of the present invention;

[0070] Figure 2 A schematic structural diagram of a news content encoder provided by an embodiment of the present invention;

[0071] Figure 3 This is a schematic diagram of the architecture of a diversified news recommendation system driven by user hierarchical interests according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:

[0073] like Figure 1 As shown in FIG, a diversified news recommendation method driven by user hierarchical interests includes:

[0074] S101: Preprocess the news dataset to obtain corresponding text vectors and entity vectors;

[0075] S102: Using the text vector and entity vector as input to a news content encoder, using a multi-head self-attention network to learn content representation from the context of a single news item, and using a cross-attention network to model the relationship between news items, to obtain an embedding representation of historical click news content;

[0076] S103: The popularity of the i-th clicked news is predicted by the unified quantitative timeliness-aware news popularity predictor and converted into a popularity embedding vector through popularity embedding. The importance of the historical clicked news content embedding is dynamically adjusted based on the historical clicked news content embedding representation and click-through rate to obtain the final historical clicked news content embedding representation.

[0077] S104: Based on the final historical click news content embedding representation and the user's news browsing behavior, a hierarchical interest modeling framework is constructed to extract user interest representations at the subcategory level, category level, and global level, and calculate the user's hierarchical interest score;

[0078] S105: Calculate the popularity score of the candidate news using a timeliness-aware news popularity predictor based on the news' timeliness, news content characteristics, and click-through rate;

[0079] S106: Adaptively aggregate the user's hierarchical interest scores and the popularity scores of the candidate news through a dynamic gating mechanism to obtain the recommendation scores of the candidate news, and perform diversified news recommendations based on the recommendation scores.

[0080] As an implementable embodiment, the method specifically includes:

[0081] (1) Preprocessing the news dataset

[0082] News articles usually contain rich text and entity information, which can capture the distinctive features of news. To this end, the news dataset is preprocessed, and 50k user logs and 200k user logs are randomly extracted from the MIND-large dataset to construct the MIND-small dataset and the MIND-mid dataset. The features such as news text content, entity information, and click-through rate are vectorized, and the pre-trained word embedding GloVe is used to initialize the text and entity embedding. Specifically, the text sequence and entity sequence are extracted from the news title and summary respectively as input, which are recorded as the text sequence T of the title part. t and entity sequence E t , and the text sequence T of the abstract part a With entity sequence E a In the input sequence, each word w and entity e is projected into a corresponding vector E through the embedding layer. w and E e .

[0083] (2) Building a news content encoder

[0084] News articles usually contain rich text and entity information, which can capture the distinctive features of news. To this end, a news content encoder is designed to efficiently learn the representation of news, such as Figure 2 As shown. Representation learning is performed on the user's historical click news. When learning the user's contextual news representation, not only the Multi-Head Self-Attention Network (MHSA) is used to capture the contextual features of a single news article, but also the Multi-Head Cross-Attention Network (MHCA) is introduced to capture the interactive relationship between different news articles. This combination can more comprehensively model the semantics and relevance of user click news. A multi-head self-attention network is used to learn the representation of the user's historical click news. The network can capture the contextual relevance in the news content and generate refined contextual news representation. The specific process is as follows:

[0085] S102.1: Use a multi-head self-attention network to learn word representation and entity representation from the context of text and entity respectively. The calculation formula is:

[0086]

[0087] S102.2: Model the interaction between word representation and entity representation through a multi-head cross attention network using the following formula

[0088]

[0089] Where MultiCrossAtt represents a multi-head cross attention network, and m and n represent the number of words and entities, respectively.

[0090] S102.3: Formulate a unified representation for each word and entity as the sum of its representations, i.e. and

[0091] S102.4: Use the attention network to learn word-based and entity-based news representations respectively, and express the attention weight of the i-th word as The calculation formula of attention weight is:

[0092]

[0093] Where V w and b w is a trainable parameter, q w represents the attention query vector.

[0094] Based on word n w The title representation of is the sum of word representations weighted by attention weights, i.e.

[0095]

[0096] Based on entity n e The title representation can be calculated in a similar way.

