Social user behavior recommendation system based on deep reinforcement learning

Abstract

The invention discloses a social user behavior recommendation system based on deep reinforcement learning, and the system comprises the following modules: a data collection and preprocessing module which is used for collecting and standardizing social original data to generate a user behavior feature vector; the vector mapping module is used for mapping the user behavior feature vector to a potential behavior space; the causal graph construction module is used for calculating the distance between the sample points based on the Minkowski distance and establishing a causal graph; the energy field construction module is used for calculating potential behavior energy corresponding to each node and forming a behavior energy field through an energy transfer mechanism; the force field construction module is used for performing Laplace optimization on the behavior energy field to obtain a behavior influence force field; and the personalized recommendation generation module is used for dividing different user groups through a DBSCAN algorithm according to the behavior influence field and generating user behavior recommendation content. According to the method, the Minkowski distance and the DBSCAN algorithm are combined to realize accurate personalized social behavior recommendation.

Description

Technical Field

[0001] This invention relates to the field of social network analysis technology, and in particular to a social user behavior recommendation system based on deep reinforcement learning. Background Technology

[0002] In current social recommendation systems, the accuracy of personalized recommendations and the improvement of user experience remain key research focuses. Most existing social user behavior recommendation methods rely on traditional algorithms based on collaborative filtering or content recommendation, which typically depend on users' historical behavioral data. However, these methods face limitations and challenges, especially in large-scale social networks. Accurately identifying users' potential interests and behavioral patterns to provide more precise personalized recommendations is a problem that urgently needs to be solved.

[0003] Traditional recommender systems primarily rely on collaborative filtering algorithms based on user-item interaction history. These methods predict a user's future interests based on past behavior. While collaborative filtering can capture some user interests, it depends on large amounts of user interaction data and cannot effectively handle the cold start problem (recommendations for new users or new content). Collaborative filtering often suffers from significant data scarcity when dealing with highly sparse social data, leading to reduced accuracy. Because collaborative filtering only focuses on user-item interactions, ignoring complex social relationships and behavioral interactions among users, it fails to fully exploit the potential causal relationships within social networks. Automatic learning of data features through neural network models can extract more effective representations from complex social data. However, despite the improvements in recommender system performance to some extent, existing deep learning recommender models still face the following challenges: first, how to effectively integrate various types of user behavior data to obtain more accurate behavioral features; and second, how to handle the complex causal relationships and influence propagation problems in social networks, which traditional deep learning models often overlook. Current technologies fail to fully utilize crucial information such as user interaction patterns, causal relationships, and behavioral energy within social networks. This information could provide recommendation systems with deeper user interest analysis and behavioral prediction. Especially on large social platforms, user behavior data is often influenced by other users; user interactions and social network relationships can significantly impact recommendation results. However, traditional recommendation algorithms have failed to effectively tap into this potential social influence, resulting in insufficient relevance and personalization of recommended content.

[0004] Therefore, how to provide a social user behavior recommendation system based on deep reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a social user behavior recommendation system based on deep reinforcement learning. This invention constructs an accurate social user behavior recommendation system by combining Minkowski distance and DBSCAN algorithm. This system generates personalized recommendations by analyzing user behavior characteristics, social relationships, and potential causal influence. It segments user groups and dynamically adjusts recommendations based on group interests, overcoming the cold start problem and data scarcity problem of traditional recommendation systems, improving recommendation accuracy and personalization, and significantly enhancing user experience and satisfaction.

[0006] A social user behavior recommendation system based on deep reinforcement learning according to an embodiment of the present invention includes the following modules:

[0007] The data acquisition and preprocessing module is used to collect and standardize raw social data from multiple heterogeneous data sources to generate user behavior feature vectors.

[0008] The vector mapping module is used to map user behavior feature vectors to the latent behavior space;

[0009] The causal graph construction module is used to calculate the distance between sample points based on the Minkowski distance and build a causal relationship graph.

[0010] The energy field construction module is used to calculate the potential behavioral energy corresponding to each node in the causal relationship graph and form a behavioral energy field through the energy transfer mechanism.

[0011] The force field construction module is used to construct a Laplace matrix based on the behavior energy field and perform Laplace optimization to obtain the behavior influence field.

[0012] The personalized recommendation generation module is used to segment different user groups based on the behavioral influence field using the DBSCAN algorithm and generate user behavior recommendation content.

[0013] A social user behavior recommendation method based on deep reinforcement learning according to an embodiment of the present invention includes the following steps:

[0014] Step 1: Collect raw social data from multiple heterogeneous data sources, and preprocess the raw social data to generate user behavior feature vectors;

[0015] Step 2: Construct a user behavior feature matrix based on the user behavior feature vector, and map the user behavior feature vector as sample points to the latent behavior space;

[0016] Step 3: Calculate the distance between any two sample points in the latent behavior space using Minkowski distance, and construct a causal relationship graph based on the distance values;

[0017] Step 4: Based on the causal relationship diagram, generate the potential behavioral energy corresponding to each node, and form a behavioral energy field through the energy transfer mechanism;

[0018] Step 5: Construct a Laplace matrix based on the behavioral energy field and perform Laplace optimization to obtain the behavioral influence field;

[0019] Step Six: Based on the behavioral influence field, segment different user groups using the DBSCAN algorithm and generate user behavior recommendation content.

