A user behavior data mining system and method based on multi-agent decision

Abstract

The application discloses a user behavior data mining system and method based on multi-agent decision, relates to the technical field of data mining, and comprises the following steps: acquiring original user behavior data, acquiring corresponding platform operation data and behavior environment data based on the original user behavior data, and constructing an environment agent based on the behavior environment data.The application realizes accurate classification of different environment states by extracting time and space multi-dimensional environment characteristics and clustering, realizes dynamic adaptive division of user behavior sessions by acquiring the probability distribution of user behavior time intervals under each environment category through a kernel density estimation method, realizes fine modeling of the user behavior decision process by constructing a user state space, realizes complete modeling of the multi-agent interactive decision process by constructing a three-agent joint Markov decision process, and ensures the accuracy of data mining.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, specifically to a user behavior data mining system and method based on multi-agent decision-making. Background Technology

[0002] User behavior data is one of the most core assets of internet platforms in the digital economy era, containing high-value information such as user needs, preferences, decision-making logic, and consumption habits. Precise and efficient user behavior data mining technology can extract users' true intentions from massive amounts of click, browsing, collection, sharing, and purchase records, providing core decision support for personalized recommendations, precision marketing, product iteration and optimization, user lifecycle management, and risk control. It can not only significantly improve the platform's user conversion rate, retention rate, and monetization capabilities, but also effectively alleviate information overload, providing users with content and services that better meet their needs, achieving a win-win situation for both platform commercial value and user experience. As the internet platform ecosystem becomes increasingly complex, user behavior is no longer an isolated individual decision, but rather the result of dynamic interaction and mutual competition among users, platforms, and the external environment. This poses a serious challenge to traditional user behavior data mining methods.

[0003] Currently, user behavior data mining suffers from several limitations. It often relies on fixed time thresholds for session segmentation, failing to consider the significant impact of environmental factors on user behavior intervals. This leads to session fragmentation or incorrect merging, making it impossible to accurately reconstruct users' continuous intentions and decision-making processes. Modeling behavioral patterns from a single user perspective completely ignores the profound influence of platform recommendation weight adjustments, resource scheduling, and external environmental changes on user decisions. Treating user behavior as an isolated, static process fails to capture the evolutionary logic of behavior under multi-subject interactions. Relying on design features and reward functions not only requires extensive domain expert knowledge, resulting in high development costs and long development cycles, but also struggles to adapt to complex and ever-changing user behavior scenarios. It is difficult to accurately predict users' next actions in complex and dynamic scenarios, thus impacting the accuracy and reliability of user behavior data mining. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides a user behavior data mining system and method based on multi-agent decision-making. This solution overcomes the limitations of the background technologies mentioned above, which often employ fixed time thresholds for session segmentation, failing to consider the significant impact of environmental factors on user behavior intervals. This leads to session fragmentation or incorrect merging, making it impossible to accurately reconstruct the user's continuous intent and decision-making process. Furthermore, it models behavioral patterns from a single user perspective, completely ignoring the profound impact of platform recommendation weight adjustments, resource scheduling, and other operational actions and external environmental changes on user decisions. It treats user behavior as an isolated, static process, failing to capture the behavioral evolution logic under multi-agent interactions. Relying on design features and reward functions not only requires extensive domain expert knowledge, resulting in high development costs and long development cycles, but also struggles to adapt to complex and ever-changing user behavior scenarios, making it difficult to accurately predict the user's next action in complex dynamic scenarios, thus affecting the accuracy and reliability of user behavior data mining.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A user behavior data mining method based on multi-agent decision-making includes: Obtain raw user behavior data, which represents the user's behavior records on the platform; Based on raw user behavior data, obtain corresponding platform operation data and behavioral environment data; Construct an intelligent environmental agent based on behavioral environmental data; Based on the original user behavior data, behavioral data feature information is obtained. The behavioral data feature information includes user identifier, behavior type, behavior occurrence time, behavior duration, and behavior associated object. The behavior type includes click, browse, favorite, share, comment, purchase, login, and logout. The behavior associated object represents the target corresponding to the user behavior. User behavior data is grouped according to user identifiers and sorted by the time of occurrence of the behavior to obtain a set of behavior sequences for each user. Based on the environmental intelligent agent, obtain the time interval of user behavior data segmentation corresponding to each user behavior; Based on the time interval between the occurrence of the behavior and the segmentation of user behavior data, the set of behavior sequences of each user is divided into multiple consecutive behavior sessions to obtain user behavior session information; If the time interval between two adjacent behaviors in the behavior sequence set is greater than the time interval of the user behavior data division, they are considered to be different sessions. Construct a user intelligent agent based on user behavior and session information; Construct a platform intelligent agent based on platform operation data; Fitting of environmental agents, user agents, and platform agents based on maximum likelihood inverse reinforcement learning; Based on the fitted environmental agent, user agent, and platform agent, user behavior data mining is performed. Preferably, the construction of the environmental intelligent agent based on behavioral environment data specifically includes: Based on behavioral environment data, temporal and spatial environmental features are extracted. Temporal environmental features include hour and day of the week, while spatial environmental features include user location, city level, weather conditions, and temperature. The extracted environmental features are standardized by converting continuous features into discrete features to obtain standardized environmental features. Based on standardized environmental characteristics, and using the K-means clustering algorithm, all environmental states are divided into multiple non-overlapping environmental categories. For each environment category, extract the original user behavior data corresponding to that environment category; Based on the original user behavior data under each environment category, a time interval sequence is obtained based on the time interval between all adjacent behaviors; Based on the time interval sequence and the kernel density estimation method, the probability density function of the user behavior time interval under this environment category is obtained; Based on the probability density function, obtain the maximum time interval corresponding to each environmental category; The maximum time interval corresponding to each environment category is used as the time interval for dividing user behavior data under that environment category; Based on the mapping relationship between environmental characteristics and the optimal time interval, an initial decision-making model for the environmental agent is constructed. Based on the standardized environmental characteristics, an environmental state space is constructed based on the combination of all standardized environmental characteristics; Based on user behavior data, a time interval range threshold is defined to construct the action space of the environmental intelligent agent; Based on the environmental state space, any combination of standardized environmental features can be used as the preceding state. Based on the original user behavior data, obtain the behavioral environment data corresponding to each user behavior; Based on the time of occurrence of user behavior, the behavioral environment data is sorted according to the time order to obtain a behavioral environment data sequence. Based on the behavioral environment data sequence, obtain the subsequent state corresponding to the preceding state. The ratio of the number of occurrences of each subsequent state to the total number of occurrences of subsequent states is taken as the environmental state transition probability between that subsequent state and the preceding state. Based on environmental decision-making needs, set an environmental discount factor; The environmental agent is obtained based on the environmental state space, action space, environmental state transition probability, initial reward function, and environmental discount factor.