[0097] S102.5: Using Attention Networks to Learn a Unified Representation of News Headlines by Aggregating Word-Based and Entity-Based Representations t ;

[0098] S102.6: Learning a Unified Representation for News Summarization by Aggregating Word- and Entity-Based Representations with Attention Networks a ;

[0099] S102.7: The final representation n of the news content learned by the news content encoder is the sum of the representations of its title and summary, i.e. n = s t +s a .

[0100] (3) Popularity bias mitigation process

[0101] The popularity of the news clicked on the i-th time predicted by the unified quantitative news popularity predictor is converted into an embedding vector p through popularity embedding i , combining news context representation and click-through rate information to alleviate popularity bias and highlight important news. And select important click news for user interest modeling, which is expressed as:

[0102]

[0103] where α i and n i denotes the attention weight of the i-th clicked news and the contextual news content embedding representation, n i Generated by news content encoder. i represents the popularity embedding vector of the i-th clicked news, which is used to capture the click rate characteristics. q and W u are trainable parameters used to generate the attention query vector and the projection of the learned representation, respectively.

[0104] Finally, the historical click news content is embedded as

[0105] This design not only focuses on the contextual semantic features of news, but also dynamically adjusts the importance of news through popularity embedding, ensuring that more informative news has a greater weight in user interest modeling, while effectively alleviating the model bias problem caused by popularity deviation.

[0106] (4) User-level interest learning

[0107] In order to measure the relevance between user clicked news and candidate news, we first calculate the similarity matrix between the two through the embedding matrix. Specifically, for user u’s clicked news embedding matrix and candidate news embedding matrix (obtained by repeatedly expanding the candidate news embeddings), the correlation matrix C is calculated as:

[0108] C=N u T W c N c (9)

[0109] in, is a trainable weight parameter used to capture the correlation between different feature dimensions. Here, d n is the dimension of news embedding, L represents the number of news clicked by users. The similarity matrix C∈R L×L It indicates the matching degree between each news clicked by the user and the candidate news, providing a basis for subsequent attention modeling.

[0110] Since users click on a variety of news, only some of which are relevant to candidate news, it is necessary to assign different weights to the clicked news through an attention mechanism to highlight the important information related to the candidate news. We introduce a candidate-aware attention network to calculate the importance weight of the news clicked by the user. First, by fusing the embedding information of the clicked news and the candidate news and the similarity matrix C, we calculate the attention vector a u ∈T L :

[0111]

[0112] in, is the attention query vector, which is used to measure the matching importance of each clicked news; is a trainable weight matrix; is a softmax activation function that maps the values of the correlation matrix to a probability distribution. Subsequently, the attention score is converted to a normalized weight α through softmax. k , for embedding weighted clickthrough news:

[0113]

[0114] Among them, a u,k Denotes the attention score of user u’s k-th clicked news. Based on the attention weight, the embeddings of the clicked news are weighted and summed to generate the candidate perceived user interest embedding y at this level. u :

[0115]

[0116] Among them, n u,k Represents the embedding of the k-th news clicked by user u.

[0117] At each level of interest, in order to more comprehensively model the user's interest distribution, we also combine the contextual embedding of that level (such as subcategory information, category information, or global information) to further optimize the interest representation. The formula for multi-granularity interest fusion is:

[0118] u=y u +p (13)

[0119] Among them, p represents the context embedding of the corresponding level (such as subcategory, category or global information), which contains the prior knowledge of the specific level; u is the final user interest representation at that level.

[0120] Use dot product to calculate the user interest score f at the subcategory level, category level, and global level respectively s 、f t 、f g , and the user hierarchical interest scores are shown in Table 1.

[0121] Table 1. User hierarchical interest scores

[0122]

[0123] Specifically, the subcategory-level score is given by f s =u s ·n c Calculate, where u s is the subcategory-level candidate perceived user interest, n c is the candidate news content embedding. The category-level score is given by f t =u t ·n c Calculate, where u t is the category-level candidate perceived user interest. The global level score is given by f g =u g ·n c Calculation of u g Is the sub-global candidate perceived user interest. If the subcategory (category) of the candidate news does not belong to any subcategory (category) clicked by the user, directly f s (ft ) is set to 0. In this case, the final interest score is based only on f s (f t ) and f g This hierarchical candidate-aware user encoding framework can effectively capture the hierarchical interest distribution of users, fully mine the users' click behavior information, and provide more accurate matching support for recommending candidate news.

