User behavior data analysis method and system based on multi-agent interconnection

Abstract

The application discloses a user behavior data analysis method and system based on multi-agent interconnection. The method comprises the following steps: constructing a space-time correlation graph based on original behavior data of a user in different scenes collected by a multi-agent device; extracting multi-modal features from the space-time correlation graph to obtain a multi-dimensional multi-source heterogeneous feature vector with a space-time label, then performing deep causal relationship analysis between cross-device data to obtain an individualized causal correlation knowledge graph, then designing a multi-layer progressive attention gate network architecture, and mapping the multi-dimensional multi-source heterogeneous feature vector into a unified behavior feature representation vector through a dynamic weight distribution mechanism driven by a user personalized portrait; and inputting the unified behavior feature representation vector into an adversarial analysis system of a generator-discriminator double network architecture to obtain a user behavior data analysis result. The method can obtain accurate, reliable and comprehensive user behavior data analysis results.

Description

Technical Field

[0001] This disclosure relates to the field of interdisciplinary technology of educational informatization and artificial intelligence, and in particular to a method and system for analyzing user behavior data based on multi-agent interconnection. Background Technology

[0002] With the rapid development of IoT, AI, and edge computing technologies, multi-agent devices have shown great potential in the field of user behavior data collection. Traditional user behavior analysis mainly relies on a single data source, such as questionnaires, single sensor monitoring, or simple log recording. This approach suffers from limited data dimensions, poor timeliness, and significant subjective bias. Furthermore, existing data analysis methods often remain at the level of correlation analysis, failing to reveal the deeper causal relationships behind user behavior.

[0003] The core technical challenges currently facing user behavior data analysis include: difficulty in effectively integrating multi-source heterogeneous data and the lack of a unified feature representation space; complexity in identifying causal relationships between cross-device data, with traditional correlation analysis failing to meet the needs of precise intervention; insufficient extraction of individual differentiated features, making it difficult to achieve personalized analysis tailored to each user; prominent issues of algorithmic fairness, with different user groups potentially facing evaluation biases; and limited real-time processing capabilities, resulting in performance bottlenecks in large-scale concurrent user analysis. These technical bottlenecks severely restrict the practicality and accuracy of user behavior analysis systems. Summary of the Invention

[0004] In view of this, the present disclosure provides a user behavior data analysis method and system based on multi-agent interconnection, which can solve the problems of single and incomplete user behavior data analysis dimensions, inaccurate analysis results, and poor reliability in the prior art.

[0005] In a first aspect, embodiments of this disclosure provide a method for analyzing user behavior data based on multi-agent interconnection, including:

[0006] A spatiotemporal correlation map is constructed based on the raw behavioral data of users in different scenarios collected by multi-agent devices.

[0007] Multimodal features are extracted from the spatiotemporal correlation map to obtain a multidimensional, multi-source heterogeneous feature vector with spatiotemporal labels;

[0008] Based on the multidimensional, multi-source, heterogeneous feature vectors, a deep causal relationship analysis is performed between cross-device data to obtain an individualized causal association knowledge graph;

[0009] Based on the individualized causal knowledge graph, a multi-layer progressive attention gating network architecture is designed. Through a dynamic weight allocation mechanism driven by user personalized profiles, the multi-dimensional, multi-source heterogeneous feature vectors are mapped into unified behavioral feature representation vectors.

[0010] The unified behavioral feature representation vector is input into the adversarial analysis system with a generator-discriminator dual-network architecture to obtain user behavior data analysis results.

[0011] Secondly, this disclosure also provides a computer device, which adopts the following technical solution:

[0012] The computer device includes:

[0013] At least one processor; and,

[0014] A memory communicatively connected to the at least one processor; wherein,

[0015] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform any of the user behavior data analysis methods based on multi-agent interconnection described above.

[0016] Thirdly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions for causing a computer to execute any of the above-described user behavior data analysis methods based on multi-agent interconnection.

[0017] Fourthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0018] The user behavior data analysis method disclosed in this application based on multi-agent interconnection first constructs a spatiotemporal correlation graph based on the original user behavior data collected by multiple agent devices in different scenarios. Multimodal features are extracted from the spatiotemporal correlation graph to obtain multidimensional, multi-source, heterogeneous feature vectors with spatiotemporal labels. Comprehensive behavior monitoring is achieved through collaborative data collection by multiple agent devices, avoiding the limitations of a single data source and effectively enriching the data dimensions. The method integrates data of different types and sources, initially realizing the fusion of multi-source heterogeneous data. Then, deep causal relationship analysis is performed between cross-device data based on the multidimensional, multi-source, heterogeneous feature vectors to obtain personalized causal relationship knowledge. The knowledge graph, constructed through cross-device data fusion to eliminate bias from single data sources, provides a scientific basis for precise personalized intervention. Secondly, based on the personalized causal knowledge graph, a multi-layered progressive attention-gated network architecture is designed. Through a dynamic weight allocation mechanism driven by user personalized profiles, multi-dimensional, multi-source heterogeneous feature vectors are mapped to unified behavioral feature representation vectors. This establishes a unified feature representation space for multi-source heterogeneous data, solving the problem of lacking unified representation after data fusion, enabling data from different sources to be analyzed and processed in the same space. Finally, the unified behavioral feature representation vector is input into an adversarial analysis system with a generator-discriminator dual-network architecture to obtain accurate and reliable user behavior data analysis results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a user behavior data analysis method based on multi-agent interconnection provided in an embodiment of this disclosure.

[0021] Figure 2 This is a flowchart illustrating the method for constructing a spatiotemporal correlation map provided in this embodiment of the disclosure.

[0022] Figure 3 A flowchart illustrating the method for constructing a personalized causal knowledge graph provided in this embodiment of the disclosure.

[0023] Figure 4 This is a schematic diagram of the architecture for a multi-agent device to collect raw behavioral data of users in different scenarios, provided in an embodiment of this disclosure.

[0024] Figure 5This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0025] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0026] Reference Figure 1 This application discloses a user behavior data analysis method based on multi-agent interconnection, including:

[0027] S100 constructs a spatiotemporal correlation map based on raw user behavior data collected by multi-agent devices in different scenarios.

[0028] S200 extracts multimodal features from the spatiotemporal correlation map to obtain multidimensional, multi-source heterogeneous feature vectors with spatiotemporal labels.

[0029] S300 performs in-depth causal relationship analysis between cross-device data based on multi-dimensional, multi-source, heterogeneous feature vectors to obtain an individualized causal association knowledge graph.

[0030] S400, based on a personalized causal knowledge graph, designs a three-layer progressive attention gating network architecture. Through a dynamic weight allocation mechanism driven by user personalized profiles, it maps multi-dimensional, multi-source heterogeneous feature vectors into unified behavioral feature representation vectors.

[0031] The S500 inputs a unified behavioral feature representation vector into an adversarial analysis system with a generator-discriminator dual-network architecture to obtain user behavior data analysis results.

