News pushing method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of news recommendation, and discloses a news pushing method and device, electronic equipment and a storage medium. The method comprises the steps of collecting user historical interaction data and news multivariate features, and constructing a multi-source fusion data set; grouping the users by adopting a community discovery algorithm based on density clustering to form a user community, and extracting a group interest mode; calculating the association strength of the news and the user community; combining the news timeliness and the social influence coefficient of the user community, and dynamically weighting the association strength to generate an initial push weight; adjusting the initial push weight based on network propagation by using a topological relation network between user communities to obtain an adjusted weight; and pushing the news according to the descending order of the corresponding news pushing weights. According to the method, the user communities are accurately divided by utilizing density clustering, and the pushing weight is adjusted through the network propagation model, so that the news distribution accuracy and foresight are improved.

Description

Technical Field

[0001] This invention relates to the field of news recommendation technology, specifically to news push methods and devices, electronic devices, and storage media. Background Technology

[0002] Existing personalized news recommendation systems typically rely on collaborative filtering or content-based recommendation algorithms. These methods primarily focus on the direct matching between an individual user's historical behavior and news content, or the aggregation of behaviors from similar user groups. For user segmentation, distance-based partitioning clustering algorithms are commonly used. These methods assume a spherical distribution of users, require pre-setting the number of groups, and struggle to effectively identify the complex, irregular, and intertwined user interest communities in the real world. As a result, user segmentation patterns are rigid, failing to accurately reflect the true structure of interest circles, and poorly handling users with unique interests or sparse behaviors, thus affecting the accuracy of extracting group interest patterns.

[0003] Current recommendation mechanisms mostly treat identified user groups as independent individuals when recommending content, lacking consideration for the dynamic relationships between groups. The popularity or relevance of news items within one group cannot effectively spread to other groups with potential interests. This isolated recommendation strategy cannot simulate the ripple effect of information dissemination in social networks, making it difficult for the system to capture emerging trends across communities, resulting in lagging and unforeseen recommendation results. The weighting calculation process for news pushes is relatively static and fails to incorporate the network topology that reflects inter-community influence, limiting the dynamic adaptability of the recommendation system and its ability to discover users' potential interests. Summary of the Invention

[0004] The purpose of this invention is to provide a news push method and apparatus, electronic device, and storage medium to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a news push method, the method comprising: Collect users' historical interaction data and diverse feature data from news articles to construct a multi-source fusion dataset; A density-based clustering-based community detection algorithm is used to group users into multiple user communities, and the group interest patterns of each user community are extracted. Calculate the correlation strength between news and user communities based on group interest patterns and diverse characteristics of news data; By combining the timeliness parameter of the news and the social influence coefficient of the user community, the correlation strength is dynamically weighted to generate the initial news push weight; By leveraging the topological relationship network between user communities, the initial news push weight is adjusted based on network propagation to obtain the adjusted news push weight; News is pushed out in descending order of its corresponding news push weight.

[0006] Preferably, the step of grouping users into multiple user communities using a density-based clustering community detection algorithm includes: Extract user behavior sequences from users' historical interaction data, including click time sequences, reading duration sequences, and comment behavior sequences; The user behavior sequence is converted into a set of points in a high-dimensional feature space, and the local density and relative distance of each point in the set are calculated. Density peaks are automatically identified based on local density and relative distance, and these peaks are used as community centers. The nearest neighbor propagation algorithm is used to assign non-central points to each community center, forming multiple initial user communities; For each initial user community, boundary optimization is performed, overlapping communities are merged and overly large communities are split, ensuring that the size of each user community is within a preset range, resulting in multiple user communities.

[0007] Preferably, the extraction of group interest patterns for each user community includes: Aggregate the historical news clicks of all users within the user community to construct a news keyword co-occurrence matrix; Singular value decomposition is performed on the co-occurrence matrix of news keywords to extract latent semantic topic vectors; Calculate the weight distribution of latent semantic topic vectors to obtain the topic preference distribution map of the user community; Generate interest vectors for user communities based on topic preference distribution maps. These interest vectors include topic intensity and topic diversity values. Extract demographic characteristics of the user community from user registration information and historical interaction data; Interest vectors are concatenated with the demographic characteristics of the user community to form a group interest pattern.

[0008] Preferably, the correlation strength between the calculated news and the user community includes: Deep neural networks are used to encode news headlines and body text, outputting news semantic vectors; Attention is calculated on the group interest patterns of the user community to generate user interest query vectors; Calculate the dot product similarity between the news semantic vector and the user interest query vector as the original relevance; Introduce a credibility rating for news sources to correct the original relevance. The decay curve of news release time is used to perform time decay processing on the corrected correlation and output the correlation strength.

[0009] Preferably, the step of dynamically weighting the correlation strength by combining the timeliness parameter of the news and the social influence coefficient of the user community to generate the initial news push weight includes: Obtain the difference between the news release time and the current time, calculate the timeliness parameter, and the timeliness parameter is inversely proportional to the difference; Degree centrality and betweenness centrality are extracted from the social relationship graph of the user community and then fused to generate a social influence coefficient. The initial news push weights are obtained by inputting the correlation strength, timeliness parameter, and social influence coefficient into a weighted summation function. Set a weight threshold. When the initial news push weight is lower than the weight threshold, trigger a weight compensation mechanism to replace the initial news push weight with a default weight value.

[0010] Preferably, the adjustment of the initial news push weight based on network propagation, utilizing the topological relationship network between user communities, includes: Construct a directed weighted graph of user communities, where nodes represent user communities and edge weights represent the frequency of information flow between communities; Perform a multi-step random walk starting from each node, and collect the weight propagation values ​​on the walk path; The cumulative sum of the weight propagation values ​​between nodes is used as the network propagation coefficient; The initial news push weights are convolved with the network propagation coefficients and then normalized to obtain the adjusted news push weights. The adjusted news push weights are normalized to ensure that the total weight is one.

[0011] Preferably, the cumulative sum of the weight propagation values ​​among the computing nodes, as the network propagation coefficient, includes: Initialize the propagation weight of each node to the initial news push weight; In each iteration of the multi-step random walk, the propagation weight of the current node is distributed to the neighboring nodes according to the edge weight ratio; The number of iterations is adjusted according to the activity level of the user community; the higher the activity level, the more iterations are required. After accumulating multiple iterations, the weight values ​​received by each node are used to generate a weight propagation value matrix; The network propagation coefficients are obtained by row normalizing the weight propagation value matrix.

[0012] Preferably, the present invention further includes a news push device, the device comprising: The data acquisition module is used to collect users' historical interaction data and diverse feature data from news articles to construct a multi-source fusion dataset. The user grouping module is used to group users into multiple user communities using a density clustering-based community detection algorithm, and extract the group interest patterns of each user community. The association strength calculation module is used to calculate the association strength between the news and the user community based on the group interest patterns and the multi-dimensional feature data of the news through a semantic matching model. The dynamic weighting module is used to dynamically weight the correlation strength by combining the timeliness parameter of the news and the social influence coefficient of the user community, and generate the initial news push weight. The weight adjustment module is used to adjust the initial news push weight based on network propagation by utilizing the topological relationship network between user communities, so as to obtain the adjusted news push weight. The sorting and push module pushes news in descending order according to the corresponding news push weight.

[0013] Preferably, the present invention also includes an electronic device, the device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program to implement the news push method as described above.

[0014] Preferably, the present invention further includes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the news push method described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This community detection algorithm, employing density-based clustering, identifies communities based on the natural distribution density in the user feature space, without requiring a pre-defined number of communities. It can discover user communities of arbitrary shapes, closely resembling the complex forms that interest groups may exhibit in reality, such as non-spherical or streamlined shapes. The algorithm automatically identifies sparsely distributed users who do not belong to any dense area as noise points, preventing these anomalous user data from interfering with the modeling of mainstream group interest patterns. This allows for the extraction of more representative and purer group interest patterns, resulting in improved accuracy and robustness in user group segmentation compared to traditional partitioning clustering methods.

[0016] By constructing a topological network of relationships between user communities and applying a network propagation model to adjust the initial news push weights, a dynamic influence transmission mechanism between communities is introduced. This mechanism simulates the process of information diffusion in social networks, enabling news with high weight within a particular community to automatically and dynamically increase its weight in related communities based on network connection strength. This adjustment allows the recommendation system to move beyond isolated community analysis, capturing potential trends in the cross-community spread of interests and achieving more forward-looking recommendations. The recommendation results not only reflect static interests within communities but also embody dynamic changes in network structure, enhancing the system's responsiveness to evolving trends in trending topics. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the news push method described in this invention. Figure 2 A flowchart for extracting user community group interest patterns; Figure 3 A flowchart for calculating the strength of the association between news and user communities; Figure 4 This is a graph showing the correlation strength calculation and analysis. Figure 5 This is a graph for dynamic weighted parameter analysis. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1This invention provides a news push method, comprising: collecting users' historical interaction data and news multi-dimensional feature data to construct a multi-source fusion dataset; the historical interaction data includes user clicks, reading time, and comment behavior records, and the news multi-dimensional feature data covers dimensions such as title, body, publication time, and source credibility. A density-based clustering community discovery algorithm is used to group users into multiple user communities, and the group interest pattern of each user community is extracted; the group interest pattern represents the common preference characteristics of users within the community. Based on the group interest pattern and the news multi-dimensional feature data, a semantic matching model is used to calculate the association strength between the news and the user community, and the association strength reflects the degree of matching between the news content and the community's interests. Combining the news's timeliness parameter and the user community's social influence coefficient, the association strength is dynamically weighted to generate an initial news push weight; the timeliness parameter is calculated based on the decay of the news publication time, and the social influence coefficient is extracted from users' social relationships. Utilizing the topological relationship network between user communities, the initial news push weight is adjusted based on network propagation to obtain the adjusted news push weight; the topological relationship network is represented by a directed weighted graph to represent the information flow between communities. The adjusted news push rankings are sorted in descending order, and news is then pushed to users accordingly.

[0020] Example 1: In specific implementation, the user grouping process of the news push method begins with extracting user behavior sequences from users' historical interaction data. These sequences include click time sequences, reading duration sequences, and comment behavior sequences. The click time sequence is obtained by parsing the timestamp field in the user's access log, recording the specific time point when a user clicks on a news article. The reading duration sequence is calculated from the session data, showing the start and end time difference of each news article read. The comment behavior sequence collects the content and frequency information of comments posted by users. During the extraction of user behavior sequences, the system cleans and preprocesses the raw data, removing duplicate records and outliers, such as excessively short reading durations or invalid click events, to ensure data quality. The user behavior sequences are organized chronologically, forming an independent behavior profile for each user. The extraction of user behavior sequences can also include sliding window analysis, dividing long-term sequences into multiple sub-sequences to capture short-term behavior patterns, but the core implementation focuses on the integrity of the overall sequence.