[0124] (5) Constructing a timeliness-aware news popularity predictor

[0125] The prediction of candidate news popularity includes the following steps:

[0126] S105.1: Use a temporal embedding layer to map the input into a vector space, enhancing the model's ability to capture valuable information and features. Convert t into a time-sensitive embedding vector t. The timeliness of a particular news item t is defined as the time (in hours) between its release time and the current prediction time.

[0127] S105.2: Use the Dense network to process the predicted time score t′, that is

[0128] t′=Dense(t) (14)

[0129] S105.3: Calculate the inverse of t′ and adjust its intensity through the parameter μ to obtain the popularity of news timeliness p t

[0130] p t =(1 / t′) μ (15)

[0131] S105.4: Embedding candidate news content through Dense network c Processing is performed to generate content-based news popularity score p c , the formula is:

[0132] p c =Dense(n c ) (16)

[0133] S105.5: Use content-specific aggregators to aggregate news based on timeliness, popularity, and t and content-based news popularity score p c The comprehensive news popularity p based on time-aware content is obtained. The specific calculation process is:

[0134] p=θ·p c +(1-θ)·p t (17)

[0135] Among them, θ is a dynamically adjusted weight used to balance the influence of content popularity and timeliness popularity. The weight θ is calculated as follows:

[0136] θ=σ(W p ·[n,t]+b p ) (18)

[0137] Among them, σ is the Sigmoid activation function, which limits the weight value to the range of [0,1]; W p and b p where [n, t] represents the concatenation of the news content embedding vector n and the real-time news embedding vector t. This approach provides an effective solution for time-aware news popularity prediction and can address the diverse lifecycle requirements of different news types.

[0138] S105.6: Use the user interaction data in the last t hours to calculate the near-real-time click-through rate, denoted as c t ;

[0139] To quantify the popularity of news, the click-through rate c t It is a key indicator for measuring the popularity of news. The calculation of click-through rate is further expanded:

[0140]

[0141] Among them, Clicks i Views is the number of clicks on news i within the specified time window. i is the number of exposures of news i in the same time window, and ε is a small positive value to avoid the denominator being zero. In order to capture the dynamic changes in click-through rate, its trend in different time periods is included in the calculation, using a weighted moving average method:

[0142]

[0143] Among them, ω t is the time weight, which is usually set to give a greater weight to click rates close to the current time. In addition, to reflect the user diversity of click behavior, user category preference adjustment is also introduced:

[0144]

[0145] Among them, U i It represents the number of independent users clicking on news i, and is used to measure the diversity of click behaviors.

[0146] S105.7: In the news recommendation system, accurately predicting the popularity of news is crucial to improving the recommendation effect. However, the popularity of news is affected by many factors, including the real-time click behavior of users and the content characteristics of the news itself. For popular news that has accumulated a large amount of interactive data, the click-through rate can directly reflect its popularity; while for news that has been newly released but has not yet received enough clicks, its content characteristics may better reflect its future popularity potential. Therefore, a dynamic fusion mechanism for the multi-dimensional attributes of candidate news is proposed, which performs a weighted fusion of the popularity based on click-through rate and the popularity based on content to comprehensively measure the timeliness popularity of news. The specific calculation method is as follows: For popular news (such as real-time hot spots), the click-through rate c t For newly released news (insufficient interactive data), content popularity p is more critical. Combining the two helps the model capture both the current popularity of the news (click-through rate) and its potential popularity trend (content characteristics). Finally, the news’s timeliness popularity score s p Expressed as p and c t The weighted sum of is as follows:

[0147] s p =τ p p+τ c c t (twenty two)

[0148] where τ p and τ c are trainable parameters, which are used to control the weights of CTR and content popularity in the final popularity. c >τ p ,, which means the model believes that click-through rate contributes more to news popularity (for example, user behavior data is relatively abundant). On the contrary, it means that the model believes that content popularity is dominant (for example, newly released news has not yet accumulated enough interaction data).

[0149] The prediction of historical click news popularity is similar to the above process.