[0032] This application discloses a user behavior data analysis method based on multi-agent interconnection. Through multi-step processing and analysis, it comprehensively considers multimodal user behavior data in different scenarios, enabling a comprehensive and accurate understanding of user behavior characteristics and potential causal relationships. Based on personalized user profiles and causal association knowledge graphs, it can provide personalized behavior predictions and suggestions for each user, improving the targeting and effectiveness of services. By employing a fairness-constrained loss function and gradient inversion layer technology, it eliminates analytical bias, ensuring the fairness and objectivity of the analysis results and avoiding unfairness caused by sensitive factors. This solution can be applied to multiple fields such as smart cities, education, and healthcare, providing strong support for decision-making and services in various sectors.

[0033] Reference Figure 2 The method for "constructing a spatiotemporal correlation graph based on raw user behavior data collected by multiple intelligent agent devices in different scenarios" in S100, i.e., the method for constructing the spatiotemporal correlation graph, specifically includes:

[0034] S110 establishes a heterogeneous collaborative data acquisition mechanism through several personal intelligent agent devices and several environmental intelligent agent devices to acquire raw behavioral data generated by users in various scenarios and generate raw data streams with timestamps.

[0035] "Acquiring raw behavioral data generated by users in multiple scenarios and generating raw data streams with timestamps" refers to collecting raw behavioral data of users in multiple scenarios through multiple types of devices, and forming a raw data stream with all raw behavioral data and their corresponding collection time (i.e., timestamps).

[0036] Specifically, sports watches are used to collect physiological data such as heart rate, steps, and sleep quality; smart blackboards are used to collect behavioral data such as writing trajectory, interaction frequency, and attention heatmap; and AR glasses are used to collect cognitive data such as eye tracking, facial expression recognition, and spatial positioning information.

[0037] The heterogeneous collaborative acquisition mechanism can comprehensively acquire user behavior data from multiple dimensions; the data collected by different types of intelligent agent devices covers multiple aspects such as physiology, behavior and cognition, enabling subsequent analysis to more comprehensively and accurately reflect the user's real state in different scenarios; the addition of timestamps gives the data a time dimension, which facilitates subsequent time series analysis and data correlation processing.

[0038] S120, based on the raw data stream with timestamps, constructs the initial map framework.

[0039] Specifically, based on the original data stream with timestamps, users are taken as core nodes and the data of each intelligent agent device are taken as data nodes to obtain an initial graph framework. The edges in the initial graph framework are used to connect the core nodes and the corresponding data nodes, representing the association between data and users.

[0040] For example, with students as the core node, heart rate data and step count data collected by sports watches, writing trajectory data and interaction frequency data collected by smart blackboards, and gaze tracking data and facial expression recognition data collected by AR glasses are connected to the core node as connecting edges, thereby constructing a preliminary graph framework.

[0041] The initial graph framework was constructed to organize the data collected by users and various intelligent devices in a structured manner, making the relationships between the data clearer. This visual structure helps to further analyze and process the data. The user-centric design highlights the user's central position in the entire data system, which facilitates various behavioral analyses and pattern mining around the user.

[0042] S130, based on the initial map framework, assigns timestamps, spatial coordinates, and behavioral label 3D attribute annotations to each data node, and determines the precise spatiotemporal location of data collection through a spatiotemporal positioning algorithm to obtain a node set.

[0043] For example, for heart rate data nodes collected by sports watches, spatial coordinates are assigned based on the student's location when wearing the watch (such as classroom, playground, etc.), and a timestamp is assigned based on the collection time. At the same time, behavioral labels are marked according to the level and changes in heart rate, such as "exercising" and "resting". For writing trajectory data nodes collected by smart blackboards, spatial coordinates are determined based on the specific location of the classroom and seat information, and a timestamp is determined based on the class time. Behavioral labels are marked according to the content and frequency of writing, such as "taking notes carefully" and "writing casually". Through spatiotemporal positioning algorithms, the precise spatiotemporal location of each data node is determined using GPS positioning and indoor positioning technology, and finally a complete set of nodes with attributes is obtained.

[0044] Three-dimensional attribute annotation provides richer information for data nodes, enabling each data node to contain not only the original data but also information in dimensions such as time, space, and behavioral characteristics. This helps to understand the meaning behind the data more deeply and to uncover the potential relationships between data. Precise spatiotemporal positioning enables data to be accurately correlated and analyzed in the spatiotemporal dimension, providing a foundation for building more complex relationships in the future.

[0045] S140, based on a set of nodes, uses a graph data structure organization algorithm to establish the relationships between nodes, and determines the node connection weights through similarity calculation and cluster analysis to obtain a structured spatiotemporal relationship graph.

[0046] Graph data structure organization algorithms are used to organize and manage nodes and edges in a graph to reveal the relationships between nodes. A graph is a data structure composed of nodes (vertices) and edges connecting these nodes, and it is very useful when dealing with complex relational data. Common graph data structure organization algorithms include Breadth-First Search (BFS), Depth-First Search (DFS), Dijkstra's algorithm, and the Floyd-Warshall algorithm.

[0047] However, in constructing spatiotemporal association graphs, we focus more on discovering potential relationships between nodes. The following explanation uses association rule mining algorithms as an example. Assume we already have a dataset containing node data, such as user heart rate, step count, and facial expression recognition data. Each data record has a timestamp to reflect its spatiotemporal characteristics; for continuous data (such as heart rate and step count), discretization is required to convert it into discrete intervals. For example, heart rate is divided into three intervals: "low," "medium," and "high." Noisy data and missing values ​​are removed to ensure data quality. Then, frequent itemsets are generated. Frequent itemsets are sets of items that appear frequently in the dataset. The Apriori algorithm can be used to generate frequent itemsets, with the following steps: 1) Initialization: Set a minimum support threshold to filter frequent itemsets; 2) Generate 1-itemsets: Count the frequency of each node's data and filter out 1-itemsets that meet the minimum support; 3) Generate k-itemsets (k>1): Generate k-itemsets by merging k-1 itemsets and calculate their support, filtering out k-itemsets that meet the minimum support, until no new frequent itemsets can be generated. Association rules are generated based on the frequent itemsets, and the confidence of the rules is calculated. Confidence represents the probability that the consequent is included if the antecedent is included. A minimum confidence threshold is set to filter out association rules that meet the conditions. Based on the generated association rules, the association relationships between nodes are established.

[0048] A graph data structure organization algorithm is used to analyze the relationships between nodes in a node set. For example, by analyzing the relationship between heart rate data nodes and step count data nodes, it is found that heart rate usually increases when step count increases, thus establishing a relationship between these two nodes.

[0049] By calculating similarity and clustering analysis, similarity scores are calculated between different nodes, and node connection weights are determined based on the scores. For example, if the similarity between heart rate data nodes and step count data nodes is high, the connection weight will be large; while if the similarity between heart rate data nodes and facial expression recognition data nodes is low, the connection weight will be small. The weights can also be set according to specific needs.