[0020] Optionally, step one specifically includes:

[0021] The raw social data includes user interaction data, social network relationship data, browsing history data, and comment and like data;

[0022] The preprocessing steps include:

[0023] Different types of data are processed by outlier removal, missing value imputation, timestamp alignment and normalization to obtain user behavior data with a unified structure.

[0024] User behavior data is input into a convolutional neural network for feature extraction to obtain a user behavior feature vector. The convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, which are used to extract user behavior features from user behavior data through convolution operations, non-linear activation functions, and max pooling operations.

[0025] Optionally, step two specifically involves:

[0026] Calculate the mean and standard deviation of each feature in the user behavior feature vector;

[0027] The centered user behavior feature vector is obtained by subtracting the mean from each feature in the user behavior feature vector and dividing by the standard deviation.

[0028] A user behavior feature matrix is ​​constructed based on the centralized user behavior feature vectors. Each row in the user behavior feature matrix represents a user, and each column represents a centralized user behavior feature vector.

[0029] Calculate the covariance between any two columns in the user behavior feature matrix to obtain the user behavior covariance matrix;

[0030] The user behavior covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvectors and corresponding eigenvalues. The eigenvectors represent the main directions, and the eigenvalues ​​represent the variance in the main directions.

[0031] The first few eigenvectors with the largest eigenvalues ​​are selected to form the basis vectors in the latent behavior space.

[0032] Optionally, step three specifically includes:

[0033] Calculate the Manhattan distance between any two sample points in the latent behavior space to form a Manhattan distance matrix, where the Manhattan distance represents the similarity of sample points in the latent behavior space;

[0034] Calculate the Minkowski distance between any two sample points in the latent behavior space to obtain the Minkowski distance matrix;

[0035] The Manhattan distance matrix and the Minkowski distance matrix are summed to obtain the combined distance matrix;

[0036] A causal relationship graph is constructed based on the comprehensive distance matrix. The construction steps are as follows:

[0037] The sample points in the potential behavior space are used as nodes in the causal relationship graph;

[0038] Each column in the integrated distance matrix is ​​used as the target column, and the remaining columns are used as candidate columns.

[0039] Linear regression is performed on the target column and candidate columns to obtain the corresponding fitting residuals and fitting coefficients;

[0040] If the fitting residual is less than the preset residual threshold and the fitting coefficient is greater than the preset coefficient threshold, then the corresponding candidate column is pointed to the target column, and a causal connection edge is established.

[0041] Traverse all columns of the comprehensive distance matrix to obtain the set of causal connection edges;

[0042] Remove causal connections with loop structures from the set of causal connections. The weight of each causal connection is the comprehensive distance value of the corresponding node in the comprehensive distance matrix.

[0043] Optionally, step four specifically involves:

[0044] The potential behavior energy is initialized based on the degree of the nodes in the causal graph, where the degree is the number of causal connection edges of the node;

[0045] For each node, the potential behavior energy is calculated based on the edge weights between it and its neighboring nodes in the causal relationship graph. The potential behavior energy is the product of the initial potential behavior energy and the corresponding edge weight.

[0046] The potential behavior energy of each node is calculated iteratively until the potential behavior energy of all nodes converges, forming the final potential behavior energy value.

[0047] The formation of a behavioral energy field through an energy transfer mechanism specifically refers to:

[0048] The initial energy of each node is obtained based on the converged potential behavioral energy value;

[0049] Based on the initial energy of a node, along the causal connection edges in the causal relationship graph, the node proportionally transfers its own energy to neighboring nodes. The proportion of energy transfer is the ratio of the weight of the causal connection edge between the current node and its neighboring nodes to the sum of the weights of all causal connection edges in the causal relationship graph.

[0050] Multiple rounds of energy transfer are performed. In each round, the energy of a node is updated based on the energy received from neighboring nodes and its own potential behavioral energy, until the potential behavioral energy of all nodes converges. Convergence is defined as the change in potential behavioral energy being lower than a preset threshold.

[0051] After the potential behavioral energy of all nodes converges, the energy distribution of the entire causal graph forms a behavioral energy field, which represents the influence of the behavioral characteristics of each user in the social network.

[0052] Optionally, step five specifically includes:

[0053] Construct the adjacency matrix and degree matrix using nodes and causal connections in the behavioral energy field;

[0054] Each element in the adjacency matrix represents the difference in potential behavioral energy between nodes;

[0055] The degree matrix represents the number of causal connections for each node in the behavioral energy field.

[0056] The Laplace matrix is ​​obtained by multiplying the square root of the inverse of the degree matrix with the adjacency matrix.

[0057] Multiply the potential behavioral energy of the current node in the behavioral energy field by the minimum element value of the corresponding column of the current node in the Laplace matrix to obtain the Laplace-optimized node energy value.

[0058] Each node in the behavioral energy field is individually optimized using Laplace to generate a behavioral influence field.

[0059] Optionally, step six specifically includes:

[0060] Calculate the number of nodes within a preset size range for each node in the behavioral influence field. If the number of nodes is greater than or equal to the preset minimum number of samples in the neighborhood, then the corresponding node is the core point.

[0061] For each core point, check if any nodes within its neighborhood can also be core points. If a node within the neighborhood is also a core point, add the corresponding node to the current cluster and continue expanding until the current cluster can no longer add any more nodes.

[0062] If a node has only one core point in its neighborhood, then the corresponding node is a boundary point, and the boundary point is assigned to the cluster of core points in the current neighborhood.

[0063] If a node is neither a core point nor a boundary point, it is marked as a noise point and is not assigned to any cluster.

[0064] The statistical behavior influence field is divided into multiple clusters, each of which represents a user behavior group;

[0065] Calculate the average behavioral characteristics of each user behavior group, where the average behavioral characteristics represent the overall interests of the current user behavior group;

[0066] Personalized recommendations are generated based on the overall interests of each user behavior group, combined with popular content and related topics on social media platforms.

[0067] The beneficial effects of this invention are:

[0068] Compared to existing technologies, this invention provides a novel recommendation method that more effectively addresses some key issues of traditional recommendation systems in social network applications. Existing recommendation systems typically rely on the interaction history between users and items, failing to delve into the social relationships and behavioral patterns among users, resulting in limited accuracy and personalization of recommendation results. This invention, by combining deep learning, causal inference, and social network analysis, can model user behavior from multiple dimensions, thereby achieving more accurate personalized recommendations. By introducing a causal relationship graph, this invention effectively captures the complex relationships between users in social networks and constructs a behavioral energy field through the calculation and energy transfer mechanism of potential behavioral energy, thus more realistically reflecting users' interactions and influence in social networks. Compared to traditional methods, this process better reveals the potential causal connections between user behaviors, thereby improving the accuracy of the recommendation system. Through Laplace optimization of the behavioral energy field, the system further adjusts the distribution of node energy, optimizing the relevance and accuracy of recommended content and avoiding common biases and errors in traditional recommendation methods. This invention uses the DBSCAN algorithm to segment user groups and analyze user behavior according to group characteristics, making recommended content more aligned with the group's interests and preferences, thereby improving the personalization level of the recommendation system. Recommended content generated based on group behavior characteristics can effectively avoid the overfitting of traditional algorithms to individual users and can dynamically adapt to changes in user interests in social networks, ensuring the timeliness and accuracy of recommended content.

[0069] This invention overcomes the shortcomings of traditional recommendation systems, particularly in cold start, data scarcity, and social relationship modeling, by combining deep reinforcement learning and causal inference with multiple heterogeneous data sources. Through in-depth mining of user behavior data and incorporating causal influence within social networks, this invention achieves more accurate and personalized recommendations, not only improving user satisfaction but also providing social platforms with a more powerful content recommendation engine. Attached Figure Description

[0070] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0071] Figure 1 This is a schematic diagram of the structure of a social user behavior recommendation system based on deep reinforcement learning proposed in this invention;

[0072] Figure 2 This is an overall flowchart of a social user behavior recommendation method based on deep reinforcement learning proposed in this invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0074] refer to Figure 1 A social user behavior recommendation system based on deep reinforcement learning includes the following modules:

[0075] The data acquisition and preprocessing module is used to collect and standardize raw social data from multiple heterogeneous data sources to generate user behavior feature vectors.

[0076] The vector mapping module is used to map user behavior feature vectors to the latent behavior space;

[0077] The causal graph construction module is used to calculate the distance between sample points based on the Minkowski distance and build a causal relationship graph.

[0078] The energy field construction module is used to calculate the potential behavioral energy corresponding to each node in the causal relationship graph and form a behavioral energy field through the energy transfer mechanism.

[0079] The force field construction module is used to construct a Laplace matrix based on the behavior energy field and perform Laplace optimization to obtain the behavior influence field.

[0080] The personalized recommendation generation module is used to segment different user groups based on the behavioral influence field using the DBSCAN algorithm and generate user behavior recommendation content.

[0081] This step, by comprehensively utilizing multi-source heterogeneous data and deep learning techniques, effectively captures user behavior and interest patterns in social networks. Through the construction of causal relationship graphs and the calculation of potential behavioral energy, this step delves into the potential connections between user behaviors, thereby providing more accurate behavior predictions. Utilizing Laplace optimization and the DBSCAN algorithm, this invention can better segment user groups and generate personalized recommendation content, improving the relevance and personalization of recommendations. It overcomes the problems of cold start, data scarcity, and insufficient social relationship modeling in traditional recommendation systems, greatly enhancing user experience and satisfaction.

[0082] refer to Figure 2 A social user behavior recommendation method based on deep reinforcement learning includes the following steps:

[0083] Step 1: Collect raw social data from multiple heterogeneous data sources, and preprocess the raw social data to generate user behavior feature vectors;

[0084] Step 2: Construct a user behavior feature matrix based on the user behavior feature vector, and map the user behavior feature vector as sample points to the latent behavior space;

[0085] Step 3: Calculate the distance between any two sample points in the latent behavior space using Minkowski distance, and construct a causal relationship graph based on the distance values;

[0086] Step 4: Based on the causal relationship diagram, generate the potential behavioral energy corresponding to each node, and form a behavioral energy field through the energy transfer mechanism;

[0087] Step 5: Construct a Laplace matrix based on the behavioral energy field and perform Laplace optimization to obtain the behavioral influence field;

[0088] Step Six: Based on the behavioral influence field, segment different user groups using the DBSCAN algorithm and generate user behavior recommendation content.

[0089] This step effectively improves the accuracy and personalization of social user behavior recommendations through a series of innovative steps. By collecting and preprocessing multi-source heterogeneous data, accurate user behavior feature vectors are generated, enabling effective mapping within the latent behavior space. The introduction of Minkowski distance makes distance calculations between sample points more precise, thus constructing an accurate causal relationship graph. Through the calculation of latent behavior energy and the application of energy transfer mechanisms, a behavior energy field is formed, and Laplace optimization enhances the assessment of user behavior influence. Finally, by combining the DBSCAN algorithm to segment user groups, the system can generate more personalized recommendations that align with user interests, solving the problems of insufficient personalization, data scarcity, and cold start in traditional recommendation systems.

[0090] In this embodiment, step one specifically includes:

[0091] The raw social data includes user interaction data, social network relationship data, browsing history data, and comment and like data;

[0092] The preprocessing steps include:

[0093] Different types of data are processed by outlier removal, missing value imputation, timestamp alignment and normalization to obtain user behavior data with a unified structure.

[0094] User behavior data is input into a convolutional neural network for feature extraction to obtain a user behavior feature vector. The convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, which are used to extract user behavior features from user behavior data through convolution operations, non-linear activation functions, and max pooling operations.

[0095] This step ensures high-quality generation of user behavior feature vectors through precise preprocessing of various types of raw social data. By cleaning and standardizing user interaction data, social network relationship data, browsing history data, and comment and like data, the accuracy issues of traditional methods under data scarcity and noise are addressed. The introduction of convolutional neural networks makes the feature extraction process more efficient, automatically extracting more representative behavioral features through convolutional layers, pooling layers, and non-linear activation functions. This approach not only improves the efficiency of feature extraction but also enables recommendation systems to better handle the complexity and diversity of data, enhancing the accuracy of personalized recommendations and user experience.

[0096] In this embodiment, step two specifically includes:

[0097] Calculate the mean and standard deviation of each feature in the user behavior feature vector;

[0098] The centered user behavior feature vector is obtained by subtracting the mean from each feature in the user behavior feature vector and dividing by the standard deviation.

[0099] A user behavior feature matrix is ​​constructed based on the centralized user behavior feature vectors. Each row in the user behavior feature matrix represents a user, and each column represents a centralized user behavior feature vector.

[0100] Calculate the covariance between any two columns in the user behavior feature matrix to obtain the user behavior covariance matrix;

[0101] The user behavior covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvectors and corresponding eigenvalues. The eigenvectors represent the main directions, and the eigenvalues ​​represent the variance in the main directions.

[0102] The first few eigenvectors with the largest eigenvalues ​​are selected to form the basis vectors in the latent behavior space.

[0103] This step improves data quality and recommender system performance through precise user behavior feature vector processing. By calculating the mean and standard deviation of each feature and centering them, scale differences between different features are eliminated, making subsequent analysis more accurate. A user behavior feature matrix is ​​constructed using the centered feature vectors, which more clearly reflects each user's performance on each behavioral feature. Based on this, by calculating the covariance matrix and performing eigenvalue decomposition, the main directions and variances of user behavior can be accurately extracted, thus selecting the most representative feature vectors. This process not only enhances the expressive power of the data but also ensures the extraction of the most important features in the latent behavior space, further improving the recommender system's accuracy in capturing user behavior patterns and providing personalized recommendations.

[0104] In this embodiment, step three specifically includes:

[0105] Calculate the Manhattan distance between any two sample points in the latent behavior space to form a Manhattan distance matrix, where the Manhattan distance represents the similarity of sample points in the latent behavior space;

[0106] Calculate the Minkowski distance between any two sample points in the latent behavior space to obtain the Minkowski distance matrix;

[0107] The Manhattan distance matrix and the Minkowski distance matrix are summed to obtain the combined distance matrix;

[0108] A causal relationship graph is constructed based on the comprehensive distance matrix. The construction steps are as follows:

[0109] The sample points in the potential behavior space are used as nodes in the causal relationship graph;

[0110] Each column in the integrated distance matrix is ​​used as the target column, and the remaining columns are used as candidate columns.