[0150] (6) Adaptive aggregation of multidimensional attribute scores

[0151] In news recommendation tasks, user interests typically have a multi-layered structure, including fine-grained interests (subcategory level), medium-grained interests (category level), and global interests. Furthermore, news popularity characteristics (such as click-through rate and content popularity) are also crucial for recommendation quality. This is especially true in cold-start scenarios, where personalized user information is insufficient and more reliance is placed on the overall popularity of news. How to effectively integrate user hierarchical interests with news popularity information to improve recommendation effectiveness is an urgent problem. To this end, we propose an adaptive aggregation network that uses a dynamic gating mechanism to learn the optimal weighting of different interest levels and popularity information, achieving adaptive fusion.

[0152] In our model, users' hierarchical interests can be divided into subcategory level, category level and global level, corresponding to the modeling of users' preferences for specific news topics, larger categories and overall interests, respectively.

[0153] The dynamic gating mechanism is used to learn the optimal weighting method of interest and popularity information at different levels, and the user's interests at different levels are represented as l1, l2 and l3 respectively, and the candidate news content is embedded as n c The calculation formula is as follows:

[0154] a i =V T tanh(W s l i +W h n c +b) (23)

[0155] Among them, l i represents the user's i-th level interest (subcategory level, category level and global level); n c represents the candidate news content vector representation; W s , W h , b is a trainable parameter, and V is the attention weight vector. Softmax normalization is used to calculate the final interest weight:

[0156]

[0157] This mechanism ensures that the interest weights at different levels can be adaptively adjusted according to the distribution of user interests.

[0158] To simultaneously consider user personalized interests and news popularity, we designed a dynamic gating mechanism that adaptively adjusts the weights of different interest levels and popularity information based on user characteristics. For cold-start users, due to insufficient historical behavior data, directly calculated α1, α2, and α3 may be unstable. Therefore, we introduce a news popularity weight α4 and use a dynamic coefficient γ for fusion:

[0159]

[0160] Here, γ controls the influence of popularity. For cold-start users, recommendation results rely on news popularity; conversely, recommendation results rely primarily on the user's hierarchical interests. This formula ensures that the sum of the weights always equals 1 and can adaptively adjust the proportion of user hierarchical interests and popularity.

[0161] Finally, the recommendation score of the candidate news is calculated as follows:

[0162] s=σ(α1f s +α2f t +α3f g +α4s p ) (27)

[0163] Where σ is a Sigmoid function, ensuring that the recommendation score is normalized to a reasonable range. This dynamic gating mechanism integrates users' hierarchical interests (category-level, subcategory-level, and global-level) with news popularity information and adaptively adjusts the weights of different dimensions. This improves the diversity of recommendations while maintaining personalized and accurate recommendations, resulting in more balanced recommendations. Furthermore, this mechanism effectively alleviates data sparsity in cold start scenarios, improving the overall user experience. The recommendation score not only reflects the content relevance of the news but also fully considers user preferences and news trends, achieving more multidimensional, accurate, and personalized news recommendations. For example, Table 2 shows the personalized recommendation scores of news N45 by different users, as well as the news' scores based on the subcategory-level, category-level, global-level, and candidate news popularity attributes.

[0164] Table 2. Recommendation scores and multi-dimensional attribute scores of News N45

[0165]

[0166] (7) Conduct method model training

[0167] The model is trained using the dataset and negative sampling technology. The loss function is the negative log-likelihood of all positive samples. The formula is as follows:

[0168]

[0169] Among them, S is the set of all positive samples, is the click probability score, are the click probability scores of the K negative samples respectively. The final diversified news recommendation results are shown in Table 3.

[0170] Table 3. Diversified news recommendation results

[0171]

[0172] Based on the above embodiments, Figure 3 As shown, the present invention also proposes a user-level interest-driven diversified news recommendation system, including:

[0173] The data preprocessing module is used to preprocess the news dataset to obtain the corresponding text vectors and entity vectors;

[0174] A news content encoding module is configured to use the text vector and entity vector as input to a news content encoder, utilize a multi-head self-attention network to learn content representation from the context of a single news item, utilize a cross-attention network to model the relationship between news items, and obtain an embedding representation of historical click news content;

[0175] The popularity bias mitigation module is used to unify and quantify the popularity of the i-th clicked news predicted by the timeliness-aware news popularity predictor. This module converts the predicted popularity into a popularity embedding vector through popularity embedding. The module then dynamically adjusts the importance of the historical clicked news content embedding representation by combining the historical clicked news content embedding representation and click-through rate to obtain the final historical clicked news content embedding representation.

[0176] The user hierarchical interest learning module is used to build a hierarchical interest modeling framework based on the final historical click news content embedding representation and the user's news browsing behavior. It extracts user interest representations at the subcategory level, category level, and global level, and calculates the user hierarchical interest score.

[0177] The candidate news popularity prediction module is used to calculate the popularity score of candidate news by integrating the timeliness, content characteristics and click-through rate of the news and using a timeliness-aware news popularity predictor;

[0178] The multi-dimensional attribute score adaptive aggregation module is used to adaptively aggregate user hierarchical interest scores and candidate news popularity scores through a dynamic gating mechanism to obtain the recommendation scores of candidate news and perform diversified news recommendations based on the recommendation scores.

[0179] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A diversified news recommendation method driven by user hierarchical interests, characterized by: include: Step 1: Preprocess the news dataset to obtain the corresponding text vectors and entity vectors; Step 2: The text vector and entity vector are used as input to the news content encoder. A multi-head self-attention network is used to learn content representation from the context of a single news item. A cross-attention network is used to model the relationship between news items to obtain an embedding representation of historical click-through news content. Step 3: Unify and quantify the popularity of the i-th clicked news by the timeliness-aware news popularity predictor, and convert it into a popularity embedding vector through popularity embedding. Combined with the historical click news content embedding representation and click-through rate, the importance of the historical click news content embedding representation is dynamically adjusted to obtain the final historical click news content embedding representation. Step 4: Based on the final historical click news content embedding representation and the user's news browsing behavior, a hierarchical interest modeling framework is constructed to extract user interest representations at the subcategory level, category level, and global level, and calculate the user's hierarchical interest score; Step 5: Based on the timeliness, content characteristics, and click-through rate of the news, a timeliness-aware news popularity predictor is used to calculate the popularity score of the candidate news. Step 6: Adaptively aggregate the user's hierarchical interest scores and the popularity scores of candidate news through a dynamic gating mechanism to obtain the recommendation scores of the candidate news, and perform diversified news recommendations based on the recommendation scores.

2. A user-level interest-driven diversified news recommendation method according to claim 1, characterized in that: The step 2 includes: Step 2.1: Use a multi-head self-attention network to learn word representations from the context of text and entities respectively and entity representation Step 2.2: Use a multi-head cross attention network to interactively model word representation and entity representation to obtain word representation and entity representation Step 2.3: Get a unified representation of the word based on the word representation and entity representation obtained in steps 2.1 and 2.2 and a unified representation of entities Step 2.4: Combine the unified representation of words and entities and use the attention network to learn the news title representation and news summary representation based on words and entities respectively; Step 2.5: Combine the unified representation of words and entities and use the attention network to learn a unified representation of news headlines by aggregating the word-based and entity-based news headline representations; Step 2.6: Utilize the attention network to learn a unified representation of news summaries by aggregating word-based and entity-based news summary representations; Step 2.7: Obtain the historical click news content embedding representation through the unified representation of news title and news summary.

3. A user-level interest-driven diversified news recommendation method according to claim 2, characterized in that: In step 2.4, the word-based news headline representation is learned as follows: Where, where n w is a word-based news headline representation, is the attention weight of the i-th word, V w and b w is a trainable parameter, q w represents the attention query vector and m represents the number of words.

4. The user-level interest-driven diversified news recommendation method according to claim 1, characterized in that: In step 3, the importance of the embedded representation of the historical click news content is dynamically adjusted in the following manner: where α i and n i denote the attention weight and news content embedding representation of the i-th click news, respectively, p i represents the popularity embedding vector of the i-th clicked news, q and W u are trainable parameters used to generate the attention query vector and the projection of the learned representation, respectively.

5. The user-level interest-driven diversified news recommendation method according to claim 1, characterized in that: In step 4, the user hierarchical interest score is calculated as follows: f s =u s ·n c f t =u t ·n c f g =u g ·n c where f s 、f t 、f g Represent the user interest scores at the subcategory level, category level, and global level, u s 、u t 、u g Represents the candidate perceived user interests at the subcategory level, category level, and global level, n c Represents candidate news content embedding representations.

6. A user-level interest-driven diversified news recommendation method according to claim 5, characterized in that: The u s 、u t 、u g It is calculated as follows: u=y u +p In the formula Where p represents the context embedding of the corresponding level, which can be subcategory level, category level or global level, and u is u s 、u t or u g , α k represents the normalized weight used to weight the embedding of click news, n u,k represents the embedding of the k-th clicked news by user u, a u,k represents the attention score of the k-th clicked news, y u represents the candidate-aware user interest embedding at the corresponding level, and L represents the number of news clicked by the user.

7. The user-level interest-driven diversified news recommendation method according to claim 1, characterized in that: The step 5 comprises: Step 5.1: Use the time embedding layer to convert the timeliness of the news into a timeliness embedding vector t. The timeliness of the news is the length of time between the news release time and the current prediction time. Step 5.2: Use the Dense network to process t and predict the time score t′; Step 5.3: Calculate the inverse of t′ and adjust its intensity through the parameter μ to obtain the popularity of news timeliness p t ; Step 5.4: Embed candidate news content n through Dense network c Processing is performed to generate content-based news popularity score p c ; Step 5.5: Use aggregators to integrate p t and p c Get the comprehensive news popularity p based on time-aware content: p=θ·p c +(1-θ)·p t Among them, θ is a dynamically adjusted weight used to balance p c and p t the impact of; Step 5.6: Use the user interaction data in the last t hours to calculate the near-real-time click-through rate c t ; Step 5.7: By calculating p and c t The popularity score s of the candidate news is obtained by taking the weighted sum of p .

8. The user-level interest-driven diversified news recommendation method according to claim 7, characterized in that: In step 5.6, c is calculated as follows: t : Clicks i,t is the number of clicks on news i in time window t, Views i,t is the number of exposures of news i in time window t, ε is a positive value to avoid the denominator being zero, ω t is the time weight, T is the total number of time windows, U i Indicates the number of independent user clicks on news i.

9. The user-level interest-driven diversified news recommendation method according to claim 1, characterized in that: In step 6, the recommendation score of the candidate news is obtained in the following manner: s=σ(α1f s +α2f t +α3f g +α4s p ) Where, a i =V T tanh(W s l i +W h n c +b) Where s is the recommendation score of the candidate news, f s 、f t 、f g They represent the user interest scores at the subcategory level, category level, and global level, respectively. α1, α2, and α3 represent f s 、f t 、f g The weight of α4 is the news popularity weight, s p is the popularity score of the candidate news, σ is the Sigmoid function, γ is the control coefficient of popularity, l i represents the user's i-th level interest, i∈{1,2,3} represents the subcategory level, category level and global level respectively, n c represents the candidate news content embedding representation, W s 、W h and b are trainable parameters, and V is the attention weight vector.

10. A user-level interest-driven diversified news recommendation system, characterized by: include: The data preprocessing module is used to preprocess the news dataset to obtain the corresponding text vectors and entity vectors; A news content encoding module is configured to use the text vector and entity vector as input to a news content encoder, utilize a multi-head self-attention network to learn content representation from the context of a single news item, utilize a cross-attention network to model the relationship between news items, and obtain an embedding representation of historical click news content; The popularity bias mitigation module is used to unify and quantify the popularity of the i-th clicked news predicted by the timeliness-aware news popularity predictor. This module converts the predicted popularity into a popularity embedding vector through popularity embedding. The module then dynamically adjusts the importance of the historical clicked news content embedding representation by combining the historical clicked news content embedding representation and click-through rate to obtain the final historical clicked news content embedding representation. The user hierarchical interest learning module is used to build a hierarchical interest modeling framework based on the final historical click news content embedding representation and the user's news browsing behavior. It extracts user interest representations at the subcategory level, category level, and global level, and calculates the user hierarchical interest score. The candidate news popularity prediction module is used to calculate the popularity score of candidate news by integrating the timeliness, content characteristics and click-through rate of the news and using a timeliness-aware news popularity predictor; The multi-dimensional attribute score adaptive aggregation module is used to adaptively aggregate user hierarchical interest scores and candidate news popularity scores through a dynamic gating mechanism to obtain the recommendation scores of candidate news and perform diversified news recommendations based on the recommendation scores.

Images