A game user behavior intelligent analysis method and system

Abstract

The application discloses a kind of game user behavior intelligent analysis method and system, belong to machine learning and data mining field, the method is: from message queue, the behavior log of each user in set time period is obtained and carries out feature engineering, obtains user feature matrix;Student-t distribution based probability density model is built, and the variational distribution family corresponding to probability density model is set;Based on user feature matrix and probability density model, each variational factor is iteratively updated until the preset iteration termination condition is met, and the group label of each user is determined and output based on the user membership probability of current iteration;Wherein, scale variable, user membership probability, mixed weight and group parameter are sequentially updated each time iteration;Each group label is compared with the historical group label of corresponding user respectively, and the behavior state change of each user is determined according to the comparison result.Therefore, by implementing the application, the accuracy and efficiency of game user behavior analysis can be improved.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and data mining, and in particular to a method and system for intelligent analysis of game user behavior. Background Technology

[0002] Game user behavior analysis is a core technical aspect of game operations. It aims to automatically segment massive numbers of users into different groups (such as "high-value users," "potential churn users," or "active non-paying users") by analyzing multi-dimensional behavioral data including login, payment, social interaction, and combat. This supports precise operational strategies (such as targeted push notifications, resource allocation, or customer complaint handling). Based on different modeling methods, user segmentation techniques are mainly divided into two categories: distance-based methods (such as K-means) and probability model-based methods (such as hybrid models).

[0003] The K-means algorithm groups users by minimizing the sum of squared distances from their feature vectors to the cluster centers. A typical application process in game operation systems includes: first, collecting raw user behavior data (such as login logs, payment records, or battle data) from the game database; then, converting the raw data into user feature vectors (such as "number of login days in the last 7 days" or "cumulative recharge amount") using a feature engineering module; next, randomly initializing K cluster centers and assigning each user to the nearest cluster based on Euclidean distance; finally, updating the cluster center positions based on the assignment results and repeating the iteration until convergence; and finally, writing the clustering results into the operation database to trigger differentiated operation strategies. However, K-means uses the sum of squared Euclidean distances as the objective function, implicitly assuming that each cluster follows a spherically equal-variance Gaussian distribution. This assumption makes the algorithm extremely sensitive to outliers and noisy data: outliers that are common in game operation data, such as users who make fraudulent transactions (e.g., making 50-100 times more purchases than normal users but spending very little money) or cheating users (e.g., having abnormal online time), will have their squared distances significantly amplified, thus seriously affecting the estimation of group centers and causing normal users to be incorrectly grouped.

[0004] To overcome the hard assignment problem of K-means, the Gaussian Mixture Model (GMM) introduces a probabilistic modeling framework, allowing users to belong to multiple groups with varying probabilities (soft assignment). In game user segmentation systems, GMM assumes that the behavioral characteristics of each user group follow a multivariate Gaussian distribution, and iteratively estimates group parameters (such as mixture weights, mean vectors, or covariance matrices) using the EM algorithm. However, this method suffers from bottlenecks in high-dimensional computation: game user behavioral characteristics typically include indicators from multiple subsystems such as spending, activity, social interaction, and combat, resulting in high dimensionality. The covariance matrix often exceeds 50. Storage required Parameters ( (where M is the number of groups), each M-step requires calculating the matrix inversion to evaluate the Gaussian density, with a complexity of O(n). In typical game scenarios with millions of users and 100-dimensional features, memory consumption can reach several gigabytes, and a single clustering operation can take more than 30 minutes, failing to meet the response requirements of real-time operations. Furthermore, the Gaussian distribution is a light-tailed distribution (the probability at the tail decays exponentially), assigning extremely low probabilities to samples far from the mean. When outliers exist in the game data, the EM algorithm may incorrectly increase the covariance matrix or create additional clusters in order to fit these points, leading to sensitivity to outliers and making it prone to overfitting or degraded clustering quality.

[0005] Variational inference (VI) is a type of approximate inference method within Bayesian inference. In game operation systems, variational Gaussian mixture models (VBGMM) introduce prior distributions (such as Dirichlet priors or Normal-Wishart priors) into the model parameters and use mean-field approximations to solve for the posterior distribution, enabling fully closed-loop parameter updates and avoiding the overhead of numerical optimization. However, VBGMM still uses light-tailed Gaussian distributions for modeling, lacking robustness to outliers in game data. Furthermore, in high-dimensional scenarios, the Wishart prior is essentially designed for the complete covariance matrix, and the issues of parameter size and computational complexity remain unresolved.