[0006] Preferably, the step of constructing a user intelligent agent based on user behavior session information specifically includes: Based on the set of behavior sequences, take any one type of behavior as the target behavior, and select another type of behavior as the labeling behavior of the target behavior; Based on the target behavior and the labeled behavior, characteristic behavior pairs are formed; Based on user behavior session information, traverse adjacent behavior pairs in each behavior session to obtain the number of feature behavior pairs and the total number of times the target behavior is obtained. The ratio of the number of times a feature behavior pair is used to the total number of times the target behavior is used as the behavior transition probability of that feature behavior pair; Based on user behavior session information, a user state space is constructed with any user behavior as the current behavior. The user state includes the current behavior type, the previous behavior type, the current session duration, and the user activity level. Based on the user state space, the action space of the user agent is defined, which includes behavior types, and each behavior type corresponds to a one-hot encoded vector. Based on the user state space and raw user behavior data, obtain the type and frequency of subsequent user behavior in each user state; For each user state, calculate the total frequency of all subsequent actions in that state; Based on the frequency of behaviors and the total frequency obtained from statistics, initialize the initial policy function of the user agent; Based on the user state space, action space, behavior transition probability, initial reward function, and user discount factor, obtain the user model feature set; The user agent is obtained based on the initial policy function and the user model feature set.

[0007] Preferably, the step of constructing a platform intelligent agent based on platform operation data specifically includes: Based on the platform's operational data, extract platform content data, platform user traffic data, platform resource usage data, and platform rule execution data; Based on the platform's content data, a platform content library is constructed, and a unique identifier is assigned to each piece of content to obtain content feature information. The content features include the content's category, tags, publication time, number of views, number of likes, number of comments, and number of shares. Based on platform user traffic data, statistics are compiled on user visits, dwell time, click-through rate, conversion rate, and repurchase rate for different time periods, different content categories, and different user groups. Based on platform resource usage data, calculate the CPU utilization, memory utilization, disk I / O rate, bandwidth utilization, and response time and hit rate of each node in the content delivery network for the platform server cluster. Based on platform rule execution data, extract the platform's content recommendation rules, user points rules, level promotion rules, violation penalty rules, and resource scheduling rules; Based on the platform's core operational objectives, the action space of the platform's intelligent agent is defined. This action space includes adjusting the recommendation weights of different content categories, adjusting the allocation ratio of platform server resources, and adjusting the traffic allocation ratio of content distribution network nodes. Based on platform operation data, a platform state space is constructed, which includes the distribution of online user numbers, server cluster load distribution, user conversion rate distribution, and platform resource utilization rate. Based on user behavior session information and time synchronization, obtain platform operation data corresponding to each user behavior session; Based on the platform operation data corresponding to each user behavior session, obtain the platform action information in each user behavior session; Based on the platform action information, obtain the platform state and environment state before and after the platform action is executed; The platform action, the platform state before the platform action is executed, and the environment state are taken as the platform triple; Based on the platform operation data, obtain the platform status information after the execution of the platform action corresponding to the platform triple; Based on the number of platform states after each platform action is executed and the total number of platform triples, the platform state transition probability is obtained using the Laplace smoothing parameter. Based on the platform state space, action space, platform state transition probability, initial reward function, and platform discount factor, obtain the platform model feature set; Based on the platform model feature groups, obtain the platform intelligent agent.

[0008] Preferably, the fitting of the environmental agent, user agent, and platform agent based on maximum likelihood inverse reinforcement learning specifically includes: Based on the constructed user agent state space, platform agent state space, and environment agent state space, obtain the joint state space; Based on the established user agent action space, platform agent action space, and environmental agent action space, obtain the joint action space; Based on the joint state space and joint action space, the product of the behavior transition probability and the environment state transition probability and the platform state transition probability is taken as the joint state transition probability. Based on the joint state space and joint action space, obtain the reward function corresponding to each agent; Based on the joint state space, joint action space, joint state transition probability, and reward function, environmental agents, user agents, and platform agents are obtained using maximum likelihood inverse reinforcement learning.

[0009] Preferably, the step of performing user behavior data mining based on the fitted environmental agent, user agent, and platform agent specifically includes: Real-time collection of user's current behavior data, platform's current operating data, and current environmental data; The preprocessed current environment data is input into the environmental agent to obtain the time interval for dividing user behavior data under the current environmental conditions; Based on the acquisition time interval, the real-time collected user behavior data is dynamically divided to obtain the current user behavior session; Update the current state of the user's intelligent agent based on the current user behavior session; Update the current state of the platform's intelligent agent based on the current platform operation data; Update the current state of the environmental agent based on the current environmental data; By combining the current states of the three agents, we obtain the current joint state of the system. The joint Nash equilibrium policy function of the three agents is invoked to predict the user's next action.

[0010] Furthermore, a user behavior data mining system based on multi-agent decision-making is proposed to implement the data mining method described above, including: The main control module is used to construct an environmental intelligent agent based on behavioral environment data, construct a user intelligent agent based on user behavior session information, construct a platform intelligent agent based on platform operation data, fit the environmental intelligent agent, user intelligent agent and platform intelligent agent based on maximum likelihood inverse reinforcement learning, and perform user behavior data mining based on the fitted environmental intelligent agent, user intelligent agent and platform intelligent agent. The information acquisition module is used to acquire raw user behavior data, acquire corresponding platform operation data and behavior environment data based on the raw user behavior data, acquire behavior data feature information according to the raw user behavior data, and collect the user's current behavior data, the platform's current operation data, and the current environment data in real time. The data analysis module is used to extract temporal and spatial environmental features based on behavioral environmental data, obtain time interval sequences based on the time intervals between all adjacent behaviors and the original user behavior data under each environmental category, group user behavior data according to user identifiers, and sort user behavior data by behavior occurrence time to obtain a set of behavior sequences corresponding to each user. The display module interacts with the main control module and is used to output and display raw user behavior data, behavior data feature information, and the user's next action.

[0011] Optionally, the main control module specifically includes: The control unit is used to fit the environmental agent, user agent and platform agent based on maximum likelihood inverse reinforcement learning, and to perform user behavior data mining based on the fitted environmental agent, user agent and platform agent. An information receiving unit interacts with an information acquisition module and a data analysis module to receive data and transmit it to an intelligent agent unit. The intelligent agent unit is used to construct an environmental intelligent agent based on behavioral environment data, a user intelligent agent based on user behavior session information, and a platform intelligent agent based on platform operation data.

[0012] Optionally, the information acquisition module specifically includes: The first acquisition unit is used to acquire raw user behavior data and, based on the raw user behavior data, acquire corresponding platform operation data and behavior environment data. The second acquisition unit is used to acquire behavioral data feature information based on the original user behavior data, and to collect the user's current behavior data, the platform's current operating data, and the current environment data in real time.

[0013] Optionally, the data analysis module specifically includes: The first analysis unit is used to extract temporal and spatial environmental features based on behavioral environmental data, and to obtain a time interval sequence based on the time interval between all adjacent behaviors according to the original user behavior data under each environmental category. The second analysis unit is used to group user behavior data according to user identifiers and sort user behavior data according to the time of occurrence of behavior to obtain a set of behavior sequences corresponding to each user.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a user behavior data mining system and method based on multi-agent decision-making. By extracting multi-dimensional environmental features in time and space and performing clustering, it achieves accurate classification of different environmental states. By using kernel density estimation, it obtains the probability distribution of user behavior time intervals under each environmental category, realizing dynamic adaptive segmentation of user behavior sessions. By constructing a user state space that includes the current behavior, the previous behavior, session duration, and user activity, it achieves refined modeling of the user behavior decision-making process. By constructing a three-agent joint Markov decision process, it achieves complete modeling of the multi-agent interactive decision-making process, ensuring the accuracy of data mining. Attached Figure Description

[0015] Figure 1 This is a flowchart of a user behavior data mining method based on multi-agent decision-making proposed in this invention; Figure 2 This is a flowchart of the data partitioning time interval acquisition process in this invention; Figure 3 This is a flowchart of the user intelligent agent acquisition process in this invention; Figure 4 This is a flowchart of the platform intelligent agent acquisition process in this invention; Figure 5 This is a block diagram of a user behavior data mining system based on multi-agent decision-making proposed in this invention. Detailed Implementation

[0016] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0017] Reference Figure 1 - Figure 4 As shown in the figure, a user behavior data mining method based on multi-agent decision-making in an embodiment of the present invention includes: Obtain raw user behavior data, which represents the user's behavior records on the platform; Based on raw user behavior data, obtain corresponding platform operation data and behavioral environment data; Construct an intelligent environmental agent based on behavioral environmental data; Specifically, based on behavioral environment data, an environmental intelligent agent is constructed, including: Based on behavioral environment data, temporal and spatial environmental features are extracted. Temporal environmental features include hour and day of the week, while spatial environmental features include user location, city level, weather conditions, and temperature. The extracted environmental features are standardized by converting continuous features into discrete features to obtain standardized environmental features. Based on standardized environmental characteristics, and using the K-means clustering algorithm, all environmental states are divided into multiple non-overlapping environmental categories. For each environment category, extract the original user behavior data corresponding to that environment category; Based on the original user behavior data under each environment category, a time interval sequence is obtained based on the time interval between all adjacent behaviors; Based on the time interval sequence and the kernel density estimation method, the probability density function of the user behavior time interval under this environment category is obtained; Based on the probability density function, obtain the maximum time interval corresponding to each environmental category; The maximum time interval corresponding to each environment category is used as the time interval for dividing user behavior data under that environment category; Based on the mapping relationship between environmental characteristics and the optimal time interval, an initial decision-making model for the environmental agent is constructed. Based on the standardized environmental characteristics, an environmental state space is constructed based on the combination of all standardized environmental characteristics; Based on user behavior data, a time interval range threshold is defined to construct the action space of the environmental intelligent agent; Based on the environmental state space, any combination of standardized environmental features can be used as the preceding state. Based on the original user behavior data, obtain the behavioral environment data corresponding to each user behavior; Based on the time of occurrence of user behavior, the behavioral environment data is sorted according to the time order to obtain a behavioral environment data sequence. Based on the behavioral environment data sequence, obtain the subsequent state corresponding to the preceding state. The ratio of the number of occurrences of each subsequent state to the total number of occurrences of subsequent states is taken as the environmental state transition probability between that subsequent state and the preceding state. Based on environmental decision-making needs, set an environmental discount factor; The environmental agent is obtained based on the environmental state space, action space, environmental state transition probability, initial reward function, and environmental discount factor.

[0018] This solution extracts environmental features from two dimensions—time (hours, days of the week) and space (weather, temperature, humidity)—simultaneously and standardizes them using equally spaced integer encoding and min-max normalization. This enables a systematic and numerical representation of all external factors influencing user behavior, providing a comprehensive environmental context for subsequent differentiated behavior analysis. It automatically optimizes the number of environmental categories, dividing a massive number of discrete environmental features into multiple non-overlapping homogeneous categories, significantly reducing the complexity of subsequent calculations and balancing the model's representation accuracy with operational efficiency. For each environmental category, based on the user behavior time interval sequence, a Gaussian kernel density estimation method with leave-one-out least squares cross-validation is used to determine the optimal bandwidth, accurately fitting the probability distribution of the time interval and calculating the optimal session segmentation threshold accordingly.

[0019] In this embodiment, the extracted environmental features are standardized, converting continuous features into discrete features. For example, hours are divided into four intervals: early morning (0-6), morning (6-12), afternoon (12-18), and evening (18-24); days of the week are divided into weekdays and weekends; and temperatures are divided into four intervals: cold (<10℃), cool (10-20℃), warm (20-30℃), and hot (>30℃). The converted discrete categories are encoded using an equally spaced integer encoding method, mapping each discrete category to a continuous integer. For example, for the hour, the code for early morning (0-6) is 0, the code for morning (6-12) is 1, the code for afternoon (12-18) is 2, and the code for evening (18-24) is 3. For the day of the week, the code for weekday is 0, and the code for weekend is 1. For the temperature, the code for cold (<10℃) is 0, the code for cool (10-20℃) is 1, the code for warm (20-30℃) is 2, and the code for hot (>30℃) is 3. The environmental state is obtained by randomly combining all discrete features; For each environmental state, the encoded values ​​of the corresponding discrete features are concatenated in order to obtain a numerical feature vector: Perform min-max normalization on all features to map the value of each feature to the interval [0,1]. For a given number of clusters The sum of squared errors (SSE) of K-means clustering is defined as the sum of the squared Euclidean distances from each sample to its cluster center: In the formula, For the first A cluster, The normalized feature vectors, For the first The center vector of each cluster It is the Euclidean norm; Based on the total number of combinations of environmental characteristics and the actual data distribution, the search range is set as follows: arrive The specific settings can be customized according to actual needs.

[0020] Calculations are different Value : For each The value was obtained by running the K-means clustering algorithm 10 times (using different random initial centers) and taking the results of the 10 runs. The minimum result is used as the The final value ,get sequence; Based on cluster number As the x-axis, with Plot the ordinate as the vertical axis. The curve that varies with k.

[0021] calculate First difference of the curve: in, When adding a cluster The decrease; Calculate the second-order difference: express The rate of change of the descent speed; The elbow point corresponds to the position where the second-order difference reaches its maximum value. The rate of decline changes abruptly, increasing the cluster pairs The improvement effect was significantly reduced; The number of environmental categories is determined based on the number of clusters corresponding to elbow points. Based on the K-means clustering algorithm and the number of environmental categories, all environmental states are divided. It is understood that clustering data using the K-means clustering algorithm is a common technique known to those skilled in the art, so it will not be described in detail here. Based on the time interval sequence, the probability density function of user behavior time intervals under this environment category is calculated using the kernel density estimation method. ,Right now: In the formula, The time interval for calculating the density value is the independent variable. Represents the first time interval in the time interval sequence A time interval, Bandwidth for kernel density estimation; This scheme uses a Gaussian kernel (normal kernel) as the kernel function, and its expression is: Substitute the kernel function into the probability density function ,get: Among them, bandwidth The core of this method, determined by leave-one-out least squares cross-validation, is minimizing the mean square integral error of the density estimation. Define the cross-validation objective function: In the formula, To use all samples, bandwidth is The kernel density estimation results at that time This means removing the first... After one sample, use the remaining The density function calculated from each sample is in The value at; Based on leave-one-out kernel density estimation, for Expanding, we get: Iterate through the feasible bandwidth range and select the bandwidth that minimizes the objective function as the optimal bandwidth. ,Right now: Will Substituting into the aforementioned probability density function formula, we can obtain the final user behavior time interval probability density function for this environment category. ; According to the probability density function Find the time interval that makes the following expression reach its maximum value. : The solution obtained This serves as the optimal time interval for segmenting user behavior data within this environmental category; Repeat the above steps until the optimal time interval is obtained for all environment categories; Based on the mapping relationship between environmental characteristics and the optimal time interval, an initial decision-making model for the environmental agent is constructed. The initial decision model represents the optimal time interval for each environmental category; Define the state space of the environmental intelligent agent. This refers to a combination of all the standardized environmental characteristics described above. Define the action space of the environmental intelligent agent Divide all possible user behavior data into time intervals, ranging from 1 minute to 120 minutes, with a step size of 1 minute; Define the initial Markov decision process model of the environmental intelligent agent as a quintuple. ,in, For the initial reward function, Environmental discount factor; Initialize reward function It is a constant of 0, and is subsequently optimized by a joint maximum likelihood inverse reinforcement learning algorithm; Set environmental discount factor It is used to balance the immediate effects and long-term adaptability of environmental decisions.

[0022] Based on the original user behavior data, behavioral data feature information is obtained. The behavioral data feature information includes user identifier, behavior type, behavior occurrence time, behavior duration, and behavior associated object. The behavior type includes click, browse, favorite, share, comment, purchase, login, and logout. The behavior associated object represents the target corresponding to the user behavior. User behavior data is grouped according to user identifiers and sorted by the time of occurrence of the behavior to obtain a set of behavior sequences for each user. Based on the environmental intelligent agent, obtain the time interval of user behavior data segmentation corresponding to each user behavior; Based on the time interval between the occurrence of the behavior and the segmentation of user behavior data, the set of behavior sequences of each user is divided into multiple consecutive behavior sessions to obtain user behavior session information; If the time interval between two adjacent behaviors in the behavior sequence set is greater than the time interval of the user behavior data division, they are considered to be different sessions. Construct a user intelligent agent based on user behavior and session information; Specifically, based on user behavior and session information, a user intelligent agent is constructed, including: Based on the set of behavior sequences, take any one type of behavior as the target behavior, and select another type of behavior as the labeling behavior of the target behavior; Based on the target behavior and the labeled behavior, characteristic behavior pairs are formed; Based on user behavior session information, traverse adjacent behavior pairs in each behavior session to obtain the number of feature behavior pairs and the total number of times the target behavior is obtained. The ratio of the number of times a feature behavior pair is used to the total number of times the target behavior is used as the behavior transition probability of that feature behavior pair; Based on user behavior session information, a user state space is constructed with any user behavior as the current behavior. The user state includes the current behavior type, the previous behavior type, the current session duration, and the user activity level. Based on the user state space, the action space of the user agent is defined, which includes behavior types, and each behavior type corresponds to a one-hot encoded vector. Based on the user state space and raw user behavior data, obtain the type and frequency of subsequent user behavior in each user state; For each user state, calculate the total frequency of all subsequent actions in that state; Based on the frequency of behaviors and the total frequency obtained from statistics, initialize the initial policy function of the user agent; Based on the user state space, action space, behavior transition probability, initial reward function, and user discount factor, obtain the user model feature set; The user agent is obtained based on the initial policy function and the user model feature set.

[0023] This scheme constructs "target behavior - labeled behavior" feature behavior pairs, statistically analyzes the co-occurrence frequency of adjacent behaviors, and calculates behavior transition probabilities, providing core data support for reconstructing the complete user decision-making chain. A user state space is constructed, comprising four dimensions: "current behavior type, previous behavior type, current session duration, and user activity," achieving a comprehensive and refined representation of the user's decision-making state. This provides rich state inputs for accurately predicting the user's next behavior. Based on historical behavior data, the frequency distribution of subsequent behaviors under different user states is statistically analyzed. Laplace smoothing is used to initialize the initial policy function of the user agent, ensuring the rationality of the initial policy and laying a good foundation for the rapid convergence of subsequent inverse reinforcement learning. Integrating the user state space, action space, behavior transition probabilities, initial reward function, and user discount factor, a complete Markov decision process model of the user agent is constructed. The user's behavioral decision-making process is formalized into a solvable mathematical model, providing a core foundation for user behavior modeling in multi-agent joint decision-making systems.

[0024] In this embodiment, the current behavior type And the previous behavior type The dimension takes the value of all possible behavior types, that is... , The total number of behavior types; Divide the duration of the current session into 5 intervals: Define the duration dimension of the current session. The value is the index corresponding to these 5 intervals, that is... ; Based on the user's total activity over the past 7 days The average total number of actions of all users on the platform over the past 7 days Calculate user activity User activity is categorized into 5 levels: extremely low activity. Low activity Moderate activity High activity Extremely high activity level Define user activity dimensions The value is the index corresponding to these 5 levels, that is... ; Based on user historical behavior conversion data, statistics are compiled for different user states. Users select different follow-up actions The frequency of, denoted as ,in Indicates the type of subsequent user behavior. ; For each user state Calculate the total frequency of all subsequent actions in this state: Based on the frequency of behavior obtained from statistics Total frequency Initialize the user agent's initial policy function: In the formula, Let the Laplace smoothing parameter be denoted as . This is used to avoid zero-probability problems; Define the initial Markov decision process model of the user agent as a quintuple. ,in, For user state space, Provide space for user actions Let the behavior transition probability be... For the initial reward function, For user discount factors; Initialize reward function It is a constant of 0, and is subsequently optimized by a joint maximum likelihood inverse reinforcement learning algorithm; Set user discount factor It is used to balance users' immediate rewards and future rewards, and can be set according to actual needs.

[0025] Construct a platform intelligent agent based on platform operation data; Specifically, based on platform operation data, platform content data, platform user traffic data, platform resource usage data, and platform rule execution data are extracted; Based on the platform's content data, a platform content library is constructed, and a unique identifier is assigned to each piece of content to obtain content feature information. The content features include the content's category, tags, publication time, number of views, number of likes, number of comments, and number of shares. Based on platform user traffic data, statistics are compiled on user visits, dwell time, click-through rate, conversion rate, and repurchase rate for different time periods, different content categories, and different user groups. Based on platform resource usage data, calculate the CPU utilization, memory utilization, disk I / O rate, bandwidth utilization, and response time and hit rate of each node in the content delivery network for the platform server cluster. Based on platform rule execution data, extract the platform's content recommendation rules, user points rules, level promotion rules, violation penalty rules, and resource scheduling rules; Based on the platform's core operational objectives, the action space of the platform's intelligent agent is defined. This action space includes adjusting the recommendation weights of different content categories, adjusting the allocation ratio of platform server resources, and adjusting the traffic allocation ratio of content distribution network nodes. Based on platform operation data, a platform state space is constructed, which includes the distribution of current online users, server cluster load distribution, popular content distribution, user conversion rate distribution, and platform resource utilization rate. Based on user behavior session information and time synchronization, obtain platform operation data corresponding to each user behavior session; Based on the platform operation data corresponding to each user behavior session, obtain the platform action information in each user behavior session; Based on the platform action information, obtain the platform state and environment state before and after the platform action is executed; The platform action, the platform state before the platform action is executed, and the environment state are taken as the platform triple; Based on the platform operation data, obtain the platform status information after the execution of the platform action corresponding to the platform triple; Based on the number of platform states after each platform action is executed and the total number of platform triples, the platform state transition probability is obtained using the Laplace smoothing parameter. Based on the platform state space, action space, platform state transition probability, initial reward function, and platform discount factor, obtain the platform model feature set; Based on the platform model feature groups, obtain the platform intelligent agent.

[0026] This solution comprehensively extracts operational data from four core dimensions: platform content, user traffic, resource usage, and rule enforcement. This achieves full coverage of the complete ecosystem generated by user behavior, constructs a standardized platform content library, extracts multi-dimensional content features, and statistically analyzes core indicators such as user visits, dwell time, and conversion rates across different dimensions. This transforms the platform content ecosystem and user response characteristics into computable, structured data. It not only accurately reflects the popularity and dissemination effect of different content but also precisely captures users' preferences for different content categories, providing refined feature support for subsequent modeling of the impact of platform content recommendation strategies on user behavior.

[0027] In this embodiment, the distribution dimension of the current number of online users is defined to have 5 intervals: The corresponding indices are 1 to 5; The server cluster load distribution dimension is defined to have 5 intervals: The corresponding indices are 1 to 5; The user conversion rate distribution dimension is defined with 5 intervals: The corresponding indices are 1 to 5; The platform resource utilization dimension is defined with 5 ranges: The corresponding indices are 1 to 5; Based on historical platform operation data, statistics were compiled on the status across different platforms. and environmental conditions Execute actions on different platforms Changes in post-platform operational metrics; Based on the statistical changes in platform operation indicators, initialize the initial state transition probability matrix of the platform agent. ,in Indicates the platform status and environmental conditions Next action Later transitioned to platform status The probability of; Define the initial Markov decision process model of the platform agent as a quintuple. ,in For the initial reward function, This is a platform discount factor; Initialize reward function It is a constant of 0, and is subsequently optimized by a joint maximum likelihood inverse reinforcement learning algorithm; Set platform discount factor This is used to balance the platform's short-term operating revenue and long-term development goals.

[0028] Fitting of environmental agents, user agents, and platform agents based on maximum likelihood inverse reinforcement learning; Specifically, the maximum likelihood inverse reinforcement learning is used to fit the environmental agent, user agent, and platform agent, including: Based on the constructed user agent state space, platform agent state space, and environment agent state space, obtain the joint state space; Based on the established user agent action space, platform agent action space, and environmental agent action space, obtain the joint action space; Based on the joint state space and joint action space, the product of the behavior transition probability and the environment state transition probability and the platform state transition probability is taken as the joint state transition probability. Based on the joint state space and joint action space, obtain the reward function corresponding to each agent; Based on the joint state space, joint action space, joint state transition probability, and reward function, environmental agents, user agents, and platform agents are obtained using maximum likelihood inverse reinforcement learning.

[0029] In this scheme, the independent state and action spaces of the user, platform, and environment are integrated through Cartesian product to construct a unified joint state and action space. The joint state transition probability is decomposed into the product of user behavior transition probability, platform state transition probability, and environment state transition probability. This effectively avoids the curse of dimensionality problem common in multi-agent modeling while ensuring the accuracy of the joint system's state evolution. Based on the observed user-platform-environment joint interaction trajectory, the optimal reward function of the three agents is solved in reverse using the maximum likelihood inverse reinforcement learning algorithm. The model can automatically learn the user's true decision-making motivation, the platform's operational goals, and the inherent evolutionary laws of the environment from massive behavioral data, effectively covering user scenarios. By solving the joint Nash equilibrium strategy of the three parties through a multi-agent value iteration algorithm, the optimal decision-making behavior of the user, platform, and environment in the dynamic game process is accurately captured, achieving forward-looking prediction of multi-party interaction behavior in complex dynamic scenarios.

[0030] It is understandable that in this embodiment, the joint state space is: In the formula, For a joint state space, For the user's intelligent agent state space. For the platform's intelligent agent state space, For the state space of the environmental intelligent agent; Wherein, any joint state , can be represented as: In the formula, In a united state, For user status, The platform is in a certain state. Environmental state; Joint Action Space: In the formula, For joint action space, For the user's intelligent agent's action space, For the action space of the platform's intelligent agents, For the action space of the environmental intelligent agent; Any joint action , can be represented as: In the formula, For joint operations, For user actions, For platform actions, Actions for the environment; Joint state transition function , indicating a joint state All agents perform joint actions Then, the system transitions to the next joint state. The probability of; The joint state transition probability is decomposed into the product of the state transition probabilities of the three agents: In the formula, Indicates the environmental state User status Through user actions Transfer to user state The probability, Indicates the environmental state Below, platform status Through platform actions Transfer to user state The probability, Indicates environmental state Transferred to The probability of; Define the reward function of the user agent as a linear function: In the formula, Let be the user reward weight vector to be solved. The user feature vector for the joint state-joint action pair; Define the reward function of the platform agent as a linear function: In the formula, Let be the platform reward weight vector to be solved. The platform feature vector is the joint state-joint action pair; Define the reward function of the environmental agent as a linear function: In the formula, Let be the environmental reward weight vector to be solved. The environment feature vector for the joint state-joint action pair; Constructing user feature vectors This includes user behavior frequency feature sub-vectors, user session feature sub-vectors, and user activity feature sub-vectors; Among them, the user behavior frequency feature sub-vector represents the user's long-term behavior preferences. Based on the user identifier corresponding to the user status, the ratio of the number of occurrences of each type of behavior to the total number of behaviors in the past 30 days is extracted as the occurrence frequency of that type of behavior. Max-min normalization is applied to the frequency of occurrence of the behavior to eliminate dimensional differences.

[0031] The normalized frequency values ​​are arranged in index order according to behavior type (click, browse, favorite, share, comment, purchase, login, logout), forming a structure of length [length missing]. subvectors, Total number of behavior types; User session feature subvectors represent the state attributes of a user's current short-term session, including: The current session duration has been normalized based on user state. Take the midpoint value of the interval corresponding to the "current session duration" dimension. Truncation and normalization are performed with a maximum of 1800 seconds (30 minutes): The diversity of current session behavior is defined as the ratio of the number of different behavior types that have appeared in the current session to the total number of behavior types. The normalized feature of the average behavior interval in the current session is the average interval between adjacent behaviors within the current active session. And truncate and normalize with a maximum of 600 seconds (10 minutes): User activity feature subvectors represent the overall activity level of users, including: The continuous value characteristic of user activity is the ratio of the total number of user behaviors in the past 7 days to the average number of user behaviors in the past 7 days across the entire platform. The activity level normalization feature maps the five levels of user activity (very low, low, medium, high, and very high) to equally spaced values: very low = 0, low = 0.25, medium = 0.5, high = 0.75, and very high = 1.0. These values ​​are directly used as feature values ​​to characterize the discrete hierarchical attributes of user activity.

[0032] Constructing platform feature vectors This includes user responsiveness improvement and user satisfaction, for each platform action. Obtain the user click-through rate one hour before and after the action is executed on the platform, and calculate the improvement in user response: In the formula, To improve user responsiveness, The user click-through rate in the hour prior to the platform's action. The user click-through rate one hour after the platform action is executed. ,but ; Statistical platform actions Average time users spend on the current page after execution (in seconds) and bounce rate (Except for the ratio of the number of users to the total number of users), obtain user satisfaction: In the formula, if ,but

[0033] Constructing environmental feature vectors This characterizes the impact of external environmental factors on user behavior benefits, specifically the normalized feature vector of the discrete feature encoding value of the environment state. Continuous user behavior is sampled from raw user behavior data, platform operation data, and behavioral environment data to generate multiple joint trajectories. Each trajectory... It consists of a series of continuous joint state-joint action pairs, and each trajectory simultaneously contains the complete decision-making process of the user, platform and environment. The joint trajectory is divided into training set and validation set. Initialize user reward weight vector It is a vector consisting entirely of 1s; Initialize the platform reward weight vector It is a vector consisting entirely of 1s; Initialize the environment reward weight vector It is a vector consisting entirely of 1s; Set the maximum number of iterations. Convergence threshold ; Set user learning rate Platform learning rate Environmental learning rate ; In the In the next iteration, based on the current reward weight vector , , A multi-agent value iterative algorithm is used to solve the Nash equilibrium strategy of a three-agent system. , , Specifically, it includes: Initialize the value function of each agent to 0, iteratively update the value function until convergence, and then calculate the Nash equilibrium policy based on the converged value function; Based on the joint trajectory, calculate the joint log-likelihood objective function value for the current iteration. : In the formula, To train the number of joint trajectories; Calculate the joint log-likelihood objective function for each pair. , , gradient: Update the reward weight vector using the gradient ascent algorithm: Calculate the update amount of the reward weight vector: Based on the joint trajectory in the validation set, calculate the joint log-likelihood value of the validation set in the current iteration. ; Determine if the convergence condition is met: If or If the joint log-likelihood value of the validation set no longer improves after three consecutive iterations, then stop the iteration and obtain the optimal reward weight vector. , , ; Otherwise, let Return to the previous step and continue iterating; Based on the optimal reward weight vector, calculate the optimal reward function for each agent: Based on the optimal reward function, the multi-agent value iteration algorithm is used again to solve for the final Nash equilibrium policy function of each agent. , , ; Based on the optimal reward function and the final Nash equilibrium policy function, the final specific decision-making model form of the user agent, platform agent, and environment agent is determined.

[0034] Based on the fitted environmental agent, user agent, and platform agent, user behavior data mining is performed.

[0035] Specifically, user behavior data mining is performed based on the fitted environmental agent, user agent, and platform agent, including: Real-time collection of user's current behavior data, platform's current operating data, and current environmental data; The preprocessed current environment data is input into the environmental agent to obtain the time interval for dividing user behavior data under the current environmental conditions; Based on the acquisition time interval, the real-time collected user behavior data is dynamically divided to obtain the current user behavior session; Update the current state of the user's intelligent agent based on the current user behavior session; Update the current state of the platform's intelligent agent based on the current platform operation data; Update the current state of the environmental agent based on the current environmental data; By combining the current states of the three agents, we obtain the current joint state of the system. The joint Nash equilibrium policy function of the three agents is invoked to predict the user's next action.

[0036] Reference Figure 5 As shown, further, combining the above-mentioned user behavior data mining method based on multi-agent decision-making, a user behavior data mining system based on multi-agent decision-making is proposed, including: The main control module is used to construct an environmental intelligent agent based on behavioral environment data, construct a user intelligent agent based on user behavior session information, construct a platform intelligent agent based on platform operation data, fit the environmental intelligent agent, user intelligent agent and platform intelligent agent based on maximum likelihood inverse reinforcement learning, and perform user behavior data mining based on the fitted environmental intelligent agent, user intelligent agent and platform intelligent agent. The information acquisition module is used to acquire raw user behavior data, acquire corresponding platform operation data and behavior environment data based on the raw user behavior data, acquire behavior data feature information according to the raw user behavior data, and collect the user's current behavior data, the platform's current operation data, and the current environment data in real time. The data analysis module is used to extract temporal and spatial environmental features based on behavioral environmental data, obtain time interval sequences based on the time intervals between all adjacent behaviors and the original user behavior data under each environmental category, group user behavior data according to user identifiers, and sort user behavior data by behavior occurrence time to obtain a set of behavior sequences corresponding to each user. The display module interacts with the main control module and is used to output and display raw user behavior data, behavior data feature information, and the user's next action.

[0037] The main control module specifically includes: The control unit is used to fit the environmental agent, user agent and platform agent based on maximum likelihood inverse reinforcement learning, and to perform user behavior data mining based on the fitted environmental agent, user agent and platform agent. An information receiving unit interacts with an information acquisition module and a data analysis module to receive data and transmit it to an intelligent agent unit. The intelligent agent unit is used to construct an environmental intelligent agent based on behavioral environment data, a user intelligent agent based on user behavior session information, and a platform intelligent agent based on platform operation data.

[0038] The information acquisition module specifically includes: The first acquisition unit is used to acquire raw user behavior data and, based on the raw user behavior data, acquire corresponding platform operation data and behavior environment data. The second acquisition unit is used to acquire behavioral data feature information based on the original user behavior data, and to collect the user's current behavior data, the platform's current operating data, and the current environment data in real time.

[0039] The data analysis module specifically includes: The first analysis unit is used to extract temporal and spatial environmental features based on behavioral environmental data, and to obtain a time interval sequence based on the time interval between all adjacent behaviors according to the original user behavior data under each environmental category. The second analysis unit is used to group user behavior data according to user identifiers and sort user behavior data according to the time of occurrence of behavior to obtain a set of behavior sequences corresponding to each user.

[0040] In summary, the advantages of this invention are as follows: by extracting multi-dimensional temporal and spatial environmental features and performing clustering, accurate classification of different environmental states is achieved; by obtaining the probability distribution of user behavior time intervals under each environmental category through kernel density estimation, dynamic adaptive segmentation of user behavior sessions is achieved; by constructing a user state space that includes the current behavior, the previous behavior, session duration, and user activity, refined modeling of the user behavior decision-making process is achieved; and by constructing a three-agent joint Markov decision process, complete modeling of the multi-agent interactive decision-making process is achieved, ensuring the accuracy of data mining.

[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for mining user behavior data based on multi-agent decision-making, characterized in that, include: Obtain raw user behavior data, which represents the user's behavior records on the platform; Based on raw user behavior data, obtain corresponding platform operation data and behavioral environment data; Construct an intelligent environmental agent based on behavioral environmental data; Based on the original user behavior data, behavioral data feature information is obtained. The behavioral data feature information includes user identifier, behavior type, behavior occurrence time, behavior duration, and behavior associated object. The behavior type includes click, browse, favorite, share, comment, purchase, login, and logout. The behavior associated object represents the target corresponding to the user behavior. User behavior data is grouped according to user identifiers and sorted by the time of occurrence of the behavior to obtain a set of behavior sequences for each user. Based on the environmental intelligent agent, obtain the time interval of user behavior data segmentation corresponding to each user behavior; Based on the time interval between the occurrence of the behavior and the segmentation of user behavior data, the set of behavior sequences of each user is divided into multiple consecutive behavior sessions to obtain user behavior session information; If the time interval between two adjacent behaviors in the behavior sequence set is greater than the time interval of the user behavior data division, they are considered to be different sessions. Construct a user intelligent agent based on user behavior and session information; Construct a platform intelligent agent based on platform operation data; Fitting of environmental agents, user agents, and platform agents based on maximum likelihood inverse reinforcement learning; Based on the fitted environmental agent, user agent, and platform agent, user behavior data mining is performed.

2. The user behavior data mining method based on multi-agent decision-making according to claim 1, characterized in that, The construction of an environmental intelligent agent based on behavioral environment data specifically includes: Based on behavioral environment data, temporal and spatial environmental features are extracted. Temporal environmental features include hour and day of the week, while spatial environmental features include user location, city level, weather conditions, and temperature. The extracted environmental features are standardized by converting continuous features into discrete features to obtain standardized environmental features. Based on standardized environmental characteristics, and using the K-means clustering algorithm, all environmental states are divided into multiple non-overlapping environmental categories. For each environment category, extract the original user behavior data corresponding to that environment category; Based on the original user behavior data under each environment category, a time interval sequence is obtained based on the time interval between all adjacent behaviors; Based on the time interval sequence and the kernel density estimation method, the probability density function of the user behavior time interval under this environment category is obtained; Based on the probability density function, obtain the maximum time interval corresponding to each environmental category; The maximum time interval corresponding to each environment category is used as the time interval for dividing user behavior data under that environment category; Based on the mapping relationship between environmental characteristics and the optimal time interval, an initial decision-making model for the environmental agent is constructed. Based on the standardized environmental characteristics, an environmental state space is constructed based on the combination of all standardized environmental characteristics; Based on user behavior data, a time interval range threshold is defined to construct the action space of the environmental intelligent agent; Based on the environmental state space, any combination of standardized environmental features can be used as the preceding state. Based on the original user behavior data, obtain the behavioral environment data corresponding to each user behavior; Based on the time of occurrence of user behavior, the behavioral environment data is sorted according to the time order to obtain a behavioral environment data sequence. Based on the behavioral environment data sequence, obtain the subsequent state corresponding to the preceding state. The ratio of the number of occurrences of each subsequent state to the total number of occurrences of subsequent states is taken as the environmental state transition probability between that subsequent state and the preceding state. Based on environmental decision-making needs, set an environmental discount factor; The environmental agent is obtained based on the environmental state space, action space, environmental state transition probability, initial reward function, and environmental discount factor.

3. The user behavior data mining method based on multi-agent decision-making according to claim 1, characterized in that, The construction of the user intelligent agent based on user behavior session information specifically includes: Based on the set of behavior sequences, take any one type of behavior as the target behavior, and select another type of behavior as the labeling behavior of the target behavior; Based on the target behavior and the labeled behavior, characteristic behavior pairs are formed; Based on user behavior session information, traverse adjacent behavior pairs in each behavior session to obtain the number of feature behavior pairs and the total number of times the target behavior is obtained. The ratio of the number of times a feature behavior pair is used to the total number of times the target behavior is used as the behavior transition probability of that feature behavior pair; Based on user behavior session information, a user state space is constructed with any user behavior as the current behavior. The user state includes the current behavior type, the previous behavior type, the current session duration, and the user activity level. Based on the user state space, the action space of the user agent is defined, which includes behavior types, and each behavior type corresponds to a one-hot encoded vector. Based on the user state space and raw user behavior data, obtain the type and frequency of subsequent user behavior in each user state; For each user state, calculate the total frequency of all subsequent actions in that state; Based on the frequency of behaviors and the total frequency obtained from statistics, initialize the initial policy function of the user agent; Based on the user state space, action space, behavior transition probability, initial reward function, and user discount factor, obtain the user model feature set; The user agent is obtained based on the initial policy function and the user model feature set.

4. The user behavior data mining method based on multi-agent decision-making according to claim 3, characterized in that, The construction of the platform intelligent agent based on platform operation data specifically includes: Based on the platform's operational data, extract platform content data, platform user traffic data, platform resource usage data, and platform rule execution data; Based on the platform's content data, a platform content library is constructed, and a unique identifier is assigned to each piece of content to obtain content feature information. The content features include the content's category, tags, publication time, number of views, number of likes, number of comments, and number of shares. Based on platform user traffic data, statistics are compiled on user visits, dwell time, click-through rate, conversion rate, and repurchase rate for different time periods, different content categories, and different user groups. Based on platform resource usage data, calculate the CPU utilization, memory utilization, disk I / O rate, bandwidth utilization, and response time and hit rate of each node in the content delivery network for the platform server cluster. Based on platform rule execution data, extract the platform's content recommendation rules, user points rules, level promotion rules, violation penalty rules, and resource scheduling rules; Based on the platform's core operational objectives, the action space of the platform's intelligent agent is defined. This action space includes adjusting the recommendation weights of different content categories, adjusting the allocation ratio of platform server resources, and adjusting the traffic allocation ratio of content distribution network nodes. Based on platform operation data, a platform state space is constructed, which includes the distribution of online user numbers, server cluster load distribution, user conversion rate distribution, and platform resource utilization rate. Based on user behavior session information and time synchronization, obtain platform operation data corresponding to each user behavior session; Based on the platform operation data corresponding to each user behavior session, obtain the platform action information in each user behavior session; Based on the platform action information, obtain the platform state and environment state before and after the platform action is executed; The platform action, the platform state before the platform action is executed, and the environment state are taken as the platform triple; Based on the platform operation data, obtain the platform status information after the execution of the platform action corresponding to the platform triple; Based on the number of platform states after each platform action is executed and the total number of times the platform triple is executed, the platform state transition probability is obtained using the Laplace smoothing parameter. Based on the platform state space, action space, platform state transition probability, initial reward function, and platform discount factor, obtain the platform model feature set; Based on the platform model feature groups, obtain the platform intelligent agent.

5. The user behavior data mining method based on multi-agent decision-making according to claim 4, characterized in that, The fitting of the environmental agent, user agent, and platform agent based on maximum likelihood inverse reinforcement learning specifically includes: Based on the constructed user agent state space, platform agent state space, and environment agent state space, obtain the joint state space; Based on the established user agent action space, platform agent action space, and environmental agent action space, obtain the joint action space; Based on the joint state space and joint action space, the product of the behavior transition probability and the environment state transition probability and the platform state transition probability is taken as the joint state transition probability. Based on the joint state space and joint action space, obtain the reward function corresponding to each agent; Based on the joint state space, joint action space, joint state transition probability, and reward function, environmental agents, user agents, and platform agents are obtained using maximum likelihood inverse reinforcement learning.

6. The user behavior data mining method based on multi-agent decision-making according to claim 5, characterized in that, The step of mining user behavior data based on the fitted environmental agent, user agent, and platform agent specifically includes: Real-time collection of user's current behavior data, platform's current operating data, and current environmental data; The preprocessed current environment data is input into the environmental agent to obtain the time interval for dividing user behavior data under the current environmental conditions; Based on the acquisition time interval, the real-time collected user behavior data is dynamically divided to obtain the current user behavior session; Update the current state of the user's intelligent agent based on the current user behavior session; Update the current state of the platform's intelligent agent based on the current platform operation data; Update the current state of the environmental agent based on the current environmental data; By combining the current states of the three agents, we obtain the current joint state of the system. The joint Nash equilibrium policy function of the three agents is invoked to predict the user's next action.

7. A user behavior data mining system based on multi-agent decision-making, used to implement the data mining method as described in any one of claims 1-6, characterized in that, include: The main control module is used to construct an environmental intelligent agent based on behavioral environment data, construct a user intelligent agent based on user behavior session information, construct a platform intelligent agent based on platform operation data, fit the environmental intelligent agent, user intelligent agent and platform intelligent agent based on maximum likelihood inverse reinforcement learning, and perform user behavior data mining based on the fitted environmental intelligent agent, user intelligent agent and platform intelligent agent. The information acquisition module is used to acquire raw user behavior data, acquire corresponding platform operation data and behavior environment data based on the raw user behavior data, acquire behavior data feature information according to the raw user behavior data, and collect the user's current behavior data, the platform's current operation data, and the current environment data in real time. The data analysis module is used to extract temporal and spatial environmental features based on behavioral environmental data, obtain time interval sequences based on the time intervals between all adjacent behaviors and the original user behavior data under each environmental category, group user behavior data according to user identifiers, and sort user behavior data by behavior occurrence time to obtain a set of behavior sequences corresponding to each user. The display module interacts with the main control module and is used to output and display raw user behavior data, behavior data feature information, and the user's next action.

8. A user behavior data mining system based on multi-agent decision-making according to claim 7, characterized in that, The main control module specifically includes: The control unit is used to fit the environmental agent, user agent and platform agent based on maximum likelihood inverse reinforcement learning, and to perform user behavior data mining based on the fitted environmental agent, user agent and platform agent. An information receiving unit interacts with an information acquisition module and a data analysis module to receive data and transmit it to an intelligent agent unit. The intelligent agent unit is used to construct an environmental intelligent agent based on behavioral environment data, a user intelligent agent based on user behavior session information, and a platform intelligent agent based on platform operation data.

9. A user behavior data mining system based on multi-agent decision-making according to claim 7, characterized in that, The information acquisition module specifically includes: The first acquisition unit is used to acquire raw user behavior data and, based on the raw user behavior data, acquire corresponding platform operation data and behavior environment data. The second acquisition unit is used to acquire behavioral data feature information based on the original user behavior data, and to collect the user's current behavior data, the platform's current operating data, and the current environment data in real time.

10. A user behavior data mining system based on multi-agent decision-making according to claim 7, characterized in that, The data analysis module specifically includes: The first analysis unit is used to extract temporal and spatial environmental features based on behavioral environmental data, and to obtain a time interval sequence based on the time interval between all adjacent behaviors according to the original user behavior data under each environmental category. The second analysis unit is used to group user behavior data according to user identifiers and sort user behavior data according to the time of occurrence of behavior to obtain a set of behavior sequences corresponding to each user.

Images