[0111] Linear regression is performed on the target column and candidate columns to obtain the corresponding fitting residuals and fitting coefficients;

[0112] If the fitting residual is less than the preset residual threshold and the fitting coefficient is greater than the preset coefficient threshold, then the corresponding candidate column is pointed to the target column, and a causal connection edge is established.

[0113] Traverse all columns of the comprehensive distance matrix to obtain the set of causal connection edges;

[0114] Remove causal connections with loop structures from the set of causal connections. The weight of each causal connection is the comprehensive distance value of the corresponding node in the comprehensive distance matrix.

[0115] This step further improves the accuracy of user behavior analysis in the latent behavior space by combining Manhattan distance and Minkowski distance to measure the similarity between sample points. Two distance matrices are generated by calculating Manhattan and Minkowski distances, and then summed to obtain a comprehensive distance matrix. This process fully considers the multi-dimensional relationships of user behavior features. Based on this, a causal relationship graph is constructed using the comprehensive distance matrix. Linear regression is used to analyze the potential causal relationships between each pair of user behavior features, establishing causal connections. This method not only reveals the mutual influence between user behaviors but also optimizes the structure of the causal graph by removing loop structures, ensuring the accuracy of causal relationships. This innovative step enables the recommender system to more accurately capture the potential causal effects of user behavior, improving the accuracy and reliability of personalized recommendations.

[0116] In this embodiment, step four specifically includes:

[0117] The potential behavior energy is initialized based on the degree of the nodes in the causal graph, where the degree is the number of causal connection edges of the node;

[0118] For each node, the potential behavior energy is calculated based on the edge weights between it and its neighboring nodes in the causal relationship graph. The potential behavior energy is the product of the initial potential behavior energy and the corresponding edge weight.

[0119] The potential behavior energy of each node is calculated iteratively until the potential behavior energy of all nodes converges, forming the final potential behavior energy value.

[0120] The formation of a behavioral energy field through an energy transfer mechanism specifically refers to:

[0121] The initial energy of each node is obtained based on the converged potential behavioral energy value;

[0122] Based on the initial energy of a node, along the causal connection edges in the causal relationship graph, the node proportionally transfers its own energy to neighboring nodes. The proportion of energy transfer is the ratio of the weight of the causal connection edge between the current node and its neighboring nodes to the sum of the weights of all causal connection edges in the causal relationship graph.

[0123] Multiple rounds of energy transfer are performed. In each round, the energy of a node is updated based on the energy received from neighboring nodes and its own potential behavioral energy, until the potential behavioral energy of all nodes converges. Convergence is defined as the change in potential behavioral energy being lower than a preset threshold.

[0124] After the potential behavioral energy of all nodes converges, the energy distribution of the entire causal graph forms a behavioral energy field, which represents the influence of the behavioral characteristics of each user in the social network.

[0125] This step initializes the node degree in the causal graph, enabling the potential behavioral energy of each node to better reflect its influence in the social network. By calculating the potential behavioral energy through edge weights, the system can accurately quantify the interaction and influence between each node and its neighbors. Through multiple iterations, the potential behavioral energy of the nodes gradually converges, forming the final energy value. This process not only enhances the dynamic adjustment of node energy in the causal graph but also effectively transmits the behavioral influence between nodes. In the energy transfer mechanism, the energy of a node is proportionally transferred to its neighboring nodes, optimizing energy flow and allowing the entire behavioral energy field to more realistically reflect the influence of each user's behavioral characteristics in the social network. In this way, the recommendation system can more accurately capture the social interactions and behavioral influences between users, improving the accuracy and relevance of personalized recommendations.

[0126] In this embodiment, step five specifically includes:

[0127] Construct the adjacency matrix and degree matrix using nodes and causal connections in the behavioral energy field;

[0128] Each element in the adjacency matrix represents the difference in potential behavioral energy between nodes;

[0129] The degree matrix represents the number of causal connections for each node in the behavioral energy field.

[0130] The Laplace matrix is ​​obtained by multiplying the square root of the inverse of the degree matrix with the adjacency matrix.

[0131] Multiply the potential behavioral energy of the current node in the behavioral energy field by the minimum element value of the corresponding column of the current node in the Laplace matrix to obtain the Laplace-optimized node energy value.

[0132] Each node in the behavioral energy field is individually optimized using Laplace to generate a behavioral influence field.

[0133] This step further optimizes the energy evaluation process of the social user behavior recommendation system by constructing an adjacency matrix and a degree matrix, based on nodes and causal connections in the behavioral energy field. The adjacency matrix effectively represents the potential behavioral energy difference between nodes, while the degree matrix reflects the number of causal connections between nodes, providing structural support for energy transfer. By calculating the Laplace matrix and combining it with the product of the inverse square root of the degree matrix and the adjacency matrix, the system achieves reasonable energy optimization. Laplace optimization effectively balances the energy distribution between nodes by multiplying the potential behavioral energy of a node by the minimum element value in the Laplace matrix, improving the accuracy of node influence assessment. Finally, the energy value of each node after Laplace optimization generates a behavioral influence field, which accurately reflects the actual influence of each user's behavioral characteristics in the social network, significantly improving the accuracy and relevance of personalized recommendations.

[0134] In this embodiment, step six specifically includes:

[0135] Calculate the number of nodes within a preset size range for each node in the behavioral influence field. If the number of nodes is greater than or equal to the preset minimum number of samples in the neighborhood, then the corresponding node is the core point.

[0136] For each core point, check if any nodes within its neighborhood can also be core points. If a node within the neighborhood is also a core point, add the corresponding node to the current cluster and continue expanding until the current cluster can no longer add any more nodes.

[0137] If a node has only one core point in its neighborhood, then the corresponding node is a boundary point, and the boundary point is assigned to the cluster of core points in the current neighborhood.

[0138] If a node is neither a core point nor a boundary point, it is marked as a noise point and is not assigned to any cluster.

[0139] The statistical behavior influence field is divided into multiple clusters, each of which represents a user behavior group;

[0140] Calculate the average behavioral characteristics of each user behavior group, where the average behavioral characteristics represent the overall interests of the current user behavior group;

[0141] Personalized recommendations are generated based on the overall interests of each user behavior group, combined with popular content and related topics on social media platforms.

[0142] This step effectively enhances the personalization level of the social user behavior recommendation system by using the DBSCAN algorithm to segment users into groups within the behavioral influence field. By calculating the number of nodes within a preset domain for each node, the system can accurately identify core points and form stable user groups by expanding the neighborhood of these core points. Labeling boundary points and noise points ensures the accuracy and rationality of clustering. Calculating the average behavioral characteristics of each user group allows the system to accurately capture the overall interests of the group, further improving the relevance and accuracy of recommended content. Combining popular content and topics on the social platform, the system provides more personalized recommendations that align with the preferences of different user groups, significantly improving user experience and the overall recommendation performance.

[0143] Example 1:

[0144] To verify the feasibility of this invention in practice, it was applied to a social media platform. This platform is a social media application where users can browse, like, comment, and share content. We analyzed and optimized the personalized recommendation system using user behavior data from this platform.

[0145] On this social platform, user behavioral data includes their interactions with others (such as comments and likes), social network relationships (such as friend relationships and number of followers), content browsing history, and reactions to posts and advertisements. Traditional recommendation systems typically provide recommendations based on users' interaction history with content, using collaborative filtering or simple content recommendation algorithms. However, these methods sometimes fail to accurately capture the complex social interactions and potential changes in interests among users. Furthermore, existing technologies struggle to handle this dynamic nature because user relationships and interests on social platforms change at different points in time.

[0146] By analyzing multi-source heterogeneous data and combining causal inference and behavioral energy fields, the system can accurately capture potential causal relationships in social networks, thereby generating personalized recommendations for each user. Through deep learning and energy field optimization, the system not only improves the accuracy of recommendations but also addresses issues such as data scarcity and cold start.

[0147] The platform collects user behavior data from multiple data sources. This data includes posts viewed, comments, likes, and interactions on social networks. To ensure data quality, the platform preprocesses this data, including outlier removal, missing value imputation, timestamp alignment, and normalization. Next, the system uses a convolutional neural network (CNN) to extract features from the user behavior data, obtaining a behavior feature vector for each user.

[0148] After feature extraction, the system uses these feature vectors to construct a user behavior feature matrix and calculates the similarity between users using Minkowski distance to build a causal relationship graph. Each node represents a user behavior feature, and the edge weights represent the potential causal relationships between users. Based on this causal relationship graph, the system calculates the potential behavior energy corresponding to each node and forms a behavior energy field through an energy transfer mechanism. After the energy field is constructed, the system uses the Laplace optimization method to adjust the energy of the nodes, ultimately obtaining an optimized behavior influence field. This influence field reflects the behavioral influence of each user in the social network and can provide a strong basis for subsequent recommendations.

[0149] The system divides users into multiple groups based on the influence field of behavior using the DBSCAN algorithm. DBSCAN is a density-based clustering algorithm that automatically identifies different user groups based on the similarity and behavioral characteristics between users. Users in each group have similar interests and behavioral patterns, thus enabling the generation of group-specific recommendations, rather than just recommendations for individual users. This group segmentation method effectively solves the problem of inaccurate recommendations caused by individual user behavioral differences in traditional recommendation systems.

[0150] In generating recommended content, the system considers not only users' historical behavioral data but also popular content and topics on social media platforms to generate recommendations relevant to group interests. For example, for a user group that enjoys tech news and product reviews, the system might recommend the latest comments or reviews related to tech products. For an active social media user group, the system might recommend social events or discussions of trending topics. Through this user-group-based recommendation approach, the system ensures the diversity and relevance of recommended content, meeting the needs of different user groups.

[0151] In practical application, we selected user data from a social media platform for validation. From the platform's 5000 active users, we randomly selected data from 1000 users, including their browsing history, likes, and comments. After preprocessing and feature extraction, the system calculated the potential behavioral energy of these users and segmented the users into groups using Laplace optimization and the DBSCAN algorithm.

[0152] In the comparative experiment, we compared the recommendation system of this invention with traditional methods, and the comparison results are shown in the table below:

[0153] Table 1. Comparison of Personalized Recommendation Data

[0154] Method types Recommended accuracy Personalized recommendation satisfaction User behavior similarity accuracy Recommended coverage System response time User engagement This invention system 85% 92% 88% 80% 1.2 seconds 85% Traditional collaborative filtering algorithm 65% 78% 70% 60% 2.5 seconds 65% Content-based recommendation systems 72% 80% 74% 68% 1.8 seconds 70% Recommendation systems based on matrix factorization 78% 82% 76% 74% 1.5 seconds 72% Recommendation system based on user-item collaboration 68% 75% 71% 65% 2.0 seconds 68%

[0155] As can be seen from the comparative data in Table 1, the deep reinforcement learning-based recommendation system of this invention outperforms traditional algorithms in all key metrics. Firstly, the recommendation accuracy and personalized recommendation satisfaction both reach 85% and 92%, respectively, significantly higher than the 65% and 78% of traditional collaborative filtering algorithms. Furthermore, this system also surpasses other recommendation methods in user behavior similarity accuracy (88%) and recommendation coverage (80%), indicating that it can more comprehensively capture user interest characteristics and provide more relevant recommended content. The system response time is 1.2 seconds, providing a smoother user experience. The user engagement rate is as high as 85%, further demonstrating that the system effectively improves user activity and engagement. Overall, the recommendation system of this invention, through deep reinforcement learning and causal inference, achieves significant improvements in accuracy, personalized recommendation, and user experience, solving problems such as cold start, data scarcity, and insufficient recommendation accuracy in traditional recommendation methods.

[0156] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep reinforcement learning based social user behavior recommendation system, characterized in that, The system comprises the following modules: a data collection and preprocessing module for collecting and standardizing social raw data from multiple heterogeneous data sources to generate user behavior feature vectors; a vector mapping module for mapping the user behavior feature vectors to a latent behavior space; a causal graph construction module for calculating the distance between sample points based on Minkowski distance and establishing a causal relationship graph; an energy field construction module for calculating the latent behavior energy corresponding to each node in the causal relationship graph and forming a behavior energy field through an energy transfer mechanism; a force field construction module for constructing a Laplacian matrix based on the behavior energy field and performing Laplace optimization to obtain a behavior influence field; a personalized recommendation generation module for dividing different user groups according to the behavior influence field through a DBSCAN algorithm and generating user behavior recommendation content. 2.The social user behavior recommendation system based on deep reinforcement learning of claim 1, wherein, The modules are realized through the following methods: Step 1: Collect social raw data from multiple heterogeneous data sources and preprocess the social raw data to generate user behavior feature vectors; Step 2: Construct a user behavior feature matrix according to the user behavior feature vectors and map the user behavior feature vectors as sample points to a latent behavior space; Step 3: Calculate the distance between any two sample points in the latent behavior space using Minkowski distance and construct a causal relationship graph according to the distance values; Step 4: Generate the latent behavior energy corresponding to each node according to the causal relationship graph and form a behavior energy field through an energy transfer mechanism; Step 5: Construct a Laplacian matrix based on the behavior energy field and perform Laplace optimization to obtain a behavior influence field; Step 6: Divide different user groups according to the behavior influence field through a DBSCAN algorithm and generate user behavior recommendation content. 3.The social user behavior recommendation system based on deep reinforcement learning of claim 2, wherein, The step 1 is specifically: The social raw data includes user interaction data, social network relationship data, browsing history data, comment and like data; The preprocessing step includes: Perform outlier rejection, missing value filling, timestamp alignment and normalization processing on different types of data to obtain user behavior data with unified structure; Input the user behavior data into a convolutional neural network to extract features and obtain user behavior feature vectors, the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer, which is used to extract user behavior features in user behavior data through convolution operation, nonlinear activation function and maximum pooling operation. 4.The social user behavior recommendation system based on deep reinforcement learning of claim 2, wherein, The step 2 is specifically: Calculate the mean and standard deviation of each feature in the user behavior feature vector; Subtract the corresponding mean from each feature in the user behavior feature vector and divide by the corresponding standard deviation to obtain the centralized user behavior feature vector, Construct a user behavior feature matrix based on the centralized user behavior feature vector, each row in the user behavior feature matrix represents a user, and each column represents a centralized user behavior feature vector; Calculate the covariance between any two columns in the user behavior feature matrix to obtain a user behavior covariance matrix; Eigenvalue decomposition is performed on the user behavior covariance matrix to obtain eigenvectors and corresponding eigenvalues, the eigenvectors represent main directions, and the eigenvalues represent variances in the main directions; The eigenvectors with the largest eigenvalues are selected to form base vectors in the latent behavior space. 5.The social user behavior recommendation system based on deep reinforcement learning of claim 2, wherein, The step three is specifically: A Manhattan distance matrix is formed by calculating Manhattan distances between any two sample points in the latent behavior space, the Manhattan distance representing the similarity of the sample points in the latent behavior space; A Minkowski distance matrix is obtained by calculating Minkowski distances between any two sample points in the latent behavior space; The Manhattan distance matrix and the Minkowski distance matrix are added to obtain a comprehensive distance matrix; A causal relationship graph is constructed according to the comprehensive distance matrix, and the construction step is: The sample points in the latent behavior space are taken as nodes of the causal relationship graph; Each column in the comprehensive distance matrix is taken as a target column, and the remaining columns are taken as candidate columns; Linear regression fitting is performed on the target column and the candidate columns to obtain corresponding fitting residuals and fitting coefficients; If the fitting residual is less than a preset residual threshold and the fitting coefficient is greater than a preset coefficient threshold, the corresponding candidate column is pointed to the target column to establish a causal connection edge; All columns in the comprehensive distance matrix are traversed to obtain a causal connection edge set; Causal connection edges with loop structures in the causal connection edge set are removed, and the weight of the causal connection edge is the comprehensive distance value of the corresponding node in the comprehensive distance matrix. 6.The social user behavior recommendation system based on deep reinforcement learning of claim 2, wherein, The step four is specifically: The latent behavior energy is initialized according to the degree of the node in the causal relationship graph, and the degree is the number of causal connection edges of the node; For each node, the latent behavior energy is calculated according to the edge weight between adjacent nodes in the causal relationship graph, and the latent behavior energy is the product of the initialized latent behavior energy and the corresponding edge weight; The latent behavior energy of each node is iteratively calculated until the latent behavior energies of all nodes converge to form final latent behavior energy values; The behavior energy field is formed through the energy transmission mechanism, and the specific process is: The initial energy of each node is obtained according to the converged latent behavior energy values; According to the initial energy of the node, the energy of the node is transmitted to the adjacent nodes along the causal connection edges in the causal relationship graph, and the transmission ratio is the ratio of the weight of the causal connection edge between the current node and the adjacent node to the sum of the weights of all causal connection edges in the causal relationship graph; Multiple rounds of energy transmission are performed, and in each round of transmission, the energy of the node is updated according to the energy received from the adjacent nodes and the latent behavior energy of the node until the latent behavior energies of all nodes converge, and the convergence is that the change value of the latent behavior energy is lower than a preset threshold; After the latent behavior energies of all nodes converge, the energy distribution of the entire causal relationship graph forms a behavior energy field, and the behavior energy field represents the influence of each user behavior feature in the social network. 7.The social user behavior recommendation system based on deep reinforcement learning of claim 2, wherein, The step five is specifically: An adjacency matrix and a degree matrix are constructed by using the nodes and the causal connection edges in the behavior energy field; Each element in the adjacency matrix represents the difference between the latent behavior energies of the nodes. The degree matrix is the number of causal connection edges of each node in the behavior energy field; The square root of the inverse of the degree matrix is multiplied with the adjacency matrix to obtain the Laplacian matrix; The potential behavior energy of the current node in the behavior energy field is multiplied with the minimum element value in the corresponding column of the current node in the Laplacian matrix to obtain the node energy value after Laplacian optimization; Each node in the behavior energy field is respectively subjected to Laplacian optimization to generate a behavior influence field. 8.The social user behavior recommendation system based on deep reinforcement learning of claim 2, wherein, The step six is specifically: Calculate the number of nodes in the preset size field range of each node in the behavior influence field, if the number of nodes is greater than or equal to the preset minimum number of samples in the neighborhood, the corresponding node is a core point; For each core point, check whether the nodes in the neighborhood range can also be core points, if the points in the neighborhood are also core points, the corresponding points are added to the current cluster, and the expansion is continued until the current cluster cannot add nodes; If there is only one core point in the neighborhood range of a node, the corresponding node is a boundary point, and the boundary point is divided into the cluster of the current neighborhood range core point; If a node is neither a core point nor a boundary point, it is marked as a noise point and is not assigned to any cluster; Statistical behavior influence field is divided into multiple clusters, each cluster represents a user behavior group; Calculate the average behavior feature of each user behavior group, the average behavior feature represents the overall interest of the current user behavior group; According to the overall interest of each user behavior group, combined with the popular content and related topics in the social platform, personalized recommendation content is generated.

Images