[0006] To address the sensitivity of Gaussian mixture models to outliers, some studies have attempted to replace the Gaussian distribution with the Student-t distribution to construct mixture models. The Student-t distribution is a heavy-tailed distribution, where the probability of the tail decays multinomially rather than exponentially, assigning relatively high probabilities to samples far from the center, thus achieving robustness against outliers. However, existing technologies provide the Student-t mixture model as a general clustering algorithm library, without integrating it with the specific business scenarios of game operations, lacking end-to-end system design. Specific problems include: as an independent algorithm module, it does not involve game log collection, feature engineering, or operational strategy execution, making it impossible to directly deploy to production systems; using a point estimation framework (EM algorithm), it cannot provide quantification of parameter uncertainty, and the anomaly tolerance parameter requires manual tuning; and it lacks optimization for game scenarios (e.g., how to identify fraudulent transactions, how to handle multi-dimensional behavioral features, or how to configure real-time update frequency). Summary of the Invention

[0007] This invention provides a method and system for intelligent analysis of game user behavior, which can improve the accuracy and efficiency of game user behavior analysis.

[0008] This invention provides a method for intelligent analysis of game user behavior, comprising: Retrieve the behavior logs of each user within a set time period from the message queue, and perform feature engineering on each behavior log to obtain a user feature matrix; wherein, each row of the user feature matrix corresponds to a user feature vector; Construct a probability density model based on the Student-t distribution, and set the variational distribution family corresponding to the probability density model; wherein, the variational factors of the variational distribution family include mixed weights, group parameters, user affiliation probabilities and scale variables, and the group parameters include group center vectors; Based on the user feature matrix and the probability density model, each variational factor is iteratively updated until a preset iteration termination condition is met. The group label of each user is determined and output based on the user affiliation probability of the current iteration. In each iteration, the scaling variable, user affiliation probability, mixture weight and group parameter are updated sequentially based on the distance between each user feature vector and the group center vector of the current iteration. Each group label is compared with the corresponding user's historical group labels, and the changes in the user's behavioral status are determined based on the comparison results.

[0009] This invention, through obtaining behavioral logs within a set time period from a message queue and performing feature engineering, can transform scattered raw events into a structured user feature matrix. By constructing a probability density model based on the Student-t distribution, a probability density model with heavy-tailed distribution characteristics can be established for game user behavior, thereby improving the robustness to outlier data during subsequent clustering calculations. By iteratively updating the variational factors of the variational distribution family based on the user feature matrix and the probability density model, the model parameters can be efficiently solved using variational inference. Compared to existing technologies that are sensitive to outliers and suffer from high-dimensional data computational bottlenecks, this application can improve the accuracy and efficiency of game user behavior analysis.

[0010] Furthermore, before retrieving the behavior logs of each user within a set time period from the message queue, the process further includes: By deploying a tracking SDK on the game server, the behavioral events of each user are monitored in real time. The behavioral events of each user are aggregated into behavioral logs based on the user ID and stored in the message queue.

[0011] This invention provides a stable data source for subsequent real-time feature calculation and analysis by deploying a tracking SDK before data acquisition and aggregating and storing behavioral events to a message queue.

[0012] Furthermore, the feature engineering performed on each of the aforementioned behavior logs to obtain a user feature matrix includes: The first user feature vector of each user is obtained by aggregating and calculating the behavior logs through a sliding window; wherein, if there are missing values ​​in the first user feature vector, the missing values ​​are filled with the median of the feature values ​​corresponding to the non-missing samples. Each of the first user feature vectors is z-score normalized to obtain multiple user feature vectors, and a user feature matrix is ​​constructed based on each of the user feature vectors.

[0013] The embodiments of the present invention employ sliding window aggregation, median filling of missing values, and z-score standardization for feature engineering, which can generate a standardized feature matrix that is uniform in scale, insensitive to outliers, and complete in information.

[0014] Furthermore, the construction of the probability density model based on the Student-t distribution includes: User behavior is modeled using the Student-t distribution. By introducing a scaling variable, the Student-t distribution is decomposed into a scaling mixture of Gaussian and Gamma distributions to obtain a probability density model. Specifically, the probability density model is as follows: ; in, Represents the user feature vector; Indicates the first The group center vector of each group; Indicates the first Behavioral concentration parameters for each group; This represents the anomaly tolerance parameter; Represents a scale variable; express The identity matrix.

[0015] This invention constructs a probability model by introducing a scaling variable to decompose the Student-t distribution into a scaling mixture of Gaussian and Gamma, which provides a mathematical model basis for automatically reducing the weight of abnormal users in subsequent variational inference.

[0016] Furthermore, the variational distribution family is set as follows: ; in, Indicate the posterior distribution; Indicates mixed weights; Indicates the first Group parameters for each group; Represents a set of group tags; Represents a set of scale variables.

[0017] By setting the variational distribution family as an approximate posterior with a specific decomposition form and factor distribution, this invention can utilize the properties of mean-field approximation and conjugate priors to enable efficient updating of the posterior inference of all model parameters through closed-form analytical formulas.

[0018] Furthermore, before iteratively updating each variational factor based on the user feature matrix and the probability density model, the method further includes: The K-means++ algorithm was used to calculate multiple initial group center vectors, and prior parameters were set.

[0019] The embodiments of the present invention use the K-means++ algorithm to initialize the group centers and set prior parameters before iterative updates, which can provide a stable starting point for variational inference and avoid the problems of slow convergence speed or poor results caused by random initialization.