[0050] Establishing relationships between nodes and determining connection weights can reveal the intrinsic connections and mutual influence between data. By analyzing these relationships and weights, patterns and trends in user behavior can be discovered, providing support for applications such as personalized recommendations and behavior prediction. Structured spatiotemporal relationship graphs present complex data relationships in an intuitive graphical way, making it easier for users and analysts to understand and analyze, thus improving the efficiency and accuracy of data analysis.

[0051] Reference Figure 3The S300 method of "conducting deep causal relationship analysis between cross-device data based on multi-dimensional, multi-source, heterogeneous feature vectors to obtain a personalized causal association knowledge graph," specifically includes the following:

[0052] S310, based on multidimensional, multi-source heterogeneous feature vectors, identifies direct causal relationships, indirect causal relationships, and conditional causal relationships between data across devices.

[0053] S320, based on direct causal relationships, indirect causal relationships, and conditional causal relationships, obtains the set of time-series causal chains and the distribution of causal strength confidence.

[0054] Specifically, the identified direct, indirect, and conditional causal relationships are merged and connected chronologically to form a set of time-series causal chains. For each causal relationship, causal effect estimation methods (such as propensity score matching, instrumental variable methods, etc.) can be used to calculate the causal strength, which represents the degree of influence of one variable on another. Confidence intervals for causal strength are calculated using sampling or statistical methods to obtain the confidence distribution of causal strength.

[0055] S330, based on the set of temporal causal chains and the distribution of causal strength confidence, constructs a causal relationship coding model that integrates individual characteristics, and obtains an individualized causal association knowledge graph.

[0056] This embodiment, through the analysis of deep causal relationships between cross-device data, enables a profound understanding of the interaction mechanisms between data from different devices, revealing the causal patterns hidden behind the data. Personalized causal association knowledge graphs provide a basis for personalized services and decision-making. For example, in smart homes, based on a user's personalized causal association knowledge graph, device control suggestions can be provided to better meet the user's needs; in the medical field, personalized treatment plans can be developed for each patient. Understanding causal relationships helps in system optimization and prediction. For example, in intelligent transportation systems, traffic signal control can be optimized based on causal association knowledge graphs to predict traffic congestion and improve traffic efficiency.

[0057] Furthermore, the method S330, which "constructs a causal relationship coding model that integrates individual characteristics based on the set of temporal causal chains and the distribution of causal strength confidence, to obtain an individualized causal association knowledge graph," specifically includes:

[0058] S331 constructs a hybrid graph neural architecture that integrates graph convolutional networks and graph attention networks, based on the set of temporal causal chains and the distribution of causal strength confidence.

[0059] S332 uses GraphSAGE as the basic framework for graph convolution and combines it with the multi-head attention mechanism of GAT (Graph Attention Network). It constructs a 3-layer GCN-GAT hybrid encoder through residual connections to perform deep encoding of causal relationships and obtain the output information of the hybrid encoder.

[0060] GraphSAGE is used as the basic framework for graph convolution. The GraphSAGE framework includes two main operations: sampling and aggregation. In the sampling stage, for each node in the graph, a certain number of neighboring nodes are selected from its neighboring nodes according to a preset strategy. In the aggregation stage, the features of the selected neighboring nodes are aggregated with the features of the current node to generate new features for the current node.

[0061] By combining GAT's multi-head attention mechanism, which uses multiple attention heads to compute in parallel, each attention head can learn different aspects of the relationship between nodes. During the computation process, different attention weights are assigned to each neighbor node, which reflect the importance of the neighbor node to the current node.

[0062] A three-layer GCN-GAT hybrid encoder is constructed using residual connections. Residual connections are used to address the vanishing gradient problem in deep neural networks, enabling the network to learn deeper features more effectively. In each GCN-GAT layer, graph convolution operations are first performed to update node features, and then a multi-head attention mechanism is used to further capture the dependencies between nodes. Through these three layers of processing, a deep encoding of causal relationships is achieved.

[0063] S333 is a graph structure update algorithm based on sliding time window, configured based on the output information of the hybrid encoder.

[0064] The sliding time window is used to analyze behavioral characteristics over a period of time. The size of the window can be set according to the actual application scenario and data characteristics.

[0065] S334, obtaining cosine similarity of feature vectors based on graph structure update algorithm;

[0066] When the cosine similarity is lower than the first preset threshold (preferably 0.8), graph topology reconstruction is triggered. New nodes and edges are added using an incremental learning method, and redundant connections with edge weights lower than the second preset threshold (preferably 0.1) are deleted to obtain the dynamically reconstructed graph structure.

[0067] For S333-S334, based on the output of the aforementioned hybrid encoder, a graph structure update algorithm based on a sliding time window is designed. The sliding time window is used to analyze behavioral features over a period of time, and the window size is set according to the actual application scenario and data characteristics. The cosine similarity of the behavioral feature vectors is calculated. When the cosine similarity of the behavioral feature vectors is lower than a threshold of 0.8, graph topology reconstruction is triggered. When calculating the cosine similarity, for each node's behavioral feature vector in the graph, the cosine similarity is calculated with the behavioral feature vectors of other nodes. New nodes and edges are added using an incremental learning approach. When a new causal relationship or individual behavior is detected, the corresponding new node is added to the graph, and corresponding edges are added according to the causal relationship. Simultaneously, redundant connections with weights lower than 0.1 are deleted. By traversing all edges in the graph and checking the weight of each edge, edges with weights lower than the threshold are removed from the graph, thereby achieving dynamic reconstruction of the graph structure.

[0068] S335, based on the dynamically reconstructed graph structure, maps the preset multi-dimensional individual features into 64-dimensional individual embedding vectors through MLP, and then splices and fuses them with causal relationship features to obtain an individualized causal association knowledge graph.

[0069] Based on the dynamically reconstructed graph structure, individual features such as cognitive style, learning preferences, circadian rhythms, and social patterns are input into an MLP (Multi-Layer Perceptron). An MLP is a feedforward neural network containing input, hidden, and output layers. Through multiple layers of nonlinear transformations, it maps the input individual features into 64-dimensional individual embedding vectors. These 64-dimensional individual embedding vectors are then concatenated with causal relationship features. Specifically, the individual embedding vectors and causal relationship feature vectors are joined together in dimensional order to form a new feature vector. Based on this concatenated and fused feature vector, an individualized causal association knowledge graph is constructed, reflecting the association information between individual features and causal relationships.

[0070] Furthermore, a distributed computing architecture can be adopted to divide the data and distribute it across multiple computing nodes for parallel processing. Communication and synchronization mechanisms can be designed to coordinate each node to complete the computing tasks in the above steps and optimize the processing.

[0071] The S400 method, which "designs a three-layer progressive attention-gated network architecture based on a personalized causal knowledge graph and maps multi-dimensional, multi-source heterogeneous feature vectors into unified behavioral feature representation vectors through a dynamic weight allocation mechanism driven by user personalized profiles," specifically includes:

[0072] S410, based on a personalized causal knowledge graph, designs a three-layer progressive attention gating network architecture.

[0073] The first layer is a feature-level attention gating, which uses a self-attention mechanism to calculate the importance weights of feature vectors in each dimension; the second layer is a modality-level attention gating, which integrates physiological, behavioral, and cognitive modal data through a cross-attention mechanism; the third layer is a temporal-level attention gating, which uses a Transformer encoder to capture feature evolution trends and outputs a hierarchical attention weight matrix.

[0074] Specifically, based on a personalized causal knowledge graph, a self-attention mechanism is employed in the first layer. This mechanism calculates the correlation between feature vectors across different dimensions and assigns importance weights to each dimension. Specifically, for the input feature vector, a linear transformation is first applied to map it to three spaces: query, key, and value. Then, the similarity score between the query and the key is calculated, and the score is converted into a probability distribution using a softmax function. This distribution serves as the importance weight for each dimension of the feature vector. Finally, the values ​​are weighted and summed according to these weights to obtain the feature representation after feature-level attention gating.

[0075] The second layer employs a cross-attention mechanism to fuse data from three modalities: physiological, behavioral, and cognitive. This mechanism focuses on the relationships between different modalities, using one modality as the query and the others as keys and values ​​to calculate cross-modal attention weights. This approach captures complementary information between different modalities, enabling modal data fusion and resulting in feature representations processed with modality-level attention gating.

[0076] The third layer utilizes a Transformer encoder to capture feature evolution trends. The Transformer encoder consists of a multi-head attention mechanism and a feedforward neural network. The multi-head attention mechanism can capture the dependencies between features from different perspectives in parallel, while the feedforward neural network performs a non-linear transformation on the output of the multi-head attention mechanism. After processing by the Transformer encoder, a hierarchical attention weight matrix is ​​output, reflecting the importance information of features at different levels and dimensions.

[0077] S420, based on a hierarchical attention weight matrix, constructs a dynamic weight allocation mechanism driven by user personalized profiles and outputs personalized dynamic weight parameters.

[0078] Specifically, based on the hierarchical attention weight matrix obtained above, a dynamic weight allocation mechanism driven by user personalization profiles is constructed. This mechanism considers individual characteristics such as user behavioral preferences, cognitive patterns, and activity patterns. By inputting these individual user characteristics into the meta-learning network, the network can learn general learning strategies from multiple tasks and adaptively adjust the attention weights of each layer according to the user's individual characteristics. Specifically, the meta-learning network continuously optimizes its internal parameters so that the output weights can better adapt to the personalized needs of different users, ultimately outputting personalized dynamic weight parameters.

[0079] S430, based on personalized dynamic weight parameters, uses residual connection and layer normalization techniques to map multi-dimensional, multi-source heterogeneous feature vectors into unified behavioral feature representation vectors through nonlinear transformation.

[0080] Based on the obtained personalized dynamic weight parameters, residual connections and layer normalization techniques are employed to process multi-dimensional, multi-source heterogeneous feature vectors. Residual connections address the vanishing gradient problem in deep neural networks, allowing the network to directly learn the residual between the input and output; that is, the output equals the input plus a residual term after nonlinear transformation. Layer normalization normalizes the features of each sample, ensuring similar distributions across different dimensions and improving model stability and training efficiency. Nonlinear transformation maps the multi-dimensional, multi-source heterogeneous feature vectors to a unified behavioral feature representation space, resulting in a unified behavioral feature representation vector. This vector comprehensively considers the user's personalized information and multi-source heterogeneous features, more accurately reflecting the user's characteristics and behavioral patterns.

[0081] Furthermore, the method of S420, which "constructs a dynamic weight allocation mechanism driven by user personalized profiles based on a hierarchical attention weight matrix and outputs personalized dynamic weight parameters," specifically includes:

[0082] S421, based on a hierarchical attention weight matrix, extracts individual features of users including behavioral preferences, cognitive patterns, and activity patterns. Through feature engineering methods, it constructs multi-dimensional user profile feature vectors to obtain personalized user profile data.

[0083] Assuming the hierarchical attention weight matrix is ​​a three-dimensional matrix, with different dimensions corresponding to different feature levels, time steps, and feature types, for example, in an online learning platform scenario, by analyzing this matrix, we can extract individual characteristics such as the distribution of learning time (behavioral preferences) across different course modules, the speed of mastering knowledge points of varying difficulty (cognitive patterns), and fixed daily learning time periods (activity patterns). For the extracted individual characteristics, feature selection is first performed; for example, among numerous behavioral preference features, only those highly correlated with learning effectiveness are selected, such as course completion rate and homework accuracy. Then, feature transformation is performed, standardizing features of different magnitudes, such as converting learning time from minutes to hours and applying z-score standardization to make the features comparable. Finally, feature combination is performed, combining learning time and homework accuracy into a new feature, "learning efficiency." Through these operations, a multi-dimensional user profile feature vector is constructed, thereby obtaining personalized user profile data.

[0084] Extracting individual features from a hierarchical attention weight matrix allows for deeper exploration of users' potential information, as the matrix has quantified the importance of different features. Feature engineering methods can improve the quality and usability of features, remove redundant information, and enhance the discriminative power between features, making the constructed personalized user profile data more accurate and comprehensive, and providing a more reliable basis for subsequent weight allocation.

[0085] S422, based on user personalized profile data, designs a meta-learning network architecture. It quickly adapts to the personalized characteristics of different users through a few-shot learning mechanism, and uses the gradient descent algorithm to optimize the meta-learning parameters to obtain user-specific weight learning results.

[0086] The meta-learning network architecture is designed based on a multilayer perceptron (MLP). This network consists of an input layer, a hidden layer, and an output layer. The input layer receives personalized user profile data, the hidden layer performs nonlinear transformations on the input data through multiple neurons, and the output layer outputs user-specific weights.

[0087] Few-shot learning mechanism: In online education scenarios, newly registered users may only have a small amount of learning behavior data. Meta-learning networks utilize the few-shot learning mechanism to learn general learning strategies and feature representations from a large amount of existing user data, and quickly adapt to the personalized characteristics of new users.

[0088] The meta-learning parameters are optimized using the stochastic gradient descent (SGD) algorithm. In each iteration, a subset of user data is randomly selected as a mini-batch, and the gradient of the loss function (such as mean squared error loss) with respect to the meta-learning parameters is calculated. Then, the parameters are updated in the opposite direction of the gradient. After multiple iterations, the loss function gradually decreases, resulting in the optimized meta-learning parameters and the acquisition of user-specific weight learning results.

[0089] In this step, the few-shot learning mechanism of the meta-learning network enables the system to quickly provide personalized services to new users with limited data, reducing the dependence on large amounts of data and improving the system's adaptability and flexibility. The gradient descent algorithm can effectively optimize the meta-learning parameters, enabling the meta-learning network to accurately learn the specific weights of different users, providing precise guidance for subsequent attention weight adjustments.

[0090] S423 adaptively adjusts the attention weights at the feature level, modality level, and temporal level based on user-specific weight learning results, and obtains personalized dynamic weight parameters through normalization processing.

[0091] Specifically, based on the user-specific weight learning results, the attention weights at the feature level, modality level, and temporal level are adjusted. The adjusted attention weights are then normalized using the softmax function, which converts the adjusted weights into a probability distribution, ensuring that all weight values ​​are within the range [0,1] and the sum of the weights is 1. For example, the feature-level attention weight vector [0.2, 0.3, 0.5] becomes [0.23, 0.32, 0.45] after softmax processing, ensuring the reasonableness of the weight distribution.

[0092] Adaptive adjustment of attention weights can dynamically allocate attention to each layer based on the user's personalized characteristics, enabling the system to focus more on important information relevant to the user and improving the model's relevance and accuracy. Normalization ensures the rationality and interpretability of the weights, making the attention weights of different layers comparable, while also conforming to the characteristics of probability distribution, which facilitates subsequent model calculation and decision-making.

[0093] The S500 method of "inputting a unified behavioral feature representation vector into an adversarial analysis system with a constructed generator-discriminator dual-network architecture to obtain user behavior data analysis results" specifically includes:

[0094] S510 constructs an adversarial analysis system with a generator-discriminator dual-network architecture. A unified behavioral feature representation vector is input into the adversarial analysis system to obtain bias detection results.

[0095] The generator network processes the input feature vector through a multi-head attention mechanism and residual fully connected layers, outputting scores across five dimensions: cognitive development, physical health, digital literacy, social-emotional engagement, and innovative practice. Specifically, the generator network employs a multi-head attention mechanism, which can extract feature information from different representation subspaces in parallel, capturing the dependencies between different parts of the feature vector and enhancing the understanding and expression of input features. Simultaneously, it combines the processing of the input feature vector with residual fully connected layers, which alleviate the gradient vanishing problem in deep networks, enabling the network to learn more complex feature mapping relationships. After processing by the generator network, the five-dimensional scores of cognitive development, physical health, digital literacy, social-emotional engagement, and innovative practice are output.

[0096] The discriminator network receives rating results and user group labels, uses the Wasserstein distance loss function to detect group bias in the ratings, and outputs the bias detection results and the initial rating matrix. Specifically, the Wasserstein distance loss function is used, which can more effectively measure the difference between the generated data distribution and the real data distribution, and is used to detect group bias in the ratings. The discriminator network outputs the bias detection results and the initial rating matrix by analyzing the rating results and user group labels.

[0097] S520, based on the bias detection results, designs a fairness constraint loss function, combines statistical equality, equal opportunity, and calibration fairness indicators, embeds fairness constraints into the optimization objective through the Lagrange multiplier method, and uses gradient inversion layer technology to invert the discriminator gradient during the backpropagation process of the generator, outputting a five-dimensional score result that has been calibrated for fairness.

[0098] Specifically, based on the bias detection results output by the discriminator network, a fairness constraint loss function is designed. This loss function combines three fairness indicators: statistical equality, equal opportunity, and calibration. Statistical equality requires that different groups have the same distribution in terms of scores; equal opportunity emphasizes that different groups have equal opportunities to obtain a specific score; and calibration ensures the consistency between the score and the actual behavioral performance.

[0099] The Lagrange multiplier method embeds fairness constraints into the optimization objective. The Lagrange multiplier method is a method for solving optimization problems under constraints. It transforms the constraints into part of the objective function, so that fairness constraints can be considered simultaneously during the optimization process.

[0100] During the backpropagation of the generator, a gradient inversion layer technique is used to invert the gradient of the discriminator. The gradient inversion layer enables the generator to optimize in the direction of reducing group bias during the learning process, thereby improving the fairness of the scoring and finally outputting a five-dimensional scoring result that has been calibrated for fairness.

[0101] S530 constructs a personalized behavior prediction module based on reinforcement learning, using the five-dimensional scoring results as the state space, the predicted measures as the action space, and the behavior improvement effect as the reward function. It outputs targeted behavior suggestions through the Actor-Critic algorithm to obtain user behavior data analysis results.

[0102] Specifically, based on the five-dimensional scoring results calibrated for fairness, a personalized behavior prediction module based on reinforcement learning is constructed. In this module, the five-dimensional scores are used as the state space, which describes the user's current performance in each dimension; the predicted measures are used as the action space, representing the different behavior improvement measures that can be taken; and the behavior improvement effect is used as the reward function to evaluate the degree of improvement in user behavior after taking a certain action.

[0103] The Actor-Critic algorithm is used for learning and decision-making. The Actor network is responsible for selecting the appropriate action based on the current state, while the Critic network is responsible for evaluating the value of the action selected by the Actor network. By continuously interacting with the environment, the Actor-Critic algorithm can learn the optimal strategy, that is, a personalized behavior prediction strategy for each user, and finally output targeted behavior suggestions, resulting in user behavior data analysis results.

[0104] The S530 method, which "builds a personalized behavior prediction module based on reinforcement learning based on five-dimensional scoring results, using the five-dimensional scoring as the state space, the prediction measures as the action space, and the behavior improvement effect as the reward function, and outputs targeted behavior suggestions through the Actor-Critic algorithm to obtain user behavior data analysis results," specifically includes:

[0105] S531, based on the five-dimensional scoring results, establishes a reinforcement learning environment model.

[0106] Taking online education platforms as an example, the five-dimensional scoring results reflect students' performance in cognitive development, physical health, digital literacy, social-emotional development, and innovative practice. This platform can be viewed as a reinforcement learning environment in which students act as intelligent agents. The environment model sets the initial state of the agent, i.e., the student's current five-dimensional scores, and defines the boundaries and rules of the environment, such as the score range for each dimension being 0-100 points. The environment model also simulates the possible changes in the five-dimensional scores after students take different actions (such as participating in learning activities or engaging in physical exercise). For example, if a student participates in a math competition (a recommended learning activity), the score in the cognitive development dimension may increase accordingly based on the competition results. Establishing a reinforcement learning environment model provides a clear framework for subsequent personalized behavior prediction, enabling the entire prediction process to take place within a defined environment.

[0107] S532 uses five dimensions of scores—cognitive development, physical health, digital literacy, social-emotional connection, and innovative practice—as the state space to construct a continuous state representation.

[0108] Continuous state representation can more accurately describe a user's state in various dimensions, avoiding the information loss that may result from discrete state representation. It also makes the differences between states more intuitive, which is beneficial for reinforcement learning algorithms to better learn the relationship between states and actions and improve the accuracy of predictions.

[0109] S533, based on continuous state representation, designs a behavior intervention action space, and maps discrete actions into continuous vector representations through an action encoding mechanism to obtain the action space of continuous vector representation;

[0110] The behavioral intervention action space includes learning activity recommendations, exercise program development, social interaction suggestions, and guidance on innovative practices.

[0111] For example, in online education platforms, learning activity recommendations could include attending online courses and reading professional books; exercise plans could include three aerobic exercise sessions per week and a certain amount of strength training per day; social interaction suggestions could include participating in study group discussions and collaborating on projects with classmates; and guidance for innovative practices could include participating in scientific research projects and innovation competitions. Regarding action encoding mechanisms, taking learning activity recommendations as an example, suppose there are three learning activities: a1 (attending online courses), a2 (reading professional books), and a3 (attending academic lectures). One-hot encoding can be used to map these discrete actions into continuous vectors; for example, a1 can be encoded as (1,0,0), a2 as (0,1,0), and a3 as (0,0,1). For the entire behavioral intervention action space, after encoding all discrete actions, a continuous vector representation of the action space is obtained. Designing a rich behavioral intervention action space provides users with more choices and can meet the personalized needs of different users. The action encoding mechanism converts discrete actions into continuous vector representations, enabling reinforcement learning algorithms to perform mathematical operations and comparisons on actions, facilitating the selection of the optimal action.

[0112] S534 constructs a reward function based on the action space represented by continuous vectors, aiming at improving behavior. It uses the Actor-Critic algorithm for policy optimization. The Actor network outputs the action probability distribution, the Critic network estimates the state value function, and the network parameters are updated through the policy gradient method to obtain the user behavior data analysis results.

[0113] Specifically, the reward function R can be defined based on the changes in the five-dimensional ratings. For example, if the sum of the five-dimensional ratings increases after a student takes an action, the reward function is positive; conversely, if the sum of the ratings decreases, the reward function is negative. Specifically, assume the five-dimensional rating vector before taking the action is s = (x1, x2, x3, x4, x5), and the rating vector after taking the action is si. ' =(x1) ' x2 ' x3 ' x4 ' x5 ' Then the reward function

[0114] For policy optimization in the Actor-Critic algorithm, the Actor network outputs the action probability distribution based on the current state. For example, in state s, the probability of choosing action a1 is P(a1|s), and the probability of choosing action a2 is P(a2|s), etc. The Critic network estimates the state value function V(s) to evaluate the quality of the current state. Using the policy gradient method, the parameters of the Actor and Critic networks are updated based on the reward function value, enabling the Actor network to learn a better action selection policy. After multiple iterations of optimization, based on the learned optimal policy, targeted behavioral suggestions are output to the user, such as "It is recommended that you participate in online programming courses to improve your digital literacy." These suggestions constitute the user behavior data analysis results. The reward function, aimed at improving behavior, provides the agent with a clear learning objective, prompting the agent to choose actions that improve user behavior. Furthermore, the Actor-Critic algorithm, combined with the policy gradient method, effectively utilizes state and reward information for policy optimization, improving learning efficiency and policy quality. The resulting user behavior data analysis results are highly targeted and practical, helping users improve their behavior.

[0115] The user behavior data analysis method disclosed in this application, based on multi-agent interconnection, collects raw user behavior data in different scenarios using multi-agent devices, breaking the limitations of traditional single data sources. In educational scenarios, multi-agent devices can be distributed across multiple learning environments such as classrooms, libraries, and homes, comprehensively collecting data on students' learning behaviors, social interactions, interests, and other aspects. For example, smart bracelets can record students' exercise and daily routines, smart cameras can capture students' expressions and participation in class, and smart learning tablets can collect data on students' homework completion and online learning records. By constructing a spatiotemporal correlation graph, these multi-source heterogeneous data are integrated, providing a rich and comprehensive data foundation for the comprehensive evaluation of students' qualities.

[0116] Extracting multimodal features from spatiotemporal correlation maps yields multidimensional, multi-source, heterogeneous feature vectors with spatiotemporal labels. This multimodal feature extraction method further enriches the data dimensions, considering not only the content of the data but also combining temporal and spatial information. For example, when analyzing students' learning behavior, it allows for simultaneous consideration of students' learning performance at different times and locations, leading to a more accurate assessment of students' learning status and development trends.

[0117] To address the issues of poor timeliness and long evaluation cycles in traditional educational assessments, this application proposes a lightweight model for real-time data processing and feedback at the edge. In this method, multi-agent devices can perform preliminary processing on collected data at the edge and promptly feed the results back to the system. For example, intelligent learning devices can analyze students' homework completion in real time, identify problems, promptly remind students, and provide corresponding learning suggestions. This real-time data processing and feedback mechanism significantly shortens the assessment cycle, enabling teachers and students to understand learning progress promptly and adjust teaching and learning strategies accordingly. By conducting deep causal relationship analysis on multi-dimensional, multi-source, heterogeneous feature vectors across device data, causal connections between different behavioral data can be quickly discovered. In educational assessment, this causal relationship analysis can help teachers promptly identify key factors affecting student academic performance, such as study time, learning methods, and social activities, thereby enabling targeted intervention and guidance.

[0118] Addressing the issue of isolated and uninterconnected data in traditional educational evaluation systems, this application emphasizes breaking down data silos through distributed intelligent agent networks to achieve cross-scenario data fusion. In this method, data collected by multiple intelligent agent devices from different scenarios can be integrated through spatiotemporal correlation graphs, eliminating barriers between data systems. For example, data from a school's teaching management system, students' home learning devices, and online learning platforms can be interconnected, enabling data sharing and interaction. This allows teachers to comprehensively understand students' learning progress in different scenarios and conduct more comprehensive and objective evaluations.

[0119] This application constructs a dynamic weighted evaluation model to adapt to the needs of different educational stages. In this method, based on an individualized causal knowledge graph, a three-layer progressive attention-gated network architecture is designed. Through a dynamic weight allocation mechanism driven by personalized user profiles, multi-dimensional, multi-source, heterogeneous feature vectors are mapped into unified behavioral feature representation vectors. In educational evaluation, this dynamic weight allocation mechanism can assign different weights to different evaluation indicators according to the characteristics and needs of different educational stages. In summary, the user behavior data analysis method based on multi-agent interconnection has significant advantages in solving traditional educational evaluation problems, providing more comprehensive, accurate, and real-time information for educational evaluation, and promoting the personalized and intelligent development of education and teaching.

[0120] Furthermore, referring to Figure 4Raw user behavior data collected through multi-agent devices in different scenarios can be acquired by several wearable and environmental intelligent agent devices. Wearable intelligent agent devices, such as sports watches / AR glasses, are used to collect and analyze the first type of multi-dimensional behavioral data when worn by the user to obtain first-dimensional information. This first type of multi-dimensional behavioral data includes physiological indicators, such as real-time heart rate changes and exercise intensity indices. Environmental intelligent agent devices, such as smart blackboards / learning machines, are used to collect and analyze the second type of multi-dimensional behavioral data in the target environment to obtain second-dimensional information. This second type of multi-dimensional behavioral data includes learning behavior information, which includes temporal changes in engagement, heatmaps, attention information, and temporal data on concentration fluctuations.

[0121] The second type of intelligent agent communicates with the second type of behavioral data acquisition module (such as environmental configuration devices like blackboards, AR glasses, and learning machines) to receive second-type multi-dimensional behavioral data. In a classroom setting, the blackboard intelligent agent analyzes participation through eye tracking and touch trajectory. After coarse processing of the raw data, the second type of intelligent agent generates temporal variation data and heatmaps of participation. The AR glasses intelligent agent records learning progress and attention focus dwell time. After coarse processing of the raw video stream data, the second type of intelligent agent analyzes and extracts attention information such as tag nodes and attention focus coordinates. In a home setting, the learning machine uses handwriting pressure sensing to monitor writing focus, learning progress, and difficulty, and analyzes learning emotions through emotion recognition. After coarse processing of the handwriting pressure data from the learning machine, the second type of intelligent agent analyzes and obtains temporal data of focus fluctuations.

[0122] Imagine a school setting where students wear smartwatches, use a smart blackboard in the classroom, and wear AR glasses while moving around campus. The smartwatches continuously record students' physiological data throughout the day, such as heart rate and steps, generating a data record every 5 minutes with a timestamp. The smart blackboard collects behavioral data such as students' writing trajectory and the frequency of their interactions with teachers and classmates during class, summarizing the data and adding a timestamp at the end of each class. The AR glasses collect students' gaze tracking and facial expression recognition data in real time, generating a timestamped data record every 10 seconds. The data collected by these different devices is then aggregated to form a raw data stream with timestamps.

[0123] Different types of intelligent agent devices are networked through a heterogeneous network and can connect to the intelligent agent control platform in the cloud. The intelligent agent control platform establishes a first downlink and a second downlink with the personal intelligent agent and the environmental intelligent agent, respectively. The personal intelligent agent and the environmental intelligent agent establish a first uplink and a second uplink with the intelligent agent control platform, respectively. The intelligent agent control platform can analyze the multi-dimensional behavioral data sent by several intelligent agents and obtain data analysis results.

[0124] Furthermore, multi-agent devices cover sports scenarios, classroom scenarios, and home scenarios; sports scenarios can correspond to several personal devices, classroom scenarios can correspond to several teaching environment devices, and home scenarios can correspond to several home environment devices.

[0125] Secondly, this application discloses a user behavior data analysis system based on multi-agent interconnection, used to execute the user behavior data analysis method based on multi-agent interconnection disclosed in the first aspect of this application, specifically including:

[0126] The data acquisition module is used to construct a spatiotemporal correlation map based on the raw behavioral data of users in different scenarios collected by multiple intelligent agent devices;

[0127] The multi-source heterogeneous feature vector extraction module is used to extract multimodal features from the spatiotemporal correlation map and obtain multidimensional multi-source heterogeneous feature vectors with spatiotemporal labels.

[0128] The personalized causal association knowledge graph acquisition module is used to perform deep causal relationship analysis between cross-device data based on multi-dimensional, multi-source, heterogeneous feature vectors to obtain a personalized causal association knowledge graph.

[0129] The unified behavioral feature representation vector acquisition module is used to design a multi-layer progressive attention gating network architecture based on the individualized causal association knowledge graph. Through a dynamic weight allocation mechanism driven by user personalized profiles, it maps multi-dimensional and multi-source heterogeneous feature vectors into unified behavioral feature representation vectors.

[0130] The analysis module is used to input the unified behavioral feature representation vector into the constructed generator-discriminator dual-network architecture adversarial analysis system to obtain user behavior data analysis results.

[0131] A computer device according to an embodiment of this disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0132] The processor may be a central processing unit (CPU) or other processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the user behavior data analysis method based on multi-agent interconnection described in the foregoing embodiments of this disclosure.

[0133] like Figure 5 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 5 The computer device illustrated is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein. The computer device may include a processor (e.g., a central processing unit, a graphics processing unit, etc.) that performs various appropriate actions and processes based on a program stored in read-only memory (ROM) or a program loaded from a storage device into random access memory (RAM). Various programs and data required for the operation of the computer device are also stored in the RAM. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0134] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 5 A computer apparatus with various devices is shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or included alternatively.

[0135] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for analyzing user behavior data based on multi-agent interconnection, characterized in that, include: A spatiotemporal correlation map is constructed based on the raw behavioral data of users in different scenarios collected by multi-agent devices. Multimodal features are extracted from the spatiotemporal correlation map to obtain a multidimensional, multi-source heterogeneous feature vector with spatiotemporal labels; Based on the multidimensional, multi-source, heterogeneous feature vectors, a deep causal relationship analysis is performed between cross-device data to obtain an individualized causal association knowledge graph; Based on the individualized causal knowledge graph, a multi-layer progressive attention gating network architecture is designed. Through a dynamic weight allocation mechanism driven by user personalized profiles, the multi-dimensional, multi-source heterogeneous feature vectors are mapped into unified behavioral feature representation vectors. The unified behavioral feature representation vector is input into the constructed generator-discriminator dual-network architecture adversarial analysis system to obtain user behavior data analysis results; The step of performing deep causal relationship analysis between cross-device data based on the multidimensional, multi-source, heterogeneous feature vectors to obtain an individualized causal association knowledge graph includes: constructing a multi-layer causal reasoning network based on deep learning based on the multidimensional, multi-source, heterogeneous feature vectors to identify direct causal relationships, indirect causal relationships, and conditional causal relationships between cross-device data; obtaining a set of temporal causal chains and a causal strength confidence distribution based on the direct causal relationships, the indirect causal relationships, and the conditional causal relationships; and constructing a causal relationship encoding model that integrates individual features based on the set of temporal causal chains and the causal strength confidence distribution to obtain an individualized causal association knowledge graph. The process of constructing a causal relationship encoding model that integrates individual features based on the set of temporal causal chains and the causal strength confidence distribution to obtain an individualized causal association knowledge graph includes: constructing a hybrid graph neural architecture that integrates graph convolutional networks and graph attention networks based on the set of temporal causal chains and the causal strength confidence distribution; using GraphSAGE as the basic framework for graph convolution and combining it with the multi-head attention mechanism of GAT, constructing a 3-layer GCN-GAT hybrid encoder through residual connections to perform deep encoding of causal relationships and obtain hybrid encoder output information; configuring a graph structure update algorithm based on a sliding time window based on the hybrid encoder output information; obtaining the cosine similarity of feature vectors based on the graph structure update algorithm; obtaining a dynamically reconstructed graph structure based on the cosine similarity; and mapping the preset multi-dimensional individual features into 64-dimensional individual embedding vectors through MLP based on the dynamically reconstructed graph structure, concatenating and fusing them with the causal relationship features to obtain the individualized causal association knowledge graph.

2. The user behavior data analysis method based on multi-agent interconnection according to claim 1, characterized in that, The construction of a spatiotemporal correlation graph based on raw user behavior data collected by multi-agent devices in different scenarios includes: The system acquires raw behavioral data generated by users in various scenarios through a heterogeneous collaborative acquisition mechanism involving multiple intelligent devices and generates raw data streams with timestamps. Based on the original data stream with timestamps, an initial map framework is constructed; Based on the initial map framework, each data node is assigned a timestamp, spatial coordinates, and three-dimensional attribute annotation of behavioral labels. The precise spatiotemporal location of data collection is determined by a spatiotemporal positioning algorithm to obtain a set of nodes. Based on the node set, the association between nodes is established, and the node connection weights are determined through similarity calculation and cluster analysis to obtain a structured spatiotemporal association graph.

3. The user behavior data analysis method based on multi-agent interconnection according to claim 1, characterized in that, Based on the individualized causal knowledge graph, a three-layer progressive attention-gated network architecture is designed. Through a dynamic weight allocation mechanism driven by user personalized profiles, the multi-dimensional, multi-source heterogeneous feature vectors are mapped into unified behavioral feature representation vectors, including: Based on the individualized causal knowledge graph, a three-layer progressive attention gating network architecture is designed; The first layer is a feature-level attention gating, which uses a self-attention mechanism to calculate the importance weights of feature vectors in each dimension; the second layer is a modality-level attention gating, which fuses physiological, behavioral, and cognitive modal data through a cross-attention mechanism; the third layer is a temporal-level attention gating, which uses a Transformer encoder to capture feature evolution trends and outputs a hierarchical attention weight matrix. Based on the hierarchical attention weight matrix, a dynamic weight allocation mechanism driven by user personalized profiles is constructed, and personalized dynamic weight parameters are output. Based on the personalized dynamic weight parameters, residual connection and layer normalization techniques are used to map the multidimensional, multi-source heterogeneous feature vectors into a unified behavioral feature representation vector through nonlinear transformation.

4. The user behavior data analysis method based on multi-agent interconnection according to claim 3, characterized in that, The process involves constructing a dynamic weight allocation mechanism driven by the user's personalized profile based on the hierarchical attention weight matrix, and outputting personalized dynamic weight parameters, including: Based on the hierarchical attention weight matrix, individual features of users including behavioral preferences, cognitive patterns, and activity patterns are extracted. Multi-dimensional user profile feature vectors are constructed through feature engineering methods to obtain personalized user profile data. Based on the user personalized profile data, a meta-learning network architecture is designed. It quickly adapts to the personalized characteristics of different users through a few-shot learning mechanism, and uses the gradient descent algorithm to optimize the meta-learning parameters to obtain user-specific weight learning results. Based on the user-specific weight learning results, the attention weights at the feature level, modality level, and temporal level are adaptively adjusted, and personalized dynamic weight parameters are obtained through normalization processing.

5. The user behavior data analysis method based on multi-agent interconnection according to claim 1, characterized in that, The process of inputting the unified behavioral feature representation vector into the constructed generator-discriminator dual-network architecture adversarial analysis system to obtain user behavior data analysis results includes: An adversarial analysis system with a generator-discriminator dual-network architecture is constructed. The unified behavioral feature representation vector is input into the adversarial analysis system to obtain the deviation detection result. Based on the deviation detection results, a fairness constraint loss function is designed, and the fairness constraint is embedded into the optimization objective through the Lagrange multiplier method, outputting a five-dimensional score result after fairness calibration. Based on the five-dimensional scoring results, a personalized behavior prediction module based on reinforcement learning is constructed. The Actor-Critic algorithm is used to output targeted behavior suggestions, and user behavior data analysis results are obtained.

6. A user behavior data analysis system based on multi-agent interconnection, characterized in that, include: The data acquisition module is used to construct a spatiotemporal correlation map based on the raw behavioral data of users in different scenarios collected by multiple intelligent agent devices; The multi-source heterogeneous feature vector extraction module is used to extract multimodal features from the spatiotemporal correlation map to obtain multidimensional multi-source heterogeneous feature vectors with spatiotemporal labels; The individualized causal association knowledge graph acquisition module is used to perform deep causal relationship analysis between cross-device data based on the multi-dimensional multi-source heterogeneous feature vectors to obtain an individualized causal association knowledge graph. The step of performing deep causal relationship analysis between cross-device data based on the multidimensional, multi-source, heterogeneous feature vectors to obtain an individualized causal association knowledge graph includes: constructing a multi-layer causal reasoning network based on deep learning based on the multidimensional, multi-source, heterogeneous feature vectors to identify direct causal relationships, indirect causal relationships, and conditional causal relationships between cross-device data; obtaining a set of temporal causal chains and a causal strength confidence distribution based on the direct causal relationships, the indirect causal relationships, and the conditional causal relationships; and constructing a causal relationship encoding model that integrates individual features based on the set of temporal causal chains and the causal strength confidence distribution to obtain an individualized causal association knowledge graph. The process of constructing a causal relationship encoding model that integrates individual features based on the set of temporal causal chains and the causal strength confidence distribution to obtain an individualized causal association knowledge graph includes: constructing a hybrid graph neural architecture that integrates graph convolutional networks and graph attention networks based on the set of temporal causal chains and the causal strength confidence distribution; using GraphSAGE as the basic framework of graph convolution and combining it with the multi-head attention mechanism of GAT, constructing a 3-layer GCN-GAT hybrid encoder through residual connections to perform deep encoding of causal relationships and obtain hybrid encoder output information; configuring a graph structure update algorithm based on a sliding time window based on the hybrid encoder output information; obtaining the cosine similarity of feature vectors based on the graph structure update algorithm; obtaining a dynamically reconstructed graph structure based on the cosine similarity; and mapping the preset multi-dimensional individual features into 64-dimensional individual embedding vectors through MLP based on the dynamically reconstructed graph structure, concatenating and fusing them with the causal relationship features to obtain the individualized causal association knowledge graph. The unified behavioral feature representation vector acquisition module is used to design a multi-layer progressive attention gating network architecture based on the individualized causal association knowledge graph, and to map the multi-dimensional, multi-source heterogeneous feature vectors into unified behavioral feature representation vectors through a dynamic weight allocation mechanism driven by user personalized profiles. The analysis module is used to input the unified behavioral feature representation vector into the constructed generator-discriminator dual-network architecture adversarial analysis system to obtain user behavior data analysis results.

7. A computer device, characterized in that, The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the user behavior data analysis method based on multi-agent interconnection as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the user behavior data analysis method based on multi-agent interconnection as described in any one of claims 1-5.

Images