Private domain traffic scheduling and content distribution method and system based on group behavior analysis

Abstract

The present invention provides a method and system for private domain traffic scheduling and content distribution based on group behavior analysis, which relates to the field of private domain traffic operation technology, including building a user social network by analyzing user historical behavior data, calculating user influence weights, and generating user group portraits in combination with user long-term and short-term interest distributions. The candidate content is semantically analyzed, a content dissemination network is constructed, the content dissemination coefficient and the matching coefficient between the content and the user group are calculated, and a target content ranking table is generated. The user active time period distribution is calculated according to the user activity prediction model, and a content delivery strategy is generated in combination with the target content ranking table and the matching coefficient, and the delivery strategy is dynamically adjusted according to the delivery effect index, thereby achieving accurate content distribution and improving the efficiency of private domain traffic operation.

Description

Technical Field

[0001] The present invention relates to private domain traffic operation technology, and in particular to a private domain traffic scheduling and content distribution method and system based on group behavior analysis. Background Art

[0002] With the development of the Internet, private traffic platforms have gradually become an important channel for efficient interaction between enterprises and users. Relying on its privacy and high stickiness, brands can establish long-term connections with users through precision marketing. However, the existing private traffic operation technology has the following shortcomings: First, the dynamic mining of user behavior data is insufficient, lacks the integration analysis of short-term and long-term interests, and cannot accurately capture changes in user needs; second, the content distribution method is mostly based on static tag matching or simple rules, and fails to effectively utilize user interaction behavior and content dissemination characteristics, resulting in serious homogeneity of recommendation results; third, there is a lack of refined operation of user grouping and judgment of the best time for delivery, and fails to push personalized content according to user activity; fourth, the content delivery strategy lacks a closed-loop optimization mechanism, and cannot adjust the strategy in real time according to the delivery effect, resulting in resource waste and reduced user experience.

[0003] To address the above issues, there is an urgent need for a private domain traffic scheduling and content distribution method based on user group behavior analysis. Through in-depth mining of user behavior data and time dimension analysis, combined with content dissemination characteristics, the recommendation strategy can be optimized, and the delivery strategy can be dynamically adjusted based on real-time feedback to improve content distribution efficiency and user interaction effects. Summary of the invention

[0004] The embodiments of the present invention provide a private domain traffic scheduling and content distribution method and system based on group behavior analysis, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] A private domain traffic scheduling and content distribution method based on group behavior analysis is provided, including:

[0007] Obtain historical behavior data of the target user group in the private domain traffic platform and perform preprocessing to obtain effective behavior data, construct a user-behavior association matrix based on the effective behavior data, obtain user interaction weights by calculating the interaction frequency and interaction depth between each user node in the user-behavior association matrix, construct a user social network based on the user interaction weights, calculate user influence weights using a node importance algorithm in the user social network, linearly combine the user-behavior association matrix with the user influence weights to obtain a user group portrait, perform time dimension analysis on the user group portrait, and extract the long-term interest distribution and short-term interest distribution of the target user group;

[0008] A set of candidate content to be distributed in a private domain traffic platform is obtained, semantic analysis is performed on the candidate content set to obtain content semantic features, similarities between contents in the candidate content set are calculated according to the content semantic features to obtain a content similarity matrix, a content propagation network is constructed based on the content similarity matrix, a propagation coefficient of each content node in the content propagation network is calculated, a matching coefficient between each content in the candidate content set and the target user group is calculated in combination with the long-term interest distribution and the short-term interest distribution, and a target content ranking table is generated according to the propagation coefficient and the matching coefficient;

[0009] The target user group is grouped according to the long-term interest distribution and the short-term interest distribution to obtain a user grouping result, a user activity prediction model is constructed according to the user grouping result, the user activity prediction model is used to calculate the active time period distribution of each group of users, a content delivery schedule is generated based on the active time period distribution and the target content ranking table, and the content delivery frequency of each user group is determined based on the matching coefficient, the content delivery schedule and the content delivery frequency are combined to generate a content delivery strategy, content delivery is executed on the private domain traffic platform according to the content delivery strategy, user interaction data on the delivered content is collected, delivery effect indicators are calculated according to the interaction data, the content delivery strategy is dynamically adjusted based on the delivery effect indicators, and content delivery is continued according to the adjusted content delivery strategy.

[0010] In an optional embodiment,

[0011] Constructing a user social network according to the user interaction weights, calculating user influence weights using a node importance algorithm in the user social network, and linearly combining the user-behavior association matrix with the user influence weights to obtain a user group portrait includes:

[0012] A user social network is constructed according to user interaction weights, including a multi-layer network structure, each layer of the network corresponds to an interaction type, and users are set as network nodes. The network edge weights are calculated according to a weighted combination of user interaction intimacy and user interest similarity, wherein the user interaction intimacy is obtained by performing time-attenuated weighted processing on the user interaction feature vector, and the user interest similarity is obtained by calculating the similarity of the topic vector of the user's historical interaction behavior sequence;

[0013] Calculating node importance indicators in the user social network, the node importance indicators include node degree centrality, node strength centrality, node betweenness centrality, node eigenvector centrality and node page ranking value;

[0014] An inter-layer influence propagation model is constructed based on the node importance index, and a random walk algorithm is used to calculate the influence propagation probability matrix between different layer networks, and the cross-layer influence diffusion coefficient is calculated in combination with the inter-layer correlation strength;

[0015] According to the influence propagation probability matrix and the cross-layer influence diffusion coefficient, an iterative algorithm is used to calculate the multi-layer network influence fusion function, and the influence weight of the user node is updated in each iteration to obtain a user influence weight vector;

[0016] Performing tensor decomposition on a pre-constructed user-behavior association matrix to obtain a user potential feature representation, wherein the user potential feature representation includes basic behavior features, temporal behavior features, and contextual behavior features;

[0017] The user-behavior association matrix is ​​linearly combined with the user influence weight vector, and the importance weights of different features are adaptively learned through the attention mechanism to finally obtain a user group portrait.

[0018] In an optional embodiment,

[0019] Constructing a content propagation network based on the content similarity matrix, calculating the propagation coefficient of each content node in the content propagation network, calculating the matching coefficient between each content in the candidate content set and the target user group in combination with the long-term interest distribution and the short-term interest distribution, and generating a target content ranking table according to the propagation coefficient and the matching coefficient includes:

[0020] Calculating the mean and standard deviation of the content similarity matrix, setting the sum of the mean and the standard deviation as a dynamic connection threshold, and constructing a content propagation network based on the dynamic connection threshold, wherein when the similarity in the content similarity matrix is ​​greater than the dynamic connection threshold, a connection edge is established between corresponding content nodes;

[0021] For each content node in the content dissemination network, the node degree value and the number of triangles associated with the node are calculated to obtain the node local connection feature, and the local propagation coefficient is calculated based on the node local connection feature; the number of shortest paths of the node and the number of shortest paths passing through the node are calculated to obtain the node global connection feature, and the global propagation coefficient is calculated based on the node global connection feature; the local propagation coefficient and the global propagation coefficient are adaptively weighted to obtain the node propagation coefficient;

[0022] Constructing a user cognitive feature vector, inputting the long-term interest distribution, the short-term interest distribution and the user cognitive feature vector into a self-attention network to obtain a user dynamic interest representation;

[0023] The semantic features of the candidate content are matched with the user's dynamic interest representation at the semantic level, cognitive level and context level, and the matching results of each level are weighted using an attention mechanism to obtain an initial matching score. The initial matching score is optimized in combination with the collaborative filtering characteristics of the user group to obtain a content matching coefficient;

[0024] The node propagation coefficient and the content matching coefficient are input into a pre-trained decision processing model to obtain a feature combination strategy, and an initial content ranking score is calculated according to the feature combination strategy; the initial content ranking score is diversifiedly adjusted in combination with content timeliness and topic popularity to generate a target content ranking table.

[0025] In an optional embodiment,

[0026] Input the node propagation coefficient and the content matching coefficient into a pre-trained decision processing model to obtain a feature combination strategy, and calculate the initial content ranking score according to the feature combination strategy; perform diversity adjustment on the initial content ranking score in combination with content timeliness and topic popularity, and generate a target content ranking table, including:

[0027] Construct a decision processing model, receive the node propagation coefficient and the content matching coefficient as input features, map the input features into a state vector, map the feature combination mode into an action vector, construct a state transition matrix based on real-time user feedback data, and calculate the ranking conversion rate according to the state transition matrix to obtain a reward vector;

[0028] Constructing a master control network and a mirror network, wherein the master control network receives a state vector and an action vector to generate a first state action value, and the mirror network receives a state vector and an action vector to generate a second state action value;

[0029] Calculate the time difference between the first state action value and the second state action value, construct a sample weight matrix based on the time difference, perform priority sampling on the training data according to the sample weight matrix to obtain experience playback data, input the experience playback data into the master control network and the mirror network respectively, and obtain a current state prediction value and a target state prediction value;

[0030] Performing time series sampling on the experience replay data to obtain a multi-step sampling sequence, calculating a target value function based on the multi-step sampling sequence and a target state prediction value, calculating a sampling calibration value according to a difference between a behavior strategy and a target strategy, and constructing a loss function based on the target value function and the sampling calibration value;

[0031] Calculating the parameter gradient of the master network according to the loss function, updating the network parameters of the master network based on the parameter gradient, copying the updated network parameters to the mirror network, and outputting a feature combination strategy based on the updated master network and the mirror network;

[0032] The feature combination strategy is multiplied by the input feature to obtain a content basic score, the content release time is obtained to calculate the time decay value, the topic heat value is calculated based on the topic propagation rate and the amount of interaction, and the content basic score is weightedly combined with the time decay value and the topic heat value to obtain a content ranking score;

[0033] The similarities between the to-be-sorted contents are calculated to obtain a content duplication matrix, the content sorting scores are diversity-adjusted based on the content duplication matrix to obtain a final sorting score, and a target sorting list is generated according to the final sorting score.

[0034] In an optional embodiment,

[0035] The formula for constructing the loss function based on the target value function and the sampled calibration value is as follows:

[0036] ;

[0037] Among them, L represents the loss function, N represents the total number of samples, and W ij represents the sample weight matrix, n represents the number of steps in the sample sequence, k represents the current time step, a k represents the content recommended by the recommendation system at time step k, s k represents the state of the environment at time step k, including user behavior features and content features, Indicates behavioral strategy, represents the target strategy, K represents the maximum number of steps for multi-step sampling, γ represents the discount factor, and r k Indicates that at time step k, due to taking action a k The reward obtained is Q 2 (s K+1 , a K+1 ) represents the future state s predicted by the mirror network K+1 Take action a K+1 The value of Q 1( s i , a i ) represents the current state s predicted by the master network i Take action a i The value of .

[0038] In an optional embodiment,

[0039] Constructing a user activity prediction model according to the user grouping result, calculating the active time period distribution of each group of users by using the user activity prediction model, and generating a content delivery schedule based on the active time period distribution and the target content ranking table includes:

[0040] Based on the user grouping results, an initial causal knowledge graph including user group entity nodes, user behavior entity nodes and environmental factor entity nodes is constructed, the causal relationship strength between each entity node is calculated using a structural causal model, a relationship strength matrix is ​​generated based on the causal relationship strength, and the initial causal knowledge graph and the relationship strength matrix are combined to form a time-varying causal knowledge graph;

[0041] Collecting the newly added node-to-node relationship edges of the time-varying causal knowledge graph within a preset time period, inputting the newly added node-to-node relationship edges into a recurrent neural network to obtain dynamic change characteristics of the node-to-node relationship, and updating the relationship strength matrix based on the dynamic change characteristics of the node-to-node relationship to obtain an updated relationship strength matrix;

[0042] Constructing a hierarchical user activity prediction model according to the user grouping result, wherein the hierarchical user activity prediction model includes a causal convolutional network layer for extracting active mode features of user groups, a multi-head causal attention network layer for calculating temporal influence features between user groups, and a neural ordinary differential equation layer for obtaining continuous dynamic features of user activity;

[0043] Input the updated relationship strength matrix into the hierarchical user activity prediction model to obtain user group activity mode features, temporal influence features between user groups and continuous dynamic features of user activity, input the user group activity mode features, temporal influence features between user groups and continuous dynamic features of user activity into the causal fusion network for weighted fusion, and introduce causal bias values ​​to obtain user activity prediction results;

[0044] A Bayesian causal network is established based on the user activity prediction result, a variational inference method is used to calculate the probability distribution of content delivery time, an initial content delivery time point is determined through importance sampling according to the content delivery time probability distribution, and a delivery effect deviation is calculated based on a counterfactual analysis method, a delivery time adjustment window is determined according to the delivery effect deviation, an online convex optimization algorithm is used to calculate a delivery parameter adjustment value, and the initial content delivery time point is optimized according to the delivery parameter adjustment value;

[0045] Multi-dimensional consistency constraint processing is performed on the user activity prediction result and the optimized content delivery time point to obtain a final content delivery schedule that satisfies the multi-dimensional consistency constraint.

[0046] In an optional embodiment,

[0047] Calculating the delivery effect deviation based on the counterfactual analysis method, determining the delivery time adjustment window according to the delivery effect deviation, calculating the delivery parameter adjustment value using an online convex optimization algorithm, and optimizing the initial content delivery time point according to the delivery parameter adjustment value includes:

[0048] Obtaining the initial content delivery time point, constructing a counterfactual evaluation model, the counterfactual evaluation model comprising a causal intervention operation module and a random sampling module, calculating the user response expectation value before and after the delivery time adjustment through the causal intervention operation module, and performing importance sampling on the user response expectation value through the random sampling module to obtain a counterfactual delivery effect evaluation value;

[0049] A time window adjustment function is constructed based on the counterfactual delivery effect evaluation value, wherein the time window adjustment function calculates the adaptive parameters of the time window boundary according to the effect evaluation deviation value and the effect fluctuation measurement value, and dynamically updates the time window boundary using the gradient descent method to obtain an optimized delivery time window;

[0050] Inputting the optimized delivery time window into an online convex optimization model, wherein the online convex optimization model includes a delivery effect loss function and a regularization constraint term, and adopting an adaptive subgradient descent algorithm with momentum term to iteratively optimize the delivery parameters to obtain optimized delivery parameters;

[0051] Constructing a parameter adjustment feedback model based on the optimized delivery parameters, wherein the parameter adjustment feedback model extracts candidate parameters from the posterior distribution of parameters using a probability sampling method, updates the parameter distribution in real time through a state estimation filter, and outputs the optimal delivery parameters;

[0052] The initial content delivery time point is adjusted according to the optimal delivery parameters to obtain an optimized content delivery time point.

[0053] According to a second aspect of the embodiments of the present invention,

[0054] A private domain traffic scheduling and content distribution system based on group behavior analysis is provided, including:

[0055] The first unit is used to obtain historical behavior data of a target user group in a private domain traffic platform and perform preprocessing to obtain effective behavior data, construct a user-behavior association matrix based on the effective behavior data, obtain user interaction weights by calculating the interaction frequency and interaction depth between each user node in the user-behavior association matrix, construct a user social network based on the user interaction weights, calculate user influence weights using a node importance algorithm in the user social network, linearly combine the user-behavior association matrix with the user influence weights to obtain a user group portrait, perform time dimension analysis on the user group portrait, and extract long-term interest distribution and short-term interest distribution of the target user group;

[0056] The second unit is used to obtain a set of candidate content to be distributed in the private domain traffic platform, perform semantic analysis on the candidate content set to obtain content semantic features, calculate the similarity between the contents in the candidate content set according to the content semantic features to obtain a content similarity matrix, build a content propagation network based on the content similarity matrix, calculate the propagation coefficient of each content node in the content propagation network, calculate the matching coefficient between each content in the candidate content set and the target user group in combination with the long-term interest distribution and the short-term interest distribution, and generate a target content ranking table according to the propagation coefficient and the matching coefficient;

[0057] The third unit is used to group the target user group according to the long-term interest distribution and the short-term interest distribution to obtain user grouping results, build a user activity prediction model according to the user grouping results, use the user activity prediction model to calculate the active time period distribution of each group of users, generate a content delivery schedule based on the active time period distribution and the target content ranking table, and determine the content delivery frequency of each user group based on the matching coefficient, combine the content delivery schedule and the content delivery frequency to generate a content delivery strategy, execute content delivery in the private domain traffic platform according to the content delivery strategy, collect user interaction data on the delivered content, calculate delivery effect indicators according to the interaction data, dynamically adjust the content delivery strategy based on the delivery effect indicators, and continue to execute content delivery according to the adjusted content delivery strategy.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] A fourth aspect of the embodiments of the present invention is:

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] In this embodiment, by analyzing the user's historical behavior data, interactive relationships and influence, a multi-dimensional user group portrait is constructed, and long-term interests and short-term interests are distinguished to achieve accurate grasp of user needs. Based on content semantic analysis and user interest matching, combined with the content propagation coefficient, a personalized target content ranking table is generated, and an accurate content delivery schedule and frequency are formulated based on the user activity prediction model to achieve personalized content recommendation and efficient distribution. Collect user interaction data in real time, and dynamically adjust the content delivery strategy based on the delivery effect indicators to achieve continuous optimization of content delivery effects and improve the efficiency of private domain traffic operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flowchart of a method for private domain traffic scheduling and content distribution based on group behavior analysis according to an embodiment of the present invention;

[0067] Figure 2 It is a structural diagram of a private domain traffic scheduling and content distribution system based on group behavior analysis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 Schematic diagram of the process of a private domain traffic scheduling and content distribution method based on group behavior analysis according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] S101. Obtain historical behavior data of the target user group in the private domain traffic platform and perform preprocessing to obtain valid behavior data, construct a user-behavior association matrix based on the valid behavior data, obtain user interaction weights by calculating the interaction frequency and interaction depth between each user node in the user-behavior association matrix, construct a user social network based on the user interaction weights, calculate user influence weights using a node importance algorithm in the user social network, linearly combine the user-behavior association matrix with the user influence weights to obtain a user group portrait, perform time dimension analysis on the user group portrait, and extract the long-term interest distribution and short-term interest distribution of the target user group;

[0072] S102. Obtain a set of candidate content to be distributed in the private domain traffic platform, perform semantic analysis on the candidate content set to obtain content semantic features, calculate the similarity between the content in the candidate content set according to the content semantic features to obtain a content similarity matrix, construct a content propagation network based on the content similarity matrix, calculate the propagation coefficient of each content node in the content propagation network, calculate the matching coefficient between each content in the candidate content set and the target user group in combination with the long-term interest distribution and the short-term interest distribution, and generate a target content ranking table according to the propagation coefficient and the matching coefficient;

[0073] S103. The target user group is grouped according to the long-term interest distribution and the short-term interest distribution to obtain a user grouping result, a user activity prediction model is constructed according to the user grouping result, the user activity prediction model is used to calculate the active time period distribution of each group of users, a content delivery schedule is generated based on the active time period distribution and the target content ranking table, and the content delivery frequency of each user group is determined based on the matching coefficient, the content delivery schedule and the content delivery frequency are combined to generate a content delivery strategy, content delivery is executed on the private domain traffic platform according to the content delivery strategy, user interaction data on the delivered content is collected, delivery effect indicators are calculated according to the interaction data, the content delivery strategy is dynamically adjusted based on the delivery effect indicators, and content delivery is continued according to the adjusted content delivery strategy.

[0074] Among them, obtaining the historical behavior data of the target user group refers to collecting users' click, browse, like, comment, share and other behavior data from the private domain traffic platform, and preprocessing these raw data, including data cleaning, missing value filling, outlier processing and normalization operations, to obtain structured and effective behavior data.

[0075] The user-behavior association matrix is ​​a matrix constructed based on valid behavior data, in which rows represent users, columns represent behavior types, and the values ​​of matrix elements are usually the frequency or weight of a user's behavior of a certain type.

[0076] Long-term interest distribution and short-term interest distribution are the results of time dimension analysis based on user group portraits. Long-term interest distribution reflects users' long-term stable interest preferences, while short-term interest distribution reflects users' recent interest changes.

[0077] The candidate content set is a set of content to be distributed, including content in multimedia forms such as text, pictures, and videos.

[0078] Content semantic features are feature representations obtained after semantic analysis of a candidate content set, and are generated using methods such as word vectors, sentence vectors, or topic models.

[0079] In an optional implementation, constructing a user social network according to the user interaction weights, calculating user influence weights using a node importance algorithm in the user social network, and linearly combining the user-behavior association matrix with the user influence weights to obtain a user group portrait includes:

[0080] A user social network is constructed according to user interaction weights, including a multi-layer network structure, each layer of the network corresponds to an interaction type, and users are set as network nodes. The network edge weights are calculated according to a weighted combination of user interaction intimacy and user interest similarity, wherein the user interaction intimacy is obtained by performing time-attenuated weighted processing on the user interaction feature vector, and the user interest similarity is obtained by calculating the similarity of the topic vector of the user's historical interaction behavior sequence;

[0081] Calculating node importance indicators in the user social network, the node importance indicators include node degree centrality, node strength centrality, node betweenness centrality, node eigenvector centrality and node page ranking value;

[0082] An inter-layer influence propagation model is constructed based on the node importance index, and a random walk algorithm is used to calculate the influence propagation probability matrix between different layer networks, and the cross-layer influence diffusion coefficient is calculated in combination with the inter-layer correlation strength;

[0083] According to the influence propagation probability matrix and the cross-layer influence diffusion coefficient, an iterative algorithm is used to calculate the multi-layer network influence fusion function, and the influence weight of the user node is updated in each iteration to obtain a user influence weight vector;

[0084] Performing tensor decomposition on a pre-constructed user-behavior association matrix to obtain a user potential feature representation, wherein the user potential feature representation includes basic behavior features, temporal behavior features, and contextual behavior features;

[0085] The user-behavior association matrix is ​​linearly combined with the user influence weight vector, and the importance weights of different features are adaptively learned through the attention mechanism to finally obtain a user group portrait.

[0086] Exemplarily, first, a user social network is constructed. The network is not a single structure, but is divided into multiple hierarchical networks according to the type of user interaction. For example, the comment interaction between users constitutes a layer of network, the joint purchase of the same product constitutes another layer of network, and so on. Each user is a node in the network. The weight of the edge connecting the users is not a simple count, but a fusion of user interaction intimacy and interest similarity. User interaction intimacy considers the time factor of user interaction, and the recent interaction weight is higher, while the long-term interaction weight decays over time. User interest similarity is measured by analyzing the user's historical interaction behavior sequence, extracting the topic vector and calculating the similarity. For example, if two users often browse and purchase technology products, their interest similarity is high. Assume that user A and user B have three comment interactions in the past week and one joint purchase behavior in the past month. The time decay weight of comment interaction is 0.8, and the time decay weight of joint purchase is 0.5. The interest similarity between user A and user B in technology products is 0.9, and the interest similarity in daily necessities is 0.2. Then their edge weight in the comment interaction network layer is 3×0.8+0.9=3.3, and their edge weight in the co-purchase network layer is 1×0.5+0.2=0.7.

[0087] Next, the user's influence in the social network is calculated. In each layer of the network, multiple importance indicators of each node are calculated, including degree centrality of the number of connections, strength centrality of the sum of connection strengths, betweenness centrality as a bridge, eigenvector centrality based on the importance of neighboring nodes, and page ranking value that simulates the importance of web page links.

[0088] Then, an inter-layer influence propagation model is constructed. The random walk algorithm is used to simulate the propagation of user influence between different layers of the network. For example, a user has a high influence in the comment interaction network layer, and his influence will propagate to the co-purchase network layer with a certain probability. The propagation probability is determined by the network structure. In addition, it is also necessary to consider the strength of the association between the layers, such as the degree of association between comment interaction and co-purchase behavior, which is represented by the cross-layer influence diffusion coefficient. Assuming that the influence of user A in the comment interaction network layer is 0.8, the propagation probability from this layer to the co-purchase network layer is 0.6, and the cross-layer influence diffusion coefficient is 0.5, then the influence obtained by user A in the co-purchase network layer is 0.8×0.6×0.5=0.24.

[0089] The influence of different network layers is integrated through an iterative algorithm. In each iteration, the comprehensive influence weight of each user is updated according to the influence propagation probability matrix and the cross-layer influence diffusion coefficient, and finally the user influence weight vector is obtained.

[0090] At the same time, analyze user behavior data. Perform tensor decomposition on the user-behavior association matrix to extract user potential feature representations, including basic behavior features (such as the type of goods purchased), temporal behavior features (such as the pattern of purchase time), and contextual behavior features (such as the purchase location). For example, user C purchased a mobile phone, headphones, and a power bank. The purchase time is concentrated in the evening and the purchase location is mostly near the residence. This information constitutes the behavioral characteristics of user C.

[0091] Finally, the influence of the user's social network is integrated with the user's behavior data. The user-behavior association matrix is ​​linearly combined with the user's influence weight vector, and the attention mechanism is used to adaptively learn the importance weights of different features. For example, for the technology enthusiast group, the purchase behavior of technology products has a higher weight; for the social expert group, the social network influence weight is higher. Finally, the user group portrait is obtained.

[0092] In this embodiment, by integrating the influence of user social networks and multi-dimensional user behavior data, the user group characteristics can be more comprehensively portrayed, avoiding the limitations of a single data source, thereby improving the accuracy of the user group portrait. The method takes into account factors such as user interaction intimacy, interest similarity, and influence propagation, making the user group portrait more interpretable and easier to understand the formation mechanism of group characteristics. More accurate user group portraits can be applied to a wider range of scenarios, such as precision marketing, personalized recommendations, risk control, etc., thereby improving the application value of user group portraits.

[0093] In an optional implementation, constructing a content propagation network based on the content similarity matrix, calculating the propagation coefficient of each content node in the content propagation network, calculating the matching coefficient between each content in the candidate content set and the target user group in combination with the long-term interest distribution and the short-term interest distribution, and generating a target content ranking table according to the propagation coefficient and the matching coefficient includes:

[0094] Calculating the mean and standard deviation of the content similarity matrix, setting the sum of the mean and the standard deviation as a dynamic connection threshold, and constructing a content propagation network based on the dynamic connection threshold, wherein when the similarity in the content similarity matrix is ​​greater than the dynamic connection threshold, a connection edge is established between corresponding content nodes;

[0095] For each content node in the content dissemination network, the node degree value and the number of triangles associated with the node are calculated to obtain the node local connection feature, and the local propagation coefficient is calculated based on the node local connection feature; the number of shortest paths of the node and the number of shortest paths passing through the node are calculated to obtain the node global connection feature, and the global propagation coefficient is calculated based on the node global connection feature; the local propagation coefficient and the global propagation coefficient are adaptively weighted to obtain the node propagation coefficient;

[0096] Constructing a user cognitive feature vector, inputting the long-term interest distribution, the short-term interest distribution and the user cognitive feature vector into a self-attention network to obtain a user dynamic interest representation;

[0097] The semantic features of the candidate content are matched with the user's dynamic interest representation at the semantic level, cognitive level and context level, and the matching results of each level are weighted using an attention mechanism to obtain an initial matching score. The initial matching score is optimized in combination with the collaborative filtering characteristics of the user group to obtain a content matching coefficient;

[0098] The node propagation coefficient and the content matching coefficient are input into a pre-trained decision processing model to obtain a feature combination strategy, and an initial content ranking score is calculated according to the feature combination strategy; the initial content ranking score is diversifiedly adjusted in combination with content timeliness and topic popularity to generate a target content ranking table.

[0099] For example, first, a content similarity matrix is ​​calculated. Assume that there are five contents, represented by A, B, C, D, and E. The similarity between them is calculated two by two, for example, using cosine similarity, to obtain a 5x5 matrix.

[0100] Next, calculate the mean and standard deviation of the similarity matrix. Taking the above matrix as an example, the mean is 0.46 and the standard deviation is 0.28. Add the mean and standard deviation to get the dynamic connection threshold, which is 0.74.

[0101] The content dissemination network is constructed based on the dynamic connection threshold. When the similarity of two contents in the content similarity matrix is ​​greater than the dynamic connection threshold, a connection is established between the two contents. For example, if the similarity between A and B is 0.8 and greater than 0.74, a connection is established between A and B. The resulting content dissemination network can be represented in the form of a graph, where nodes represent content and edges represent connections.

[0102] Then, the propagation coefficient of each content node in the content propagation network is calculated. For each node, its degree value (the number of nodes directly connected to it) and the number of associated triangles (the number of mutual connections between nodes directly connected to it) are first calculated. For example, the degree value of node A is 1 (only connected to B) and the number of associated triangles is 0. These features constitute the local connection features of the node. The local propagation coefficient is calculated based on these local connection features. At the same time, the number of shortest paths to the node (the number of shortest paths from all other nodes in the network to the node) and the number of shortest paths passing through the node (the number of shortest paths between all other nodes in the network that pass through the node) are calculated. These features constitute the global connection features of the node. The global propagation coefficient is calculated based on these global connection features. Finally, the local propagation coefficient and the global propagation coefficient are adaptively weighted to obtain the node propagation coefficient.

[0103] Next, construct a user cognitive feature vector. For example, the user's age, gender, occupation, education level and other information can be used as user cognitive features. Then, the user's long-term interest distribution (for example, the user's long-term interest in sports, music, movies and other content), short-term interest distribution (for example, the user's interest in a specific topic or event in the past week) and user cognitive feature vector are input into the self-attention network to obtain the user's dynamic interest representation. The dynamic interest representation captures the weight changes of interest features in different time periods and situations by inputting interest features into the self-attention network, and dynamically reflects the user's current interest status.

[0104] After that, the content semantic features of the candidate content are matched with the user's dynamic interest representation. The matching process is divided into semantic layer matching (matching of content theme with user interest), cognitive layer matching (matching of content difficulty with user cognitive level) and contextual layer matching (matching of content time with user's current context). The attention mechanism is used to weight the matching results of each layer to obtain the initial matching score. Combined with the collaborative filtering characteristics of the user group (such as the ratings of the content by other users with similar interests to the user), the initial matching score is optimized to obtain the content matching coefficient.

[0105] Finally, the node propagation coefficient and content matching coefficient are input into the pre-trained decision processing model to obtain the feature combination strategy. The initial content ranking score is calculated based on the feature combination strategy. The initial content ranking score is adjusted in a diversified manner by combining the content timeliness (e.g., the release time of the news) and the topic popularity (e.g., the discussion degree of the current hot topic), and finally the target content ranking table is generated.

[0106] In this embodiment, by considering the correlation between content to construct a content dissemination network and calculating the node dissemination coefficient, content with high dissemination potential can be effectively identified, thereby improving content dissemination efficiency. By combining the user's long-term interests, short-term interests and cognitive characteristics, constructing a user's dynamic interest representation, and performing multi-level matching, the user's interests can be captured more accurately, thereby improving the accuracy of content recommendations. By combining the content dissemination coefficient and the user interest matching coefficient, and considering factors such as content timeliness and topic popularity, better content ranking results can be generated to improve user experience.

[0107] In an optional implementation, the node propagation coefficient and the content matching coefficient are input into a pre-trained decision processing model to obtain a feature combination strategy, and an initial content ranking score is calculated according to the feature combination strategy; the initial content ranking score is adjusted in a diversified manner in combination with content timeliness and topic popularity, and generating a target content ranking table includes:

[0108] Construct a decision processing model, receive the node propagation coefficient and the content matching coefficient as input features, map the input features into a state vector, map the feature combination mode into an action vector, construct a state transition matrix based on real-time user feedback data, and calculate the ranking conversion rate according to the state transition matrix to obtain a reward vector;

[0109] Constructing a master control network and a mirror network, wherein the master control network receives a state vector and an action vector to generate a first state action value, and the mirror network receives a state vector and an action vector to generate a second state action value;

[0110] Calculate the time difference between the first state action value and the second state action value, construct a sample weight matrix based on the time difference, perform priority sampling on the training data according to the sample weight matrix to obtain experience playback data, input the experience playback data into the master control network and the mirror network respectively, and obtain a current state prediction value and a target state prediction value;

[0111] Performing time series sampling on the experience replay data to obtain a multi-step sampling sequence, calculating a target value function based on the multi-step sampling sequence and a target state prediction value, calculating a sampling calibration value according to a difference between a behavior strategy and a target strategy, and constructing a loss function based on the target value function and the sampling calibration value;

[0112] Calculating the parameter gradient of the master network according to the loss function, updating the network parameters of the master network based on the parameter gradient, copying the updated network parameters to the mirror network, and outputting a feature combination strategy based on the updated master network and the mirror network;

[0113] The feature combination strategy is multiplied by the input feature to obtain a content basic score, the content release time is obtained to calculate the time decay value, the topic heat value is calculated based on the topic propagation rate and the amount of interaction, and the content basic score is weightedly combined with the time decay value and the topic heat value to obtain a content ranking score;

[0114] The similarities between the to-be-sorted contents are calculated to obtain a content duplication matrix, the content sorting scores are diversity-adjusted based on the content duplication matrix to obtain a final sorting score, and a target sorting list is generated according to the final sorting score.

[0115] Exemplarily, first, a decision processing model is constructed. The input of the model is the node propagation coefficient and the content matching coefficient. The node propagation coefficient can be understood as the scope and speed of content propagation in the social network, such as the number of forwarding times, the number of comments, etc. The content matching coefficient represents the degree of matching between the content and the user's interests, which can be calculated by the user's historical behavior data, such as clicks, favorites, likes, etc. These input features are converted into a state vector, representing the state of the current sorting strategy. At the same time, different feature combination methods, such as weighted summation, product, etc., are mapped to different action vectors. A state transition matrix is ​​constructed using a large amount of real-time user feedback data, such as click-through rate, dwell time, etc. This matrix describes the conversion rate of the sorting results after taking different actions under different states. According to the state transition matrix, the sorting conversion rate brought by different state-action combinations is calculated to obtain a reward vector. The larger the value of the reward vector, the better the sorting strategy.

[0116] Next, build the master network and the mirror network. The two networks have the same structure but different parameters. Both the master network and the mirror network receive the state vector and action vector as input, and output the first state action value and the second state action value respectively. These two values ​​represent the expected benefits of taking a certain action in the current state.

[0117] Then, the time difference between the action value of the first state and the action value of the second state is calculated. This difference reflects the gap between the prediction of the master network and the actual benefit. The sample weight matrix is ​​constructed based on the time difference. The larger the weight of the sample, the greater the impact on the model training. The training data is sampled according to the sample weight matrix to obtain the experience replay data. These experience replay data are input into the master network and the mirror network respectively to obtain the current state prediction value and the target state prediction value.

[0118] Perform time series sampling on the experience replay data to obtain a multi-step sampling sequence. Calculate the target value function based on the multi-step sampling sequence and the target state prediction value. The target value function represents the cumulative benefit that can be obtained after taking a series of actions in the current state. Calculate the sampling calibration value based on the difference between the behavioral strategy and the target strategy. The sampling calibration value is used to correct the deviation caused by the difference in strategy. Construct the loss function based on the target value function and the sampling calibration value.

[0119] The parameter gradient of the master network is calculated according to the loss function, and the network parameters of the master network are updated based on the parameter gradient. The updated network parameters are copied to the mirror network. The feature combination strategy is output based on the updated master network and the mirror network. For example, the strategy finally learned by the master network may be to multiply the node propagation coefficient by 0.8 and add the content matching coefficient by 0.2.

[0120] Apply the learned feature combination strategy to the input features to obtain the content basic score. For example, if the node propagation coefficient of a certain content is 10 and the content matching coefficient is 0.9, then according to the above strategy, its basic score is 10×0.8+0.9×0.2=8.18. Get the content release time and calculate the time decay value. For example, the older the content is released, the smaller the time decay value. Calculate the topic heat value based on the topic propagation rate and interaction volume. For example, the faster the topic propagates and the greater the interaction volume, the higher the topic heat value. Combine the content basic score with the time decay value and the topic heat value to obtain the content ranking score. For example, assuming the time decay value is 0.9 and the topic heat value is 1.2, and the weights are 0.1 and 0.2 respectively, the final ranking score is 8.18+0.9×0.1+1.2×0.2=8.51.

[0121] Calculate the similarity between the contents to be sorted and obtain the content duplication matrix. For example, if the topics and keywords of two contents are highly similar, their corresponding element values ​​in the content duplication matrix will be larger. Based on the content duplication matrix, adjust the content sorting scores for diversity and obtain the final sorting scores. For example, reduce the sorting scores of duplicate contents to ensure the diversity of sorting results. Generate the target sorting list based on the final sorting scores.

[0122] Suppose there are two pieces of content A and B. The basic content score of A is 9, the time decay value is 0.9, and the topic heat value is 1.1; the basic content score of B is 8, the time decay value is 0.8, and the topic heat value is 1.2. Assume that the weights of the time decay value and the topic heat value are 0.1 and 0.2 respectively. Then the ranking score of A is 9+0.9×0.1+1.1×0.2=9.31, and the ranking score of B is 8+0.8×0.1+1.2×0.2=8.32. Assume that the similarity between A and B is high, resulting in a diversity adjustment coefficient of 0.9 for A and a diversity adjustment coefficient of 0.8 for B. Then the final ranking score of A is 9.31×0.9=8.379, and the final ranking score of B is 8.32×0.8=6.656. Therefore, the final ranking list is that A is ranked before B.

[0123] In this embodiment, the node propagation coefficient, content matching coefficient, time decay value, topic heat value and similarity between contents are comprehensively considered, which can more accurately evaluate the value of the content, thereby generating a sorted list that better meets user needs. By introducing the content duplication matrix, the sorting scores of similar content are effectively reduced, the diversity of sorting results is ensured, the information cocoon effect is avoided, and the user experience is improved. This method uses reinforcement learning technology to dynamically adjust the sorting strategy according to real-time user feedback data, realize personalized sorting, and can continuously learn and optimize to adapt to changing user needs.

[0124] In an optional implementation, a formula for constructing a loss function based on the target value function and the sampled calibration value is as follows:

[0125] ;

[0126] Among them, L represents the loss function, N represents the total number of samples, and W ij represents the sample weight matrix, n represents the number of steps in the sample sequence, k represents the current time step, and a k represents the content recommended by the recommendation system at time step k, s k represents the state of the environment at time step k, including user behavior features and content features, Indicates behavioral strategy, represents the target strategy, K represents the maximum number of steps for multi-step sampling, γ represents the discount factor, and r k Indicates that at time step k, due to taking action a k The reward obtained is Q 2 (s K+1 , a K+1 ) represents the future state s predicted by the mirror network K+1 Take action a K+1 The value of Q 1( s i , ai ) represents the current state s predicted by the master network i Take action a i The value of .

[0127] In order to build a recommendation system that can learn based on the target value function and the sampled calibration value, a new loss function construction method is proposed. The core idea of ​​this method is to use multi-step sampling technology and mirror network to more accurately estimate future rewards, and combine sample weights to adjust the contribution of different samples to the loss function.

[0128] First, collect user behavior data and content data to build the environment state. The environment state includes user behavior features, such as clicks, favorites, purchases, etc., and content features, such as titles, descriptions, tags, etc. For example, user A clicks on movie B, and movie B is a science fiction film with a duration of 120 minutes. Then the current environment state includes the behavior of user A clicking on movie B, as well as information such as the type and duration of movie B.

[0129] Next, based on the current state of the environment, a behavior strategy is used to select a recommended content. A behavior strategy can be a rule-based strategy, such as recommending similar content based on the user's historical behavior, or a model-based strategy, such as using a deep learning model to predict the user's interest in different content. Suppose the behavior strategy selects movie C to recommend to user A.

[0130] Then, the multi-step sampling technique is used to simulate the user's behavior in multiple future time steps. For example, assume that the maximum number of sampling steps is 3 and the discount factor is 0.9. In the first step, user A watches the recommended movie C and receives a reward, such as user A's rating of movie C. In the second step, based on the updated environment state (including user A's behavior of watching movie C), the behavior strategy is used again to select a recommendation and receive the corresponding reward. This process is repeated until the maximum number of sampling steps is reached.

[0131] In order to estimate future rewards more accurately, a mirror network is used to predict the value of taking different actions in future states. The structure and parameters of the mirror network are the same as those of the main control network, but different data sets are used for training. The main control network is trained with current data, while the mirror network is trained with past data. For example, the main control network is trained with the behavior data of user A in the last month, while the mirror network is trained with the behavior data of user A in the previous month.

[0132] The target value function is calculated using the collected multi-step sampling data and the prediction results of the mirror network. The target value function is calculated by weighted summing the rewards of multiple future time steps, with the weights determined by the discount factor. For example, the reward of the first step is multiplied by 0.9, the reward of the second step is multiplied by 0.9 squared, and so on.

[0133] At the same time, according to the target strategy and the current state, the sampling calibration value is calculated. The target strategy is an ideal recommendation strategy that can be designed according to the actual application scenario. For example, the target strategy can be a strategy that maximizes the long-term satisfaction of users. The sampling calibration value is used to measure the gap between the behavior strategy and the target strategy.

[0134] Finally, the target value function and the sampled calibration value are substituted into the loss function for calculation. The loss function is calculated by weighted summing the square of the difference between the target value function and the sampled calibration value, where the weight is determined by the sample weight matrix. The sample weight matrix is ​​used to adjust the contribution of different samples to the loss function. For example, higher weights can be given to important samples.

[0135] By minimizing the loss function, the parameters of the recommendation system can be continuously optimized to make it closer to the target strategy.

[0136] In this embodiment, through multi-step sampling and mirror network, future rewards can be estimated more accurately, thereby improving the accuracy of the recommendation system. The sample weight matrix can balance the contribution of different samples to the loss function, avoid overfitting of certain samples, and thus enhance the stability of the model. By combining the target strategy and sampling calibration value, the model learning can be guided more effectively, thereby improving the learning efficiency.

[0137] In an optional implementation, constructing a user activity prediction model according to the user grouping result, using the user activity prediction model to calculate the active time period distribution of each group of users, and generating a content delivery schedule based on the active time period distribution and the target content ranking table includes:

[0138] Based on the user grouping results, an initial causal knowledge graph including user group entity nodes, user behavior entity nodes and environmental factor entity nodes is constructed, the causal relationship strength between each entity node is calculated using a structural causal model, a relationship strength matrix is ​​generated based on the causal relationship strength, and the initial causal knowledge graph and the relationship strength matrix are combined to form a time-varying causal knowledge graph;

[0139] Collecting the newly added node-to-node relationship edges of the time-varying causal knowledge graph within a preset time period, inputting the newly added node-to-node relationship edges into a recurrent neural network to obtain dynamic change characteristics of the node-to-node relationship, and updating the relationship strength matrix based on the dynamic change characteristics of the node-to-node relationship to obtain an updated relationship strength matrix;

[0140] Constructing a hierarchical user activity prediction model according to the user grouping result, wherein the hierarchical user activity prediction model includes a causal convolutional network layer for extracting active mode features of user groups, a multi-head causal attention network layer for calculating temporal influence features between user groups, and a neural ordinary differential equation layer for obtaining continuous dynamic features of user activity;

[0141] Input the updated relationship strength matrix into the hierarchical user activity prediction model to obtain user group activity mode features, temporal influence features between user groups and continuous dynamic features of user activity, input the user group activity mode features, temporal influence features between user groups and continuous dynamic features of user activity into the causal fusion network for weighted fusion, and introduce causal bias values ​​to obtain user activity prediction results;

[0142] A Bayesian causal network is established based on the user activity prediction result, a variational inference method is used to calculate the probability distribution of content delivery time, an initial content delivery time point is determined through importance sampling according to the content delivery time probability distribution, and a delivery effect deviation is calculated based on a counterfactual analysis method, a delivery time adjustment window is determined according to the delivery effect deviation, an online convex optimization algorithm is used to calculate a delivery parameter adjustment value, and the initial content delivery time point is optimized according to the delivery parameter adjustment value;

[0143] Multi-dimensional consistency constraint processing is performed on the user activity prediction result and the optimized content delivery time point to obtain a final content delivery schedule that satisfies the multi-dimensional consistency constraint.

[0144] Exemplarily, first, a causal knowledge graph of user behavior is constructed. Taking the e-commerce platform as an example, users are divided into groups such as new users, loyal users, and lost users. The causal knowledge graph contains entity nodes such as user groups (such as new users), user behaviors (such as browsing products, adding to shopping carts, and placing orders), and environmental factors (such as holidays and promotions). Using expert knowledge or data-driven methods, the causal relationship between entity nodes is judged and an initial causal relationship strength value is assigned. For example, "promotional activities" have a strong positive causal relationship with "new user orders", and the strength value is set to 0.8. These causal relationship strength values ​​are combined into a relationship strength matrix. The initial causal knowledge graph and the relationship strength matrix together constitute a time-varying causal knowledge graph.

[0145] Next, capture the dynamic changes of the causal knowledge graph. For example, if a new association is observed between "holidays" and "loyal users browsing products" during a certain period of time, this association is used as a new node-to-node relationship edge. These new node-to-node relationship edges are input into a recurrent neural network (such as an LSTM network), and the network learns the dynamic change characteristics of these new relationship edges. Based on the learned features, update the relationship strength matrix. For example, if it is found that the impact of holidays on loyal users browsing products is gradually increasing, the corresponding value in the relationship strength matrix is ​​increased accordingly.

[0146] Then, a hierarchical user activity prediction model is constructed. The model consists of three layers: the causal convolutional network layer extracts the active pattern characteristics of user groups, for example, the active pattern of new user groups may be concentrated at night; the multi-head causal attention network layer calculates the temporal influence characteristics between user groups, for example, the activeness of loyal user groups may drive the activeness of new user groups; the neural ordinary differential equation layer obtains the continuous dynamic characteristics of user activity, such as the smooth change trend of user activity within a day.

[0147] The updated relationship strength matrix is ​​input into the hierarchical user activity prediction model to obtain the user group activity mode features, the time series influence features between user groups, and the continuous dynamic features of user activity. These three features are input into the causal fusion network, and the causal bias value is introduced (for example, according to expert experience, the impact of promotional activities on user activity is more important), and weighted fusion is performed to finally obtain the user activity prediction results, such as predicting the activity probability of new users in the next week.

[0148] Based on the user activity prediction results, a Bayesian causal network is established, and the variational inference method is used to calculate the probability distribution of content delivery time. For example, the probability of delivering content to new users at 9 am is calculated to be 0.7, and the probability of delivering content at 2 pm is 0.3. According to the probability distribution of content delivery time, the initial content delivery time point is determined by importance sampling, for example, 9 am is selected as the initial delivery time point. Then, based on the counterfactual analysis method, the delivery effect of other delivery time points (such as 2 pm) is simulated, and the effect deviation from the initial delivery time point is calculated. According to the delivery effect deviation, the delivery time adjustment window is determined, for example, plus or minus 1 hour. Finally, the online convex optimization algorithm is used to calculate the delivery parameter adjustment value, for example, to advance the delivery time by 30 minutes. The initial delivery time point is optimized according to the calculated adjustment value to obtain the final delivery time point, for example, 8:30 am.

[0149] Finally, multi-dimensional consistency constraints are applied to the user activity prediction results and the optimized content delivery time points. For example, constraints such as content delivery budget and target user coverage are considered to ensure that the final generated content delivery schedule meets multi-dimensional consistency requirements.

[0150] Assume that during the "Double Eleven" promotion, the e-commerce platform predicts that the probability of new users being active at 9 am on November 11th is the highest. Through counterfactual analysis, it is found that if the delivery time is adjusted to 8 am, the delivery effect deviation is positive, indicating that the effect of early delivery is better. Considering the budget constraint and user coverage target, it is finally decided to deliver promotional information to new users at 8:30 am on November 11th.

[0151] In this embodiment, by constructing a time-varying causal knowledge graph and a hierarchical user activity prediction model, user activity can be predicted more accurately, thereby improving the accuracy of content delivery. Through the Bayesian causal network and online convex optimization algorithm, the content delivery time can be optimized to maximize the delivery effect, such as improving click-through rate, conversion rate, etc. Automatically generate a content delivery schedule, reduce manual intervention, improve decision-making efficiency, and reduce operating costs.

[0152] In an optional implementation, calculating the delivery effect deviation based on a counterfactual analysis method, determining the delivery time adjustment window according to the delivery effect deviation, calculating the delivery parameter adjustment value using an online convex optimization algorithm, and optimizing the initial content delivery time point according to the delivery parameter adjustment value includes:

[0153] Obtaining the initial content delivery time point, constructing a counterfactual evaluation model, the counterfactual evaluation model comprising a causal intervention operation module and a random sampling module, calculating the user response expectation value before and after the delivery time adjustment through the causal intervention operation module, and performing importance sampling on the user response expectation value through the random sampling module to obtain a counterfactual delivery effect evaluation value;

[0154] A time window adjustment function is constructed based on the counterfactual delivery effect evaluation value, wherein the time window adjustment function calculates the adaptive parameters of the time window boundary according to the effect evaluation deviation value and the effect fluctuation measurement value, and dynamically updates the time window boundary using the gradient descent method to obtain an optimized delivery time window;

[0155] Inputting the optimized delivery time window into an online convex optimization model, wherein the online convex optimization model includes a delivery effect loss function and a regularization constraint term, and adopting an adaptive subgradient descent algorithm with momentum term to iteratively optimize the delivery parameters to obtain optimized delivery parameters;

[0156] Constructing a parameter adjustment feedback model based on the optimized delivery parameters, wherein the parameter adjustment feedback model extracts candidate parameters from the posterior distribution of parameters using a probability sampling method, updates the parameter distribution in real time through a state estimation filter, and outputs the optimal delivery parameters;

[0157] The initial content delivery time point is adjusted according to the optimal delivery parameters to obtain an optimized content delivery time point.

[0158] Exemplarily, after obtaining the initial content delivery time point, it is necessary to evaluate the potential impact of delivery time adjustment on user response. To this end, it is necessary to first build a counterfactual evaluation model, which includes a causal intervention operation module and a random sampling module. The causal intervention operation module calculates the difference in user response expectations before and after the adjustment by simulating changes in intervention conditions, such as adjusting the delivery time to different time periods. This difference reflects the potential impact of time adjustment on user behavior. On this basis, the random sampling module performs importance sampling on user response expectations, thereby evaluating the possible results of delivery in different time periods, and finally obtaining a counterfactual delivery effect evaluation value.

[0159] Based on the counterfactual delivery effect evaluation value, a time window adjustment function needs to be constructed. This function is used to define the boundary range of the delivery time window, and dynamically adjust these boundaries by combining the deviation of the delivery effect and the fluctuation range of the effect. The adjustment process can be achieved through an iterative method, such as gradually adjusting the range before and after the delivery time until an optimized time window is found. The boundaries of the time window will be updated based on dynamic feedback. This process achieves dynamic convergence of the time window through iterative optimization techniques, such as a step-by-step adjustment strategy that simulates decreasing returns.

[0160] The optimized time window is used as input to further build an online optimization model. The goal of the online optimization model is to iteratively optimize the delivery parameters according to the dynamically adjusted time window to ensure the flexibility of the delivery plan in actual operation. The model consists of a delivery effect loss evaluation method and regularization constraints. The delivery effect loss method is used to measure the gap between the content delivered at the current time and the user response, and the regularization constraints ensure that the optimization process remains stable and avoids over-adjustment. Through repeated adjustment and optimization, the optimized delivery parameters are finally output.

[0161] Based on the optimized delivery parameters, a parameter adjustment feedback model needs to be established. The feedback model uses a probabilistic sampling method to extract the posterior distribution of delivery parameters from historical data as samples to simulate possible delivery parameter combinations. Subsequently, these candidate parameters are updated in real time through the state estimation filter to select the optimal delivery parameters. The function of the state estimation filter is to dynamically track changes in delivery effects and adjust the parameter distribution based on real-time feedback to ensure that the selected parameters can adapt to the current user behavior pattern.

[0162] Adjust the initial content delivery time point based on the optimal delivery parameters. The principle of adjustment is to set the content delivery time in the period when users are most likely to respond based on the optimized time window and parameters, while avoiding overly intensive or repeated delivery strategies. The final optimized content delivery time point can maximize the effect of user response while maintaining the overall efficiency of the delivery plan.

[0163] Suppose a content platform plans to publish a technology article, and the initial publishing time is set to 10 a.m. Through the counterfactual evaluation model, the user response when the publishing time is adjusted to other time periods (such as 3 p.m. and 8 p.m.) is simulated, and it is found that the user click rate and reading completion rate at 3 p.m. are significantly higher than those at the initial time.

[0164] We further constructed a time window adjustment function and determined the delivery time window to be from 2pm to 4pm by analyzing the fluctuation of delivery effects. Based on this window, the online optimization model evaluated the dynamic changes of delivery parameters, such as selecting 3pm as the delivery time point and slightly adjusting the push frequency. The parameter adjustment feedback model screened out the optimal display frequency and method of delivery content through user response information extracted from historical data.

[0165] The final adjusted delivery time point was 3 pm, and a repeated push schedule was set up every three hours to ensure that the target user group could be exposed to the content multiple times. The optimized delivery strategy was continuously fine-tuned by real-time monitoring of user behavior data, such as click-through rate, number of likes and number of comments, further improving the dissemination effect of the content.

[0166] In this embodiment, dynamic adjustment and optimization of content delivery strategies are achieved through counterfactual analysis and online optimization methods. By evaluating potential changes in user responses through causal intervention, the impact of different delivery times on user behavior is accurately identified, thereby improving the accuracy of content delivery. The combination of time window adjustment and parameter optimization methods enables flexible and dynamic control of delivery periods and frequencies, effectively reducing the risk of user churn due to fixed delivery strategies. The delivery strategy is updated in real time through the parameter adjustment feedback model, so that the system can adapt to changes in user behavior and needs, and improve the efficiency and effectiveness of content dissemination. The overall solution optimizes user response rate and interaction rate, enhances the match between content and users, and at the same time improves the utilization efficiency of delivery resources, providing strong support for the intelligent operation of private domain traffic platforms.

[0167] Figure 2 FIG. 1 is a structural diagram of a private domain traffic scheduling and content distribution system based on group behavior analysis according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0168] The first unit is used to obtain historical behavior data of a target user group in a private domain traffic platform and perform preprocessing to obtain effective behavior data, construct a user-behavior association matrix based on the effective behavior data, obtain user interaction weights by calculating the interaction frequency and interaction depth between each user node in the user-behavior association matrix, construct a user social network based on the user interaction weights, calculate user influence weights using a node importance algorithm in the user social network, linearly combine the user-behavior association matrix with the user influence weights to obtain a user group portrait, perform time dimension analysis on the user group portrait, and extract long-term interest distribution and short-term interest distribution of the target user group;

[0169] The second unit is used to obtain a set of candidate content to be distributed in the private domain traffic platform, perform semantic analysis on the candidate content set to obtain content semantic features, calculate the similarity between the contents in the candidate content set according to the content semantic features to obtain a content similarity matrix, build a content propagation network based on the content similarity matrix, calculate the propagation coefficient of each content node in the content propagation network, calculate the matching coefficient between each content in the candidate content set and the target user group in combination with the long-term interest distribution and the short-term interest distribution, and generate a target content ranking table according to the propagation coefficient and the matching coefficient;

[0170] The third unit is used to group the target user group according to the long-term interest distribution and the short-term interest distribution to obtain user grouping results, build a user activity prediction model according to the user grouping results, use the user activity prediction model to calculate the active time period distribution of each group of users, generate a content delivery schedule based on the active time period distribution and the target content ranking table, and determine the content delivery frequency of each user group based on the matching coefficient, combine the content delivery schedule and the content delivery frequency to generate a content delivery strategy, execute content delivery in the private domain traffic platform according to the content delivery strategy, collect user interaction data on the delivered content, calculate delivery effect indicators according to the interaction data, dynamically adjust the content delivery strategy based on the delivery effect indicators, and continue to execute content delivery according to the adjusted content delivery strategy.

[0171] According to a third aspect of the embodiments of the present invention,

[0172] An electronic device is provided, comprising:

[0173] processor;

[0174] a memory for storing processor-executable instructions;

[0175] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0176] A fourth aspect of the embodiments of the present invention is:

[0177] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0178] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A private domain traffic scheduling and content distribution method based on group behavior analysis, characterized in that: include: Obtain historical behavior data of the target user group in the private domain traffic platform and perform preprocessing to obtain effective behavior data, construct a user-behavior association matrix based on the effective behavior data, obtain user interaction weights by calculating the interaction frequency and interaction depth between each user node in the user-behavior association matrix, construct a user social network based on the user interaction weights, calculate user influence weights using a node importance algorithm in the user social network, linearly combine the user-behavior association matrix with the user influence weights to obtain a user group portrait, perform time dimension analysis on the user group portrait, and extract the long-term interest distribution and short-term interest distribution of the target user group; A set of candidate content to be distributed in a private domain traffic platform is obtained, semantic analysis is performed on the candidate content set to obtain content semantic features, similarities between contents in the candidate content set are calculated according to the content semantic features to obtain a content similarity matrix, a content propagation network is constructed based on the content similarity matrix, a propagation coefficient of each content node in the content propagation network is calculated, a matching coefficient between each content in the candidate content set and the target user group is calculated in combination with the long-term interest distribution and the short-term interest distribution, and a target content ranking table is generated according to the propagation coefficient and the matching coefficient; The target user group is grouped according to the long-term interest distribution and the short-term interest distribution to obtain a user grouping result, a user activity prediction model is constructed according to the user grouping result, the user activity prediction model is used to calculate the active time period distribution of each group of users, a content delivery schedule is generated based on the active time period distribution and the target content ranking table, and the content delivery frequency of each user group is determined based on the matching coefficient, the content delivery schedule and the content delivery frequency are combined to generate a content delivery strategy, content delivery is executed on the private domain traffic platform according to the content delivery strategy, user interaction data on the delivered content is collected, delivery effect indicators are calculated according to the interaction data, the content delivery strategy is dynamically adjusted based on the delivery effect indicators, and content delivery is continued according to the adjusted content delivery strategy.

2. The method according to claim 1, characterized in that Constructing a user social network according to the user interaction weights, calculating user influence weights using a node importance algorithm in the user social network, and linearly combining the user-behavior association matrix with the user influence weights to obtain a user group portrait includes: A user social network is constructed according to user interaction weights, including a multi-layer network structure, each layer of the network corresponds to an interaction type, and users are set as network nodes. The network edge weights are calculated according to a weighted combination of user interaction intimacy and user interest similarity, wherein the user interaction intimacy is obtained by performing time-attenuated weighted processing on the user interaction feature vector, and the user interest similarity is obtained by calculating the similarity of the topic vector of the user's historical interaction behavior sequence; Calculating node importance indicators in the user social network, the node importance indicators include node degree centrality, node strength centrality, node betweenness centrality, node eigenvector centrality and node page ranking value; An inter-layer influence propagation model is constructed based on the node importance index, and a random walk algorithm is used to calculate the influence propagation probability matrix between different layer networks, and the cross-layer influence diffusion coefficient is calculated in combination with the inter-layer correlation strength; According to the influence propagation probability matrix and the cross-layer influence diffusion coefficient, an iterative algorithm is used to calculate the multi-layer network influence fusion function, and the influence weight of the user node is updated in each iteration to obtain a user influence weight vector; Performing tensor decomposition on a pre-constructed user-behavior association matrix to obtain a user potential feature representation, wherein the user potential feature representation includes basic behavior features, temporal behavior features, and contextual behavior features; The user-behavior association matrix is ​​linearly combined with the user influence weight vector, and the importance weights of different features are adaptively learned through the attention mechanism to finally obtain a user group portrait.

3. The method according to claim 1, characterized in that Constructing a content propagation network based on the content similarity matrix, calculating the propagation coefficient of each content node in the content propagation network, calculating the matching coefficient between each content in the candidate content set and the target user group in combination with the long-term interest distribution and the short-term interest distribution, and generating a target content ranking table according to the propagation coefficient and the matching coefficient includes: Calculating the mean and standard deviation of the content similarity matrix, setting the sum of the mean and the standard deviation as a dynamic connection threshold, and constructing a content propagation network based on the dynamic connection threshold, wherein when the similarity in the content similarity matrix is ​​greater than the dynamic connection threshold, a connection edge is established between corresponding content nodes; For each content node in the content dissemination network, the node degree value and the number of triangles associated with the node are calculated to obtain the node local connection feature, and the local propagation coefficient is calculated based on the node local connection feature; the number of shortest paths of the node and the number of shortest paths passing through the node are calculated to obtain the node global connection feature, and the global propagation coefficient is calculated based on the node global connection feature; the local propagation coefficient and the global propagation coefficient are adaptively weighted to obtain the node propagation coefficient; Constructing a user cognitive feature vector, inputting the long-term interest distribution, the short-term interest distribution and the user cognitive feature vector into a self-attention network to obtain a user dynamic interest representation; The semantic features of the candidate content are matched with the user's dynamic interest representation at the semantic level, cognitive level and context level, and the matching results of each level are weighted using an attention mechanism to obtain an initial matching score. The initial matching score is optimized in combination with the collaborative filtering characteristics of the user group to obtain a content matching coefficient; The node propagation coefficient and the content matching coefficient are input into a pre-trained decision processing model to obtain a feature combination strategy, and an initial content ranking score is calculated according to the feature combination strategy; the initial content ranking score is diversifiedly adjusted in combination with content timeliness and topic popularity to generate a target content ranking table.

4. The method according to claim 3, characterized in that Inputting the node propagation coefficient and the content matching coefficient into a pre-trained decision processing model to obtain a feature combination strategy, and calculating an initial content ranking score according to the feature combination strategy; The initial scores of the content ranking are adjusted in a diversified manner in combination with the content timeliness and topic popularity to generate a target content ranking table including: Construct a decision processing model, receive the node propagation coefficient and the content matching coefficient as input features, map the input features into a state vector, map the feature combination mode into an action vector, construct a state transition matrix based on real-time user feedback data, and calculate the ranking conversion rate according to the state transition matrix to obtain a reward vector; Constructing a master control network and a mirror network, wherein the master control network receives a state vector and an action vector to generate a first state action value, and the mirror network receives a state vector and an action vector to generate a second state action value; Calculate the time difference between the first state action value and the second state action value, construct a sample weight matrix based on the time difference, perform priority sampling on the training data according to the sample weight matrix to obtain experience playback data, input the experience playback data into the master control network and the mirror network respectively, and obtain a current state prediction value and a target state prediction value; Performing time series sampling on the experience replay data to obtain a multi-step sampling sequence, calculating a target value function based on the multi-step sampling sequence and a target state prediction value, calculating a sampling calibration value according to a difference between a behavior strategy and a target strategy, and constructing a loss function based on the target value function and the sampling calibration value; Calculating the parameter gradient of the master network according to the loss function, updating the network parameters of the master network based on the parameter gradient, copying the updated network parameters to the mirror network, and outputting a feature combination strategy based on the updated master network and the mirror network; The feature combination strategy is multiplied by the input feature to obtain a content basic score, the content release time is obtained to calculate the time decay value, the topic heat value is calculated based on the topic propagation rate and the amount of interaction, and the content basic score is weightedly combined with the time decay value and the topic heat value to obtain a content ranking score; The similarities between the to-be-sorted contents are calculated to obtain a content duplication matrix, the content sorting scores are diversity-adjusted based on the content duplication matrix to obtain a final sorting score, and a target sorting list is generated according to the final sorting score.

5. The method according to claim 4, characterized in that The formula for constructing the loss function based on the target value function and the sampled calibration value is as follows: ； Among them, L represents the loss function, N represents the total number of samples, and W ij represents the sample weight matrix, n represents the number of steps in the sample sequence, k represents the current time step, a k represents the content recommended by the recommendation system at time step k, s k represents the state of the environment at time step k, including user behavior features and content features, Indicates behavioral strategy, represents the target strategy, K represents the maximum number of steps for multi-step sampling, γ represents the discount factor, and r k Indicates that at time step k, due to taking action a k The reward obtained, Q2(s K+1 , a K+1 ) represents the future state s predicted by the mirror network K+1 Take action a K+1 The value of Q 1( s i , a i ) represents the current state s predicted by the master network i Take action a i The value of .

6. The method according to claim 1, characterized in that Constructing a user activity prediction model according to the user grouping result, calculating the active time period distribution of each group of users by using the user activity prediction model, and generating a content delivery schedule based on the active time period distribution and the target content ranking table includes: Based on the user grouping results, an initial causal knowledge graph including user group entity nodes, user behavior entity nodes and environmental factor entity nodes is constructed, the causal relationship strength between each entity node is calculated using a structural causal model, a relationship strength matrix is ​​generated based on the causal relationship strength, and the initial causal knowledge graph and the relationship strength matrix are combined to form a time-varying causal knowledge graph; Collecting the newly added node-to-node relationship edges of the time-varying causal knowledge graph within a preset time period, inputting the newly added node-to-node relationship edges into a recurrent neural network to obtain dynamic change characteristics of the node-to-node relationship, and updating the relationship strength matrix based on the dynamic change characteristics of the node-to-node relationship to obtain an updated relationship strength matrix; Constructing a hierarchical user activity prediction model according to the user grouping result, wherein the hierarchical user activity prediction model includes a causal convolutional network layer for extracting active mode features of user groups, a multi-head causal attention network layer for calculating temporal influence features between user groups, and a neural ordinary differential equation layer for obtaining continuous dynamic features of user activity; Input the updated relationship strength matrix into the hierarchical user activity prediction model to obtain user group activity mode features, temporal influence features between user groups and continuous dynamic features of user activity, input the user group activity mode features, temporal influence features between user groups and continuous dynamic features of user activity into the causal fusion network for weighted fusion, and introduce causal bias values ​​to obtain user activity prediction results; A Bayesian causal network is established based on the user activity prediction result, a variational inference method is used to calculate the probability distribution of content delivery time, an initial content delivery time point is determined through importance sampling according to the content delivery time probability distribution, and a delivery effect deviation is calculated based on a counterfactual analysis method, a delivery time adjustment window is determined according to the delivery effect deviation, an online convex optimization algorithm is used to calculate a delivery parameter adjustment value, and the initial content delivery time point is optimized according to the delivery parameter adjustment value; Multi-dimensional consistency constraint processing is performed on the user activity prediction result and the optimized content delivery time point to obtain a final content delivery schedule that satisfies the multi-dimensional consistency constraint.

7. The method according to claim 6, characterized in that Calculating the delivery effect deviation based on the counterfactual analysis method, determining the delivery time adjustment window according to the delivery effect deviation, calculating the delivery parameter adjustment value using an online convex optimization algorithm, and optimizing the initial content delivery time point according to the delivery parameter adjustment value includes: Obtaining the initial content delivery time point, constructing a counterfactual evaluation model, the counterfactual evaluation model comprising a causal intervention operation module and a random sampling module, calculating the user response expectation value before and after the delivery time adjustment through the causal intervention operation module, and performing importance sampling on the user response expectation value through the random sampling module to obtain a counterfactual delivery effect evaluation value; A time window adjustment function is constructed based on the counterfactual delivery effect evaluation value, wherein the time window adjustment function calculates the adaptive parameters of the time window boundary according to the effect evaluation deviation value and the effect fluctuation measurement value, and dynamically updates the time window boundary using the gradient descent method to obtain an optimized delivery time window; Inputting the optimized delivery time window into an online convex optimization model, wherein the online convex optimization model includes a delivery effect loss function and a regularization constraint term, and adopting an adaptive subgradient descent algorithm with momentum term to iteratively optimize the delivery parameters to obtain optimized delivery parameters; Constructing a parameter adjustment feedback model based on the optimized delivery parameters, wherein the parameter adjustment feedback model extracts candidate parameters from the posterior distribution of parameters using a probability sampling method, updates the parameter distribution in real time through a state estimation filter, and outputs the optimal delivery parameters; The initial content delivery time point is adjusted according to the optimal delivery parameters to obtain an optimized content delivery time point.

8. A private domain traffic scheduling and content distribution system based on group behavior analysis, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain historical behavior data of a target user group in a private domain traffic platform and perform preprocessing to obtain effective behavior data, construct a user-behavior association matrix based on the effective behavior data, obtain user interaction weights by calculating the interaction frequency and interaction depth between each user node in the user-behavior association matrix, construct a user social network based on the user interaction weights, calculate user influence weights using a node importance algorithm in the user social network, linearly combine the user-behavior association matrix with the user influence weights to obtain a user group portrait, perform time dimension analysis on the user group portrait, and extract long-term interest distribution and short-term interest distribution of the target user group; The second unit is used to obtain a set of candidate content to be distributed in the private domain traffic platform, perform semantic analysis on the candidate content set to obtain content semantic features, calculate the similarity between the contents in the candidate content set according to the content semantic features to obtain a content similarity matrix, build a content propagation network based on the content similarity matrix, calculate the propagation coefficient of each content node in the content propagation network, calculate the matching coefficient between each content in the candidate content set and the target user group in combination with the long-term interest distribution and the short-term interest distribution, and generate a target content ranking table according to the propagation coefficient and the matching coefficient; The third unit is used to group the target user group according to the long-term interest distribution and the short-term interest distribution to obtain user grouping results, build a user activity prediction model according to the user grouping results, use the user activity prediction model to calculate the active time period distribution of each group of users, generate a content delivery schedule based on the active time period distribution and the target content ranking table, and determine the content delivery frequency of each user group based on the matching coefficient, combine the content delivery schedule and the content delivery frequency to generate a content delivery strategy, execute content delivery in the private domain traffic platform according to the content delivery strategy, collect user interaction data on the delivered content, calculate delivery effect indicators according to the interaction data, dynamically adjust the content delivery strategy based on the delivery effect indicators, and continue to execute content delivery according to the adjusted content delivery strategy.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images