[0020] Further, the step of iteratively updating each variational factor based on the user feature matrix and the probability density model until a preset iteration termination condition is met, and determining and outputting the group label of each user based on the user affiliation probability of the current iteration, includes: During each iteration, the scaling variable, user affiliation probability, mixture weights, and group parameters are updated sequentially based on the distance between each user feature vector and the group center vector of the current iteration. For each user affiliation probability, calculate the change between the value of the current iteration and the value of the previous iteration. If the maximum value of each change is less than a preset threshold, then the current iteration is determined to meet the iteration termination sub-condition. If the iterations for a preset number of consecutive iterations all satisfy the iteration termination sub-condition or the total number of iterations reaches the preset maximum number of iterations, then the iteration termination condition is determined to be satisfied, and the group label of each user is determined and output based on the user affiliation probability of the current iteration.

[0021] This invention provides a clear convergence criterion for the variational inference process by designing an iteration termination condition based on the maximum change in user affiliation probability. This ensures that the algorithm automatically stops when it reaches sufficient accuracy and protects against non-convergence by using the maximum number of iterations.

[0022] Furthermore, in each iteration, after sequentially updating the scale variable, user affiliation probability, mixture weights, and group parameters based on the distance between each user feature vector and the group center vector of the current iteration, the process further includes: The Newton-Raphson algorithm is used to iteratively solve the fixed-point equation constructed from the anomaly tolerance parameters to obtain the updated anomaly tolerance parameters.

[0023] The embodiments of the present invention use the Newton-Raphson algorithm to adaptively update the anomaly tolerance parameter after the main iteration, which can automatically learn the optimal robustness based on the noise level of the current data.

[0024] Further, the step of comparing each group tag with the corresponding user's historical group tags and determining the change in each user's behavioral status based on the comparison results includes: Each group tag is stored in a database and compared with the historical group tags stored in the database; If a user's group tag is different from their historical group tag, it is determined that the user's behavior status has changed; when the user's behavior status changes, the game server's operation interface is invoked to execute a preset action.

[0025] This invention enables an automated closed loop from user behavior analysis to operational strategy execution by comparing the grouping results in a database and automatically calling the operation interface when changes in user groups are detected.

[0026] Another embodiment of the present invention provides a game user behavior intelligent analysis system, including: a data acquisition module, a probability modeling module, a variational inference module, and a behavior state module; The data acquisition module is used to acquire the behavior logs of each user within a set time period from the message queue, and perform feature engineering on each behavior log to obtain a user feature matrix; wherein, each row of the user feature matrix corresponds to a user feature vector. The probability modeling module is used to construct a probability density model based on the Student-t distribution and set the variational distribution family corresponding to the probability density model; wherein, the variational factors of the variational distribution family include mixed weights, group parameters, user affiliation probabilities and scale variables, and the group parameters include group center vectors; The variational inference module is used to iteratively update each variational factor based on the user feature matrix and the probability density model until a preset iteration termination condition is met, and to determine and output the group label of each user based on the user affiliation probability of the current iteration; wherein, in each iteration, the scaling variable, user affiliation probability, mixture weight and group parameter are updated sequentially based on the distance between each user feature vector and the group center vector of the current iteration. The behavior status module is used to compare each group tag with the corresponding user's historical group tags, and determine the change in the behavior status of each user based on the comparison results. Attached Figure Description

[0027] Figure 1 A flowchart illustrating one embodiment of the intelligent analysis method for game user behavior provided by the present invention; Figure 2 This is a schematic diagram of the structure of one embodiment of the intelligent analysis system for game user behavior provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0030] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0031] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0032] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0033] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0034] See Figure 1 To address the issues of existing technologies being sensitive to outliers and suffering from high-dimensional data computation bottlenecks, an embodiment of the present invention provides a method for intelligent analysis of game user behavior, including steps S101 to S104: Step S101: Obtain the behavior logs of each user within a set time period from the message queue, and perform feature engineering on each behavior log to obtain a user feature matrix; wherein, each row of the user feature matrix corresponds to a user feature vector.

[0035] Preferably, before retrieving the behavior logs of each user within a set time period from the message queue, the method further includes: By deploying a tracking SDK on the game server, the behavioral events of each user are monitored in real time. The behavioral events of each user are aggregated into behavioral logs based on the user ID and stored in the message queue.

[0036] In one embodiment, user behavior events may include login behavior, payment behavior, combat behavior, and social behavior; the message queue may be a Kafka message queue, and an asynchronous non-blocking IO mode may be used to avoid affecting the performance of the game server; the behavior log may be in JSON format, and the log content may include fields such as timestamp, user_id, event_type, and event_data.

[0037] It should be noted that retrieving the behavior logs of each user within a specified time period from the message queue and performing feature engineering on each of the behavior logs to obtain the user feature matrix means: consuming the behavior logs reported by all users within a specified time range from the message queue; then, parsing, aggregating, and transforming these behavior logs to organize the multi-dimensional behavior features of each user into a feature vector, and arranging all user feature vectors in rows to form a structured user feature matrix, which serves as the input data for subsequent probabilistic modeling and variational inference.