[0021] Converting user behavior sequences into a point set in a high-dimensional feature space involves feature engineering steps, where each user's behavior sequence is mapped to a numerical vector. Click time series are transformed into numerical features using time encoding techniques, such as converting timestamps into cyclic encodings of hours or days of the week. Reading duration sequences are summarized using statistics such as mean, variance, and maximum value. Comment behavior sequences are extracted using natural language processing to extract keyword frequency or sentiment scores. The dimensionality of the high-dimensional feature space is determined by the number of features and is typically reduced using principal component analysis or autoencoders to retain key information. Each point in the point set corresponds to a user's coordinates in the feature space. When calculating the local density and relative distance of each point in the point set, the local density is calculated using a kernel density estimation method, employing a Gaussian kernel function based on the Euclidean distance between a point and its neighbors. The bandwidth parameter is optimized through cross-validation. The relative distance is obtained by traversing the point set and finding the minimum distance to points with higher density for each point. In practice, the calculation of local density and relative distance can be parallelized to improve efficiency, especially suitable for large-scale user datasets.

[0022] Density peaks are automatically identified based on local density and relative distance, characterized by high local density and high relative distance. The system scans the point set, calculates the product or weighted sum of local density and relative distance for each point, and sets a threshold to filter out peak points. This threshold is dynamically adjusted based on the distribution of the point set. Density peaks are selected as community centers, representing the core behavioral patterns within the user group. Non-center points are assigned to community centers using the nearest neighbor propagation algorithm. This algorithm iteratively updates the affiliation value based on the similarity matrix between points, calculated using cosine similarity or the reciprocal of Euclidean distance. The algorithm runs until convergence, forming stable community assignments. In practice, parameters of the nearest neighbor propagation algorithm, such as the damping factor, can be adjusted to control the convergence speed and avoid oscillations. After forming the initial user communities, boundary optimization is performed. Boundary optimization includes merging overlapping communities and splitting overly large communities. Overlapping communities are detected by calculating the Jaccard similarity of user behavior sequences between communities; communities with similarity higher than a preset threshold are merged. Overly large communities are split into multiple sub-communities using the k-means clustering algorithm based on a user number threshold. Understandably, boundary optimization ensures that the size of each user community is within a preset range, which is set by the system administrator according to application needs, such as a community size between 50 and 500 users, in order to balance computational efficiency and group representativeness.

[0023] In practical implementation, the extraction of user behavior sequences is further refined. Click time series can include click frequency and pattern recognition, such as identifying user active periods. Reading duration sequences can be normalized to eliminate individual differences, and comment behavior sequences utilize text analysis to extract topic consistency. The construction of high-dimensional feature spaces can employ deep learning models such as variational autoencoders to capture nonlinear features. In local density calculations, the kernel function selection includes the Epanechnikov kernel or a uniform kernel, determined based on the data distribution. Relative distance calculations are optimized by using a KD-tree data structure to accelerate nearest neighbor search. Identification of density peak points can be combined with visualization tools such as decision graphs to assist in threshold setting. The implementation of the nearest neighbor propagation algorithm includes asynchronous update strategies to reduce memory usage. In the boundary optimization stage, overlapping communities can be merged using hierarchical clustering methods, while excessively large communities can be split using spectral clustering techniques. Optionally, the preset interval for community size can be dynamically adjusted based on changes in the total number of users. The user grouping process can also integrate a real-time update mechanism to periodically recalculate communities to reflect changes in user behavior.

[0024] It is understandable that each step of the user grouping process relies on configurable parameters, which are initialized through experimentation or domain knowledge and adaptively adjusted during operation. The entire implementation emphasizes the automation and scalability of the algorithm, ensuring applicability to news push systems of varying sizes. The transformation of the high-dimensional feature space point set ensures a multi-dimensional representation of user behavior; the calculation of local density and relative distance provides the basis for clustering; density peak point identification enables automatic center selection; the nearest neighbor propagation algorithm guarantees the stability of community formation; and boundary optimization handles the rationality of the community structure. In specific implementations, point set transformation may involve feature scaling such as standardization or normalization to eliminate the influence of dimensions. In local density calculation, the choice of neighbor radius affects the density value, determined through the elbow rule. The calculation of relative distance can introduce weighted distance to handle feature importance. Density peak point identification can add constraints such as a minimum density threshold. The convergence criterion of the nearest neighbor propagation algorithm is set to the number of iterations or a change threshold. In boundary optimization, the threshold for merging overlapping communities is calibrated based on historical data, and the k-value for splitting excessively large communities is optimized through the silhouette coefficient.

[0025] The extraction and processing of user behavior sequences involves large-scale data streams, utilizing distributed computing frameworks such as Apache Spark for parallel processing. The high-dimensional feature space point set is stored in matrix format for rapid computation. Algorithms for calculating local density and relative distance are optimized for incremental updates to support online learning. The density peak point identification module outputs a community center list for subsequent use. The nearest neighbor propagation algorithm includes a fault recovery mechanism to ensure robustness. Boundary optimization operations are logged for auditing and debugging. In the specific implementation, the output of the user grouping process is a user community mapping table, linking each user to their respective community. The entire process is integrated into the data preprocessing pipeline of the news push system as a pre-step for extracting group interest patterns. It is understood that the accuracy of user grouping directly affects the subsequent push effect; therefore, the implementation emphasizes the stability and interpretability of the algorithm. The dimensionality of the high-dimensional feature space point set can be reduced through feature selection methods. Local density calculation uses approximate algorithms to handle large datasets. Relative distance calculation caches intermediate results to accelerate processing. Density peak point identification uses heuristic rules to improve efficiency. The nearest neighbor propagation algorithm optimizes the message passing protocol, and boundary optimization incorporates graph partitioning theory to enhance quality. The user community formation process can be fine-tuned based on human feedback.

[0026] Example 2: See Figure 2 In practical implementation, the extraction of group interest patterns for each user community is illustrated through a specific example. Assume the system processes a user community of 500 users, formed from a group of users across the entire user base. The system aggregates the historical news clicks of all users within the community to construct a news keyword co-occurrence matrix. For example, if all users in the community clicked on 1000 news articles in the past three months, and keywords for each article are extracted using natural language processing (NLP), including keywords such as "election," "economy," and "sports," the news keyword co-occurrence matrix is ​​a symmetric matrix where rows and columns represent keywords, and matrix elements record the frequency of keyword co-occurrence within the same news article. The system uses a sliding window method to process the news text, with a window size of 5 words, and counts the number of keyword co-occurrences. The dimension of the news keyword co-occurrence matrix is ​​determined by the number of unique keywords, and noise is reduced through stop word filtering and stemming.

[0027] Singular value decomposition (SVD) is performed on the news keyword co-occurrence matrix. SVD decomposes the matrix into a product of three matrices to reduce dimensionality and capture latent semantic structure. Latent semantic topic vectors are extracted from the first k columns of matrix V, with each vector representing the distribution of a latent topic in the keyword space. In practice, SVD uses an iterative algorithm to process large-scale matrices. The weight distribution of the latent semantic topic vectors is calculated by statistically analyzing the frequency of each latent semantic topic in clicked news articles within the user community. Keywords are extracted individually from each historically clicked news article within the user community to construct a keyword frequency vector for each article, resulting in a news vector whose keyword set is consistent with the keyword set of the news keyword co-occurrence matrix. Frequency calculation is based on the dot product similarity between the topic vector and the news vector, and after normalization, a topic preference distribution map of the user community is obtained. The topic preference distribution map is represented in vector form, for example, [Topic 1: 0.3, Topic 2: 0.5, Topic 3: 0.2], reflecting the community's preference strength for each topic.

[0028] Interest vectors for the user community are generated based on the topic preference distribution map. These interest vectors include topic strength and topic diversity values. The topic strength value is calculated by weighted summation of the latent semantic topic vectors, with the weights being the values ​​in the topic preference distribution map. The topic diversity value is measured using Shannon entropy to determine the dispersion of the topic distribution. The Shannon entropy formula is used to calculate the topic diversity value.

[0029] in: Indicates the theme diversity value. This represents the weight of the i-th topic in the topic preference distribution graph. Indicates the total number of topics. This is the base of the logarithm, usually set to 2. Thematic diversity value. The higher the value, the more dispersed the interests of the user community. The interest vector, as a multidimensional numerical feature, is used to characterize the community's interest characteristics. Demographic features of the user community are extracted from user registration information and historical interaction data. These features, including average user age, gender ratio, and geographic information, are concatenated with the interest vector. Examples of demographic features extracted from user profile data include an average age of 35, a gender ratio of 60% male, and a predominantly urban geographic distribution. The demographic features are encoded, such as using one-hot encoding for categorical variables and standardizing numerical variables. The concatenation operation is performed through vector joins to form a group interest pattern. This group interest pattern is stored as a fixed-length vector for subsequent calculation of association strength.

[0030] In practical implementation, the construction of the news keyword co-occurrence matrix is ​​further refined. Keyword extraction uses TF-IDF weighting to distinguish importance, and co-occurrence statistics introduce a decay factor to handle time proximity. Singular value decomposition can be accelerated using stochastic singular value decomposition. The extraction of latent semantic topic vectors includes a topic naming step to enhance interpretability. The calculation of weight distribution can introduce smoothing techniques to prevent zero-frequency problems. The generation of topic preference distribution maps can be visualized, but the core is still numerical vectors. The topic intensity value of interest vectors is calculated using a weighted average, and the logarithmic base b in the calculation of topic diversity value can be chosen as the natural logarithm e. The concatenation of demographic features can be dimensionality-reduced to avoid the curse of dimensionality. Each step in the process of extracting group interest patterns involves parameter tuning and verification. For example, the construction of the news keyword co-occurrence matrix requires selecting an appropriate number of keywords, and the order of singular value decomposition affects the number of topics. The sparsity of the news keyword co-occurrence matrix can be mitigated by padding or thresholding. The numerical stability of singular value decomposition is guaranteed by regularization. The calculation of the weight distribution of latent semantic topic vectors can use a soft assignment method, and the normalization of the topic preference distribution map ensures probabilistic properties. Interest vector generation emphasizes statistical representativeness, while the splicing of demographic features enhances pattern diversity. In practical implementation, an example of constructing a news keyword co-occurrence matrix shows that for a community of 500 users, processing 1000 news articles extracts 200 keywords, resulting in a 200x200 matrix. Singular value decomposition is set to k=10, extracting 10 potential topics. The topic preference distribution map is calculated based on a weighted average of the topic's frequency of occurrence in the news.

[0031] Example 3: See Figure 3In practical implementation, the calculation of the correlation strength between news and user communities is illustrated through a specific example. Assume the system processes a user community with group interest patterns extracted from Example 2, and a news article to be pushed, containing a title and body text such as "Global Climate Change Summit Reaches New Agreement." A deep neural network is used to encode the news title and body text, outputting a news semantic vector. The deep neural network employs a pre-trained Transformer model. The input news text is segmented into word sequences, converted into word vector sequences through an embedding layer, and the encoding layer uses a multi-head self-attention mechanism to capture contextual information. Finally, a fixed-dimensional news semantic vector is generated through pooling operations. The news semantic vector is a high-dimensional real-number vector, for example, 512-dimensional, representing the semantic content of the news. In practical implementation, the inference process of the deep neural model can batch process multiple news articles to improve efficiency, but the core implementation focuses on the encoding of a single news article. Attention is calculated on the group interest patterns of the user community to generate user interest query vectors. The attention mechanism is based on a query-key-value model. The group interest pattern serves as the query vector, which is transformed linearly to generate a query matrix. The key and value matrices are derived from the intermediate representation of the news semantic vector. The attention weights are calculated using a softmax function to determine the similarity between the query and the key. The weighted summation of the value vectors is then used to output the user interest query vector. The user interest query vector has the same dimension as the news semantic vector, for example, 512 dimensions, focusing on semantic features related to the user community.

[0032] The dot product similarity between the news semantic vector and the user interest query vector is calculated as the original relevance. The dot product similarity is calculated as the inner product of the two vectors, and the result is a scalar value reflecting the degree of alignment of the vectors in space. The dot product similarity formula is:

[0033] in: Indicates the original degree of correlation. Represents the semantic vector of news. This represents a user interest query vector, and the dot product operation calculates the sum of the products of corresponding elements. The original relevance value has an uncertain range and can be any real number, depending on the vector normalization. A credibility level for the news source is introduced to correct the original relevance; the credibility level is obtained from a predefined credibility database, for example, news sources are divided into high, medium, and low levels, corresponding to correction factors of 1.2, 1.0, and 0.8. The correction operation is implemented through multiplication, and the corrected relevance... ,in This is a credibility correction factor. The decay curve of the news release time is applied to the corrected relevance for time decay processing, outputting the relevance strength. The decay curve uses an exponential decay model, calculating the decay factor based on the time difference between the news release time and the current time. The time decay processing formula is:

[0034] in: Indicates the strength of the association. This indicates the time difference (in hours) between the time the news was released and the current time. This is the attenuation rate parameter, which controls the attenuation rate. The setting is determined by fitting historical data; for example, a value of 0.01 indicates a decay of approximately 63% every 100 hours. Correlation strength. As the final output, it is used for subsequent dynamic weighting. It can be understood that the association strength calculation integrates deep learning and rule base correction to ensure that the result reflects both semantic matching and external factors.

[0035] In practical implementation, when deep neural networks encode news headlines and body text, long texts can be truncated or segmented to adapt to the model's input length limitations. Residual connections can be introduced to prevent information loss during the generation of user interest query vectors. L2 normalization is applied to the vectors before calculating dot product similarity, limiting the values ​​to the [-1, 1] range to improve stability. The credibility correction factor c is assigned based on the source's authority assessment; for example, mainstream news organizations are assigned a high level. In the time decay processing, the time difference Δt is calculated to the hour, and the decay rate α can be adjusted according to the news category; for example, a smaller α value slows down decay for breaking news. Optionally, smoothing processing can be added to the association strength calculation to prevent outliers.

[0036] It is understandable that each step in the association strength calculation process depends on parameter configuration. For example, the choice of deep neural network model affects the quality of semantic vectors, and the number of heads in the attention mechanism affects the representational ability of query vectors. The generation of news semantic vectors ensures a deep understanding of the content, the generation of user interest query vectors focuses on community preferences, dot product similarity provides a basic matching metric, credibility correction enhances the reliability of the results, and time decay processing reflects the importance of timeliness. In a specific implementation, an example shows that for user community interest pattern vectors and news semantic vectors, the dot product similarity S is calculated to be 0.75, indicating a high credibility level for the news source. After correction Time difference Hours, decay rate Correlation strength The entire process example demonstrates the complete calculation from the original news to the strength of association, emphasizing the algorithm's practicality and scalability.

[0037] See Figure 4 This chart comprehensively illustrates the interaction of various factors in the correlation strength calculation process using scatter plots and trend lines. The horizontal axis represents the time difference between the news release time and the current time, while the vertical axis displays the calculated correlation strength value. Color coding is used to distinguish news sources with different credibility levels: red represents high-credibility news, blue represents medium-credibility news, and green represents low-credibility news. Several important phenomena can be observed from the chart: First, the correlation strength generally decreases gradually as the time difference increases, reflecting the significant impact of news timeliness. Second, high-credibility news maintains a high correlation strength value under the same time conditions, indicating that the credibility correction mechanism effectively improves the push priority of authoritative news sources. The trend line clearly reflects the application effect of the exponential decay model, consistent with the objective law that news value decreases over time. This visualization analysis provides an intuitive basis for algorithm optimization, helping developers understand the rationality of the time decay parameter settings and the effectiveness of credibility weight allocation. By adjusting these parameters, the accuracy of news pushes and user satisfaction can be further optimized.

[0038] Example 4: In specific implementation, the process of dynamically weighting the association strength by combining the timeliness parameter of the news and the social influence coefficient of the user community is illustrated through a specific example. Assume the system processes a user community and a news article. The association strength of the news, calculated from Example 3, is 0.556. The news was published on October 1, 2023 at 12:00 PM, and the current time is October 3, 2023 at 12:00 PM. The difference between the news's publication time and the current time is obtained, and the timeliness parameter is calculated. The timeliness parameter is inversely proportional to the difference. The time difference is calculated in hours; for example, in this example, the time difference is 48 hours. The timeliness parameter is calculated using an inverse proportional function, the formula of which is:

[0039] in: Indicates timeliness parameter, This indicates the time difference between the news release time and the current time (in hours). This is the attenuation coefficient, which controls the strength of the inverse relationship. The attenuation coefficient is set to 0.02 based on historical data calibration. The timeliness parameter value ranges from 0 to 1; the larger the time difference, the smaller the timeliness parameter. In practical implementation, the calculation of the time difference takes into account time zone unification and leap second adjustment to ensure accuracy. Refer to Table 1 for the calculation of the timeliness parameter, which shows the timeliness parameter values ​​under different time differences. The attenuation coefficient is fixed at 0.02.

[0040] Table 1: Calculation Table of Timeliness Parameters

[0041] The degree centrality and betweenness centrality of a user community are extracted from its social relationship graph and then fused to generate a social influence coefficient. The social relationship graph is constructed based on the attention or interaction relationships between users. The graph can be directed or undirected, with nodes representing users and edges representing relationships. Degree centrality calculates the number of connections (neighbors) for each user within the community, and the average value is taken as the community degree centrality. Betweenness centrality calculates the frequency with which each user appears on the shortest path between other users, and the average value is taken as the community betweenness centrality. The social influence coefficient is generated using a weighted average formula, for example:

[0042] in: Indicates degree centrality. Indicates betweenness centrality, and These are the weighting coefficients, which are set to 0.6 and 0.4 using principal component analysis. In this specific example, assuming the user community has a degree centrality of 150 and a betweenness centrality of 0.05, the social influence coefficient is calculated as follows: In some embodiments, the construction of a social relationship graph may include a dynamic update mechanism, but the core implementation focuses on static extraction.

[0043] The initial news push weights are obtained by inputting the correlation strength, timeliness parameter, and social influence coefficient into a weighted summation function; the weighted summation function has the following form:

[0044] in: Indicates the initial news push weight. Indicates the strength of the association. Indicates timeliness parameter, Indicates the social influence coefficient. , , These are weighting coefficients, which are optimized using grid search and set to 0.5, 0.3, and 0.2. In this specific example, the association strength is 0.556, the timeliness parameter is 0.510, the social influence coefficient is 90.02, and the initial news push weight is calculated as follows: .

[0045] Before weighted summation, the input values ​​are standardized. For example, the correlation strength and timeliness parameters are divided by their maximum values, and the social influence coefficient is divided by the baseline value of 100 to eliminate the influence of dimensions. Therefore, the standardized value of the correlation strength is 0.556, the standardized value of the timeliness parameter is 0.510, and the standardized value of the social influence coefficient is 0.9002. The final news push weight is calculated as follows: However, the basic implementation directly uses the original values. It's understandable that the weighting coefficient affects the push notification bias and needs to be adjusted according to the application scenario. A weight threshold is set; when the initial news push weight is lower than the threshold, a weight compensation mechanism is triggered, replacing it with the default weight value. The weight threshold is set to 1.0 based on historical quantiles. In the specific example, the initial news push weight of 18.435 is higher than the threshold, so compensation is not triggered. Optionally, the weight compensation mechanism can include a gradual replacement strategy, but the core implementation is direct replacement.

[0046] In practical implementation, the calculation of timeliness parameters can incorporate news type factors; for example, breaking news has a smaller decay coefficient, but a uniform decay coefficient is used in the basic implementation. The integration of social influence coefficients can incorporate community size normalization to prevent large communities from dominating. The weight coefficients of the weighted summation function can be dynamically updated through online learning. The weight threshold settings can be adjusted based on user feedback. The calculation of the initial news push weight can be recorded in audit logs for analysis.

[0047] See Figure 5 The system employs a dual Y-axis design, simultaneously presenting the complex relationship between timeliness parameters and initial push weights over time. The left vertical axis represents the timeliness parameter, whose value decreases over time following an inverse proportional function; the right vertical axis displays the initially push weight calculated comprehensively. Analysis of the charts reveals several key characteristics: the timeliness parameter curve exhibits a typical inverse proportional function characteristic, decreasing rapidly in the initial stage and then leveling off, reflecting the non-linear nature of news value decay. The initial push weight curve displays a more complex pattern of change, influenced by varying weights of the timeliness parameter and social influence coefficient across different time intervals. When the news release time is short, the timeliness parameter dominates the push weight; as time progresses, the weights of other factors such as social influence gradually increase. This dynamic adjustment mechanism ensures that the news push system can maintain timeliness while also considering the social value of the content and the relevance to user interests. The comparative analysis of the two curves provides an important perspective for understanding the system's decision-making logic and helps to further optimize the weighting coefficient settings.

[0048] Example 5: In a specific implementation, the initial news push weights are adjusted based on network propagation using the topological relationship network between user communities. A concrete example illustrates this: assuming the system processes a topological relationship network formed by three user communities, labeled Community A, Community B, and Community C, with initial news push weights calculated from Example 4 as 18.435, 12.700, and 9.800 respectively. A directed weighted graph of user communities is constructed, where nodes represent user communities and edge weights represent the frequency of information flow between communities. The directed weighted graph is constructed based on historical interaction data; for example, the information flow frequency from Community A to Community B is 0.3, from Community B to Community C is 0.5, and from Community C to Community A is 0.2. The directed weighted graph is stored using an adjacency matrix, where matrix elements represent the edge weights pointing from one community to another. In a specific implementation, the information flow frequency is calculated by statistically analyzing the number of news forwards, comments, or shares between communities, and normalized to the range of 0 to 1. A multi-step random walk is performed starting from each node, collecting weight propagation values ​​along the walk path. The random walk begins at each community node, selecting a transition path based on edge weight probabilities. The number of walk steps is set to 3 steps, based on the network diameter. The weight propagation value records the initial cumulative news push weight value of the nodes traversed during the walk. The cumulative sum of weight propagation values ​​between nodes is calculated as the network propagation coefficient; this cumulative sum is obtained by iteratively summing the weight values ​​received by each node from all walk paths. The network propagation coefficient matrix represents the overall effect of weight propagation between communities. In the specific implementation, the random walk process uses a Markov chain model, and the transition probability is determined by the edge weights. The collection of weight propagation values ​​involves multi-path sampling; for example, a walk path starting from community A could be A→B→C→A.

[0049] Initialize the propagation weight of each node to the initial news push weight. In each walk iteration, distribute the propagation weight of the current node to neighboring nodes according to the edge weight ratio. After accumulating multiple iterations, the weight values ​​received by each node are used to generate a weight propagation value matrix. The rows of the weight propagation value matrix represent source nodes, and the columns represent target nodes. The matrix elements record the weight values ​​propagated from the source node to the target node. Perform row normalization on the weight propagation value matrix to obtain the network propagation coefficients; row normalization ensures that the sum of the output weights of each source node is one. The network propagation coefficient matrix is ​​used to adjust the initial news push weight. The number of iterations is adjusted according to the activity level of the user community; the higher the activity level, the more iterations are needed. The activity level is measured by the average daily login frequency of community users; for example, the number of iterations is set to 5 for communities with high activity levels and 3 for communities with low activity levels. In the specific example, the calculation of the network propagation coefficients uses the following formula to describe the iteration process:

[0050] in: This represents the weight propagation matrix after the t-th iteration. This represents the weight propagation matrix after the (t-1)th iteration. This represents the transition probability matrix of a directed weighted graph; matrix multiplication is performed row-by-row. The initial weight propagation value matrix is ​​also shown. The diagonal elements are the initial news push weights, while the off-diagonal elements are zero.

[0051] After a preset number of iterations, the final weight propagation value matrix is ​​row-normalized to obtain the network propagation coefficient matrix. The initial news push weights are convolved with the network propagation coefficients to obtain the adjusted news push weights. A one-dimensional convolution operation is used, with the network propagation coefficient vector as the kernel. For example, the initial news push weights of each community are convolved with their network propagation coefficient vectors. The adjusted news push weights reflect the comprehensive weights after the network propagation effect. The adjusted news push weights are then normalized to ensure the total weight sum is one. Normalization uses the softmax function or simple scaling. In a specific example, assuming the initial news push weight of community A is 18.435 and the network propagation coefficient vector is [0.6, 0.3, 0.1], the convolution calculation is a weighted sum of 18.435 × 0.6 + 12.700 × 0.3 + 9.800 × 0.1 = 16.061, and the final normalization yields the total weight ratio. It can be understood that network propagation adjustment utilizes the inter-community relationships to enhance the rationality of the weight distribution.

[0052] In practical implementation, the construction of the directed weighted graph can include self-loop edges to represent information flow within the community. Random walks can introduce restart probabilities to handle sparse networks. The generation of the weight propagation value matrix can be accelerated using parallel computing. The normalization of the network propagation coefficients can adopt different norms. Convolution operations can choose different kernel sizes and strides. The adjustment of the number of iterations can be based on convergence detection rather than fixed values. Optionally, the calculation of the adjusted news push weights can introduce a decay factor to prevent weight diffusion. In some embodiments, the topological relationship network can be dynamically updated to reflect changes in community relationships. The adjustment process based on network propagation depends on the choice of graph structure and propagation model. The accuracy of the directed weighted graph affects the propagation effect, the number of steps in the random walk controls the propagation range, the accumulation and summation of weight propagation values ​​provide a global view, the network propagation coefficients quantify the degree of influence, and the convolution operation integrates local and global information. In a specific example, the topological relationship networks of the three user communities constitute a strongly connected graph, ensuring that the weights can be fully propagated. The entire process example demonstrates the complete calculation from the initial weights to the adjusted weights, emphasizing the network-aware characteristics of the algorithm.

[0053] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A news push method, characterized in that, The method includes: Collect users' historical interaction data and diverse feature data from news articles to construct a multi-source fusion dataset; A density-based clustering-based community detection algorithm is used to group users into multiple user communities, and the group interest patterns of each user community are extracted. Calculate the correlation strength between news and user communities based on group interest patterns and diverse characteristics of news data; By combining the timeliness parameter of the news and the social influence coefficient of the user community, the correlation strength is dynamically weighted to generate the initial news push weight; By leveraging the topological relationship network between user communities, the initial news push weight is adjusted based on network propagation to obtain the adjusted news push weight; News is pushed out in descending order of its corresponding news push weight.

2. The news push method according to claim 1, characterized in that, The method of grouping users into multiple user communities using a density-based clustering community detection algorithm includes: Extract user behavior sequences from users' historical interaction data, including click time sequences, reading duration sequences, and comment behavior sequences; The user behavior sequence is converted into a set of points in a high-dimensional feature space, and the local density and relative distance of each point in the set are calculated. Density peaks are automatically identified based on local density and relative distance, and these peaks are used as community centers. The nearest neighbor propagation algorithm is used to assign non-central points to each community center, forming multiple initial user communities; For each initial user community, boundary optimization is performed, overlapping communities are merged and overly large communities are split, ensuring that the size of each user community is within a preset range, resulting in multiple user communities.

3. The news push method according to claim 1, characterized in that, The extraction of group interest patterns for each user community includes: Aggregate the historical news clicks of all users within the user community to construct a news keyword co-occurrence matrix; Singular value decomposition is performed on the co-occurrence matrix of news keywords to extract latent semantic topic vectors; Calculate the weight distribution of latent semantic topic vectors to obtain the topic preference distribution map of the user community; Generate interest vectors for user communities based on topic preference distribution maps. These interest vectors include topic intensity and topic diversity values. Extract demographic characteristics of the user community from user registration information and historical interaction data; Interest vectors are concatenated with the demographic characteristics of the user community to form a group interest pattern.

4. The news push method according to claim 1, characterized in that, The strength of the correlation between the computational news and the user community includes: Deep neural networks are used to encode news headlines and body text, outputting news semantic vectors; Attention is calculated on the group interest patterns of the user community to generate user interest query vectors; Calculate the dot product similarity between the news semantic vector and the user interest query vector as the original relevance; Introduce a credibility rating for news sources to correct the original relevance. The decay curve of news release time is used to perform time decay processing on the corrected correlation and output the correlation strength.

5. The news push method according to claim 1, characterized in that, The method of dynamically weighting the correlation strength by combining the timeliness parameter of the news and the social influence coefficient of the user community to generate the initial news push weight includes: Obtain the difference between the news release time and the current time, calculate the timeliness parameter, and the timeliness parameter is inversely proportional to the difference; Degree centrality and betweenness centrality are extracted from the social relationship graph of the user community and then fused to generate a social influence coefficient. The initial news push weights are obtained by inputting the correlation strength, timeliness parameter, and social influence coefficient into a weighted summation function. Set a weight threshold. When the initial news push weight is lower than the weight threshold, trigger a weight compensation mechanism to replace the initial news push weight with a default weight value.

6. A news push method according to claim 5, characterized in that, The method of adjusting the initial news push weight based on network propagation by utilizing the topological relationship network between user communities includes: Construct a directed weighted graph of user communities, where nodes represent user communities and edge weights represent the frequency of information flow between communities; Perform a multi-step random walk starting from each node, and collect the weight propagation values ​​on the walk path; The cumulative sum of the weight propagation values ​​between nodes is used as the network propagation coefficient; The initial news push weights are convolved with the network propagation coefficients and then normalized to obtain the adjusted news push weights.

7. A news push method according to claim 6, characterized in that, The cumulative sum of the weight propagation values ​​among the computing nodes, as the network propagation coefficient, includes: Initialize the propagation weight of each node to the initial news push weight; In each iteration of the multi-step random walk, the propagation weight of the current node is distributed to the neighboring nodes according to the edge weight ratio; After accumulating multiple iterations, the weight values ​​received by each node are used to generate a weight propagation value matrix; The number of iterations is adjusted according to the activity level of the user community; the higher the activity level, the more iterations are required. The network propagation coefficients are obtained by row normalizing the weight propagation value matrix.

8. A news push device, characterized in that, The device includes: The data acquisition module is used to collect users' historical interaction data and diverse feature data from news articles to construct a multi-source fusion dataset. The user grouping module is used to group users into multiple user communities using a density clustering-based community detection algorithm, and extract the group interest patterns of each user community. The association strength calculation module is used to calculate the association strength between the news and the user community based on the group interest patterns and the multi-dimensional feature data of the news through a semantic matching model. The dynamic weighting module is used to dynamically weight the correlation strength by combining the timeliness parameter of the news and the social influence coefficient of the user community, and generate the initial news push weight. The weight adjustment module is used to adjust the initial news push weight based on network propagation by utilizing the topological relationship network between user communities, so as to obtain the adjusted news push weight. The sorting and push module pushes news in descending order according to the corresponding news push weight.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program to implement the news push method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the news push method according to any one of claims 1-7.

Images