[0038] Preferably, the step of performing feature engineering on each of the behavior logs to obtain a user feature matrix includes: The first user feature vector of each user is obtained by aggregating and calculating the behavior logs through a sliding window; wherein, if there are missing values ​​in the first user feature vector, the missing values ​​are filled with the median of the feature values ​​corresponding to the non-missing samples. Each of the first user feature vectors is z-score normalized to obtain multiple user feature vectors, and a user feature matrix is ​​constructed based on each of the user feature vectors.

[0039] In one embodiment, user behavior logs from the past 7 days are consumed hourly from a Kafka message queue, and a first user feature vector is calculated by aggregation using a sliding window. The dimensions of the first user feature vector may include payment dimensions (such as cumulative recharge amount, number of recharges, or the time of the most recent recharge), activity dimensions (such as login days, online time, or activity rating), social dimensions (such as number of friends, guild contribution, or chat frequency), and combat dimensions (such as number of battles, win rate, or average combat power), totaling approximately 50-100 dimensions.

[0040] Furthermore, let the first user feature matrix, composed of the feature vectors of each first user, be... ,in Total number of users This represents the total number of dimensions. Because some users in the game operation data may have missing behaviors (such as not participating in social activities or having empty social characteristics), it is necessary to impute the missing values ​​in each first-user feature vector. Since mean imputation is easily affected by extreme values, while the median is not sensitive to outliers, a median imputation strategy can be used. For example, if a first-user feature vector has missing values... Then calculate the first feature vector of other first users. The median of the eigenvalues and use Fill in the missing value.

[0041] Furthermore, the mean and standard deviation of each feature in each first user feature vector are calculated, and first user feature vectors exceeding three times the standard deviation are marked. For example, if... Then the first user feature vector In features The above is marked as "suspected abnormality", among which, For the first The first user feature vector of the first user feature vector 1 eigenvalue, For each first user feature vector, the first... The mean of the eigenvalues, For each first user feature vector, the first... The standard deviation of each feature value. In particular, the first user feature vector containing outliers is only marked and not deleted, and the outlier data is retained for automatic weight reduction in subsequent modeling.

[0042] Furthermore, to eliminate the influence of different dimensions of features, z-score standardization can be performed on each first user feature vector. The standardization formula is as follows: ,in for The standardized user feature vector. After standardization, each feature has a mean of 0 and a variance of 1, which significantly reduces the modeling error of the spherical covariance assumption. Finally, the first user feature matrix... After feature engineering, it is converted into a standardized user feature matrix. And mark the anomaly vector. .

[0043] Step S102: Construct a probability density model based on the Student-t distribution and set the variational distribution family corresponding to the probability density model; wherein, the variational factors of the variational distribution family include mixed weights, group parameters, user affiliation probabilities and scale variables, and the group parameters include group center vectors.

[0044] It should be noted that constructing a probability density model based on the Student-t distribution means: to establish a mixture probability model with the Student-t distribution as its basic component to describe the statistical distribution of game user behavior characteristics. This model divides all users into... There are several potential groups, and it is assumed that the behavioral feature vectors of users in each group follow a Student-t distribution with a specific center, precision parameter, and degrees of freedom. By introducing a latent scaling variable, the Student-t distribution can be expressed as a scaling mixture of the Gaussian and Gamma distributions, which is mathematically convenient for subsequent variational Bayesian inference.

[0045] Preferably, the construction of the probability density model based on the Student-t distribution includes: User behavior is modeled using the Student-t distribution. By introducing a scaling variable, the Student-t distribution is decomposed into a scaling mixture of Gaussian and Gamma distributions to obtain a probability density model. Specifically, the probability density model is as follows: ; in, Represents the user feature vector; Indicates the first The group center vector of each group; Indicates the first Behavioral concentration parameters for each group; This represents the anomaly tolerance parameter; Represents a scale variable; express The identity matrix.

[0046] In one embodiment, the group center vector represents the typical behavioral characteristics of a user group, such as the typical behavioral characteristics of a "high-value user" being a monthly recharge of 2000 yuan, logging in 25 days / month, or online for 100 hours / month. The behavioral concentration parameter characterizes the degree of similarity in user behavior within the group. The larger the value, the more concentrated the user behavior. The anomaly tolerance parameter is a degree-of-freedom parameter and can be learned automatically. The smaller the value, the more robust it is to anomalies. It degenerates into Gaussian over time. The scaling variable is the user reliability weight, which can be automatically calculated, and abnormal users are considered. Smaller size to reduce its impact on the group center.

[0047] The embodiments of the present invention are passed through The posterior expectation is used to implement a data-driven automatic weight reduction mechanism. The formula for the posterior expectation is as follows: ; For abnormal users who are far from the group center ( (very large) It automatically shrinks, reducing its contribution to the group center. For normal users ( (smaller) Keep the weight close to 1 and maintain normal weight.

[0048] The embodiments of the present invention assign reasonable probabilities to outliers through heavy-tailed distribution (polynomial decay rather than exponential decay), and process outlier data by reducing weights rather than isolating them, thereby maintaining the stability of the group structure.

[0049] Specifically, targeting high-dimensional user behavior characteristics ( To address the computational bottleneck caused by [the aforementioned issue], this embodiment of the invention employs the assumption of spherical covariance. A scalar precision parameter is used. Replace the complete covariance matrix : ; in, for The identity matrix, which means that each feature dimension has the same variance. Furthermore, the features are uncorrelated (covariance is 0). After z-score standardization, the variance of each dimension of the game user behavior features is close to 1. The modeling error of the spherical assumption is controllable, the information loss in the game scenario is small, and the cost-effectiveness is high (the improvement in computational efficiency far outweighs the loss in accuracy).

[0050] It should be noted that setting the variational distribution family corresponding to the probability density model refers to: for efficient variational Bayesian inference of the probability density model, a set of approximate posterior distributions with a defined form and controlled by a set of variational parameters is predefined, i.e., the variational distribution family. Specifically, this step uses the mean-field approximation to approximately decompose the joint posterior distribution of all unknown variables in the model into the product of several independent variational factors. Each variational factor is constrained to a conjugate distribution of a specific form.

[0051] Preferably, the variational distribution family is configured as follows: ; in, Indicate the posterior distribution; Indicates mixed weights; Indicates the first Group parameters for each group; Represents a set of group tags; Represents a set of scale variables.

[0052] In one embodiment, each variational factor adopts a posterior form of a conjugate prior. Specifically, the mixed-weight posterior is: ,in , These are prior parameters. For groups The effective number of samples. The posterior group parameter is: That is, the Normal-Gamma distribution, where , , For the posterior mean center, For scalar precision parameters, For the Gamma shape parameter, Here is the Gamma rate parameter. The posterior user attribution is: ,in For the degree of responsibility, i.e., user Belongs to group The probability of user affiliation. The posterior of the scaling variable is: .

[0053] Step S103: Based on the user feature matrix and the probability density model, iteratively update each variational factor until a preset iteration termination condition is met. Determine and output the group label of each user based on the user affiliation probability of the current iteration. In each iteration, based on the distance between each user feature vector and the group center vector of the current iteration, update the scale variable, user affiliation probability, mixture weight and group parameters in sequence.

[0054] It should be noted that, based on the user feature matrix and the probability density model, iteratively updating each variational factor until a preset iteration termination condition is met, and determining and outputting the group label for each user based on the user affiliation probability of the current iteration, refers to: using the user feature matrix as observation data, and within the framework of the probability density model and its variational distribution family, performing the core calculation process of variational Bayesian inference. Specifically, this process employs the Coordinate Ascent Variational Inference (CAVI) method. In each iteration, a set of closed-form update formulas are executed sequentially based on the current parameter values. This iterative cycle continues until a preset convergence condition is met. When the iteration terminates, the group index corresponding to the maximum user affiliation probability of each user in the last iteration is taken as the final group label for that user and output, thus completing the mapping from the probability model to the determined grouping result.

[0055] Preferably, before iteratively updating each variational factor based on the user feature matrix and the probability density model, the method further includes: The K-means++ algorithm was used to calculate multiple initial group center vectors, and prior parameters were set.

[0056] In one embodiment, the prior parameter can be set to .

[0057] Preferably, the step of iteratively updating each variational factor based on the user feature matrix and the probability density model until a preset iteration termination condition is met, and determining and outputting the group label of each user based on the user affiliation probability of the current iteration, includes: During each iteration, the scaling variable, user affiliation probability, mixture weights, and group parameters are updated sequentially based on the distance between each user feature vector and the group center vector of the current iteration. For each user affiliation probability, calculate the change between the value of the current iteration and the value of the previous iteration. If the maximum value of each change is less than a preset threshold, then the current iteration is determined to meet the iteration termination sub-condition. If the iterations for a preset number of consecutive iterations all satisfy the iteration termination sub-condition or the total number of iterations reaches the preset maximum number of iterations, then the iteration termination condition is determined to be satisfied, and the group label of each user is determined and output based on the user affiliation probability of the current iteration.

[0058] In one embodiment, the first step of the iterative update is to update the scale variable. Given a user Belonging to a group When the scaling variable is used, the formula for calculating the posterior parameter is: ; ; The formula for calculating the posterior expectation used to reduce the weight of abnormal users is: .

[0059] Furthermore, the second step of the iterative update is to update the user affiliation probability, i.e., the responsibility score. First, the logarithmic responsibility score is calculated (using the log-softmax form, which is numerically stable): ; Then, the logarithmic responsibility score is normalized to obtain the responsibility score: ; Subtracting the maximum value can prevent exponential overflow.

[0060] Furthermore, the third step of the iterative update is to update the mixture weights. The formula for calculating the Dirichlet posterior parameters is: ; in, For groups The number of valid samples (soft count).

[0061] Furthermore, the fourth step of the iterative update is to update the group parameters. First, the weighted statistic is calculated: (Weighted valid sample size); (Weighted mean); (Weighted divergence); Then calculate the Normal-Gamma posterior parameters: ; (Posterior mean center); (Gamma shape parameter); (Gamma rate parameter); Finally, calculate the posterior expectation for the next iteration: ; .

[0062] In this embodiment of the invention, the first and second steps of the iterative update require traversing all users and groups and calculating... and The computational complexity is The third and fourth steps of the iterative update only traverse the group, calculate statistics, and update parameters, resulting in a computational complexity of O(n log n). The total computational complexity of a single iteration is... Linear to the number of users Number of groups and feature dimensions .

[0063] Preferably, in each iteration, after sequentially updating the scale variable, user affiliation probability, mixture weight, and group parameters based on the distance between each user feature vector and the group center vector of the current iteration, the method further includes: The Newton-Raphson algorithm is used to iteratively solve the fixed-point equation constructed from the anomaly tolerance parameters to obtain the updated anomaly tolerance parameters.

[0064] In one embodiment, the optimal anomaly tolerance parameter is automatically learned through fixed-point iteration. To eliminate the overhead of manual tuning. Regarding the lower bound of evidence (ELBO)... Taking the partial derivative and setting it to zero, we obtain the fixed-point equation: ; in, For the digamma function (the logarithmic derivative of the Gamma function); For the average term, This equation represents the optimal Make the theoretical value With data statistics They are equal. Since the fixed-point equation is nonlinear, iteratively solving it using the Newton-Raphson method is as follows: initialization: (Moderate robustness); Calculate the function value: ; Calculate the derivative: ,in Let be the derivative of the trigamma function and digamma. renew: ; To prevent crossing the boundary: (make sure ); Convergence criterion: If Then it converges.

[0065] In this embodiment of the invention, after completing one outer iteration of variational inference, that is, after updating all parameters, one... The Newton-Raphson inner loop achieves adaptive learning while avoiding overcomputation.

[0066] Step S104: Compare each group tag with the corresponding user's historical group tags, and determine the change in the user's behavior status based on the comparison results.

[0067] It should be noted that comparing each group label with the corresponding user's historical group labels and determining the change in each user's behavioral state based on the comparison results means: taking the group label of each user output after the convergence of this iteration as the user's current group label, writing it into the database, and comparing it with the user's historical group labels stored in the database one by one. Specifically, for each user, their latest group label is read and compared with the group label determined by the user in the last grouping calculation. If the two match, it is determined that the user's behavioral state has not undergone a significant migration across groups; if the two do not match, it is determined that the user's behavioral state has undergone a change sufficient to change their group affiliation, and the specific direction of the change is recorded.

[0068] Preferably, the step of comparing each group tag with the corresponding user's historical group tags and determining the change in each user's behavioral status based on the comparison results includes: Each group tag is stored in a database and compared with the historical group tags stored in the database; If a user's group tag is different from their historical group tag, it is determined that the user's behavior status has changed; when the user's behavior status changes, the game server's operation interface is invoked to execute a preset action.

[0069] In one embodiment, the clustering results can be written to a PostgreSQL operational database (persistent storage) and a Redis cache (for fast queries). The data table structure can be: user_cluster_result(user_id BIGINT, cluster_id INT, cluster_name VARCHAR, cluster_center JSONB, update_timeTIMESTAMP). B-tree indexes can also be created on user_id and cluster_id to support fast join queries when implementing operational strategies.

[0070] In one embodiment, differentiated operational strategies can be automatically triggered based on user grouping results. Specifically, using PostgreSQL database triggers (TRIGGER) or scheduled tasks (Cron), when user groups change, the game server's operational interface is invoked to execute corresponding actions. For example, a "return gift" can be sent to "potential churned users" (such as those who haven't logged in for three consecutive days but previously had high activity); a dedicated customer service representative can be assigned to "high-value users" (such as those with monthly recharges > 1000 yuan) with increased priority; and a first-time recharge discount can be sent to "active non-paying users" (such as those who log in frequently but haven't recharged).

[0071] In one embodiment, a web-based visualization dashboard (based on ECharts or D3.js) can also be provided to operations personnel to display statistical analysis of the clustering results. Functions may include: group profiles, i.e., the characteristic mean of each group (e.g., the average monthly recharge of "high-value users" is 2000 yuan); user distribution, i.e., the number and percentage of users in each group (displayed using pie charts or bar charts); behavioral patterns, i.e., the migration path of users between different groups (displayed using Sankey diagrams); and operational effectiveness, i.e., indicators such as conversion rate and retention rate after strategy implementation (displayed using line charts).

[0072] This invention, through obtaining behavioral logs within a set time period from a message queue and performing feature engineering, can transform scattered raw events into a structured user feature matrix. By constructing a probability density model based on the Student-t distribution, a probability density model with heavy-tailed distribution characteristics can be established for game user behavior, thereby improving the robustness to outlier data during subsequent clustering calculations. By iteratively updating the variational factors of the variational distribution family based on the user feature matrix and the probability density model, the model parameters can be efficiently solved using variational inference. Compared to existing technologies that are sensitive to outliers and suffer from high-dimensional data computational bottlenecks, this application can improve the accuracy and efficiency of game user behavior analysis.

[0073] Optionally, in this embodiment of the invention, before obtaining the behavior logs of each user within a set time period from the message queue, the method further includes: By deploying a tracking SDK on the game server, the behavioral events of each user are monitored in real time. The behavioral events of each user are aggregated into behavioral logs based on the user ID and stored in the message queue.

[0074] This invention provides a stable data source for subsequent real-time feature calculation and analysis by deploying a tracking SDK before data acquisition and aggregating and storing behavioral events to a message queue.

[0075] Optionally, in this embodiment of the invention, the step of performing feature engineering on each of the behavior logs to obtain a user feature matrix includes: The first user feature vector of each user is obtained by aggregating and calculating the behavior logs through a sliding window; wherein, if there are missing values ​​in the first user feature vector, the missing values ​​are filled with the median of the feature values ​​corresponding to the non-missing samples. Each of the first user feature vectors is z-score normalized to obtain multiple user feature vectors, and a user feature matrix is ​​constructed based on each of the user feature vectors.

[0076] The embodiments of the present invention employ sliding window aggregation, median filling of missing values, and z-score standardization for feature engineering, which can generate a standardized feature matrix that is uniform in scale, insensitive to outliers, and complete in information.

[0077] Optionally, in this embodiment of the invention, constructing a probability density model based on the Student-t distribution includes: User behavior is modeled using the Student-t distribution. By introducing a scaling variable, the Student-t distribution is decomposed into a scaling mixture of Gaussian and Gamma distributions to obtain a probability density model. Specifically, the probability density model is as follows: ; in, Represents the user feature vector; Indicates the first The group center vector of each group; Indicates the first Behavioral concentration parameters for each group; This represents the anomaly tolerance parameter; Represents a scale variable; express The identity matrix.

[0078] This invention constructs a probability model by introducing a scaling variable to decompose the Student-t distribution into a scaling mixture of Gaussian and Gamma, which provides a mathematical model basis for automatically reducing the weight of abnormal users in subsequent variational inference.

[0079] Optionally, in an embodiment of the present invention, the variational distribution family is configured as follows: ; in, Indicate the posterior distribution; Indicates mixed weights; Indicates the first Group parameters for each group; Represents a set of group tags; Represents a set of scale variables.

[0080] By setting the variational distribution family as an approximate posterior with a specific decomposition form and factor distribution, this invention can utilize the properties of mean-field approximation and conjugate priors to enable efficient updating of the posterior inference of all model parameters through closed-form analytical formulas.

[0081] Optionally, in this embodiment of the invention, before iteratively updating each variational factor based on the user feature matrix and the probability density model, the method further includes: The K-means++ algorithm was used to calculate multiple initial group center vectors, and prior parameters were set.

[0082] The embodiments of the present invention use the K-means++ algorithm to initialize the group centers and set prior parameters before iterative updates, which can provide a stable starting point for variational inference and avoid the problems of slow convergence speed or poor results caused by random initialization.

[0083] Optionally, in this embodiment of the invention, the step of iteratively updating each variational factor based on the user feature matrix and the probability density model until a preset iteration termination condition is met, and determining and outputting the group label of each user based on the user affiliation probability of the current iteration, includes: During each iteration, the scaling variable, user affiliation probability, mixture weights, and group parameters are updated sequentially based on the distance between each user feature vector and the group center vector of the current iteration. For each user affiliation probability, calculate the change between the value of the current iteration and the value of the previous iteration. If the maximum value of each change is less than a preset threshold, then the current iteration is determined to meet the iteration termination sub-condition. If the iterations for a preset number of consecutive iterations all satisfy the iteration termination sub-condition or the total number of iterations reaches the preset maximum number of iterations, then the iteration termination condition is determined to be satisfied, and the group label of each user is determined and output based on the user affiliation probability of the current iteration.

[0084] This invention provides a clear convergence criterion for the variational inference process by designing an iteration termination condition based on the maximum change in user affiliation probability. This ensures that the algorithm automatically stops when it reaches sufficient accuracy and protects against non-convergence by using the maximum number of iterations.

[0085] Optionally, in this embodiment of the invention, after updating the scale variable, user affiliation probability, mixture weight, and group parameters sequentially based on the distance between each user feature vector and the group center vector of the current iteration in each iteration, the method further includes: The Newton-Raphson algorithm is used to iteratively solve the fixed-point equation constructed from the anomaly tolerance parameters to obtain the updated anomaly tolerance parameters.

[0086] The embodiments of the present invention use the Newton-Raphson algorithm to adaptively update the anomaly tolerance parameter after the main iteration, which can automatically learn the optimal robustness based on the noise level of the current data.

[0087] Optionally, in this embodiment of the invention, comparing each group tag with the corresponding user's historical group tags and determining the change in each user's behavioral status based on the comparison results includes: Each group tag is stored in a database and compared with the historical group tags stored in the database; If a user's group tag is different from their historical group tag, it is determined that the user's behavior status has changed; when the user's behavior status changes, the game server's operation interface is invoked to execute a preset action.

[0088] This invention enables an automated closed loop from user behavior analysis to operational strategy execution by comparing the grouping results in a database and automatically calling the operation interface when changes in user groups are detected.

[0089] like Figure 2 As shown, based on the above method embodiments, corresponding system embodiments are provided; One embodiment of the present invention provides a game user behavior intelligent analysis system, including: a data acquisition module 201, a probability modeling module 202, a variational inference module 203, and a behavior state module 204; The data acquisition module 201 is used to acquire the behavior logs of each user within a set time period from the message queue, and perform feature engineering on each behavior log to obtain a user feature matrix; wherein, each row of the user feature matrix corresponds to a user feature vector. The probability modeling module 202 is used to construct a probability density model based on the Student-t distribution and set the variational distribution family corresponding to the probability density model; wherein, the variational factors of the variational distribution family include mixed weights, group parameters, user affiliation probabilities and scale variables, and the group parameters include group center vectors; The variational inference module 203 is used to iteratively update each variational factor based on the user feature matrix and the probability density model until a preset iteration termination condition is met, and to determine and output the group label of each user based on the user affiliation probability of the current iteration; wherein, in each iteration, the scaling variable, user affiliation probability, mixture weight and group parameter are updated sequentially based on the distance between each user feature vector and the group center vector of the current iteration. The behavior status module 204 is used to compare each group tag with the corresponding user's historical group tags, and determine the change in the behavior status of each user based on the comparison results.

[0090] Optionally, in this embodiment of the invention, a behavior monitoring submodule and a behavior storage submodule are further included before the data acquisition module 201; The behavior monitoring submodule is used to monitor the behavior events of each user in real time by deploying a tracking SDK on the game server. The behavior storage submodule is used to aggregate the behavior events of each user into behavior logs based on the user ID and store them in the message queue.

[0091] This invention provides a stable data source for subsequent real-time feature calculation and analysis by deploying a tracking SDK before data acquisition and aggregating and storing behavioral events to a message queue.

[0092] Optionally, in this embodiment of the invention, the data acquisition module 201 includes: a first feature engineering submodule and a second feature engineering submodule; The first feature engineering submodule is used to aggregate and calculate each of the behavior logs through a sliding window to obtain the first user feature vector of each user; wherein, if there are missing values ​​in the first user feature vector, the missing values ​​are filled with the median of the feature values ​​corresponding to the non-missing samples. The second feature engineering submodule is used to perform z-score normalization on each of the first user feature vectors to obtain multiple user feature vectors, and to construct a user feature matrix based on each of the user feature vectors.

[0093] The embodiments of the present invention employ sliding window aggregation, median filling of missing values, and z-score standardization for feature engineering, which can generate a standardized feature matrix that is uniform in scale, insensitive to outliers, and complete in information.

[0094] Optionally, in this embodiment of the invention, the probability modeling module 202 includes: a probability modeling submodule; The probability modeling submodule is used to model user behavior using the Student-t distribution, and by introducing a scaling variable, decomposes the Student-t distribution into a scale mixture of Gaussian and Gamma distributions to obtain a probability density model; wherein, the probability density model is specifically as follows: ; in, Represents the user feature vector; Indicates the first The group center vector of each group; Indicates the first Behavioral concentration parameters for each group; This represents the anomaly tolerance parameter; Represents a scale variable; express The identity matrix.

[0095] This invention constructs a probability model by introducing a scaling variable to decompose the Student-t distribution into a scaling mixture of Gaussian and Gamma, which provides a mathematical model basis for automatically reducing the weight of abnormal users in subsequent variational inference.

[0096] Optionally, in this embodiment of the invention, a parameter setting submodule is further included before the variational inference module 203; The parameter setting submodule is used to calculate multiple initial group center vectors using the K-means++ algorithm and set prior parameters.

[0097] The embodiments of the present invention use the K-means++ algorithm to initialize the group centers and set prior parameters before iterative updates, which can provide a stable starting point for variational inference and avoid the problems of slow convergence speed or poor results caused by random initialization.

[0098] Optionally, in this embodiment of the invention, the variational inference module 203 includes: a parameter update submodule, a first termination condition submodule, and a second termination condition submodule; The parameter update submodule is used to update the scale variable, user affiliation probability, mixture weight and group parameters sequentially based on the distance between each user feature vector and the group center vector of the current iteration in each iteration. The first termination condition submodule is used to calculate the change between the value of the current iteration and the value of the previous iteration for each user's belonging probability. If the maximum value of each change is less than a preset threshold, the current iteration is determined to meet the iteration termination sub-condition. The second termination condition submodule is used to determine that the iteration termination condition is met if the iteration termination subcondition is met for a preset number of consecutive iterations or the total number of iterations reaches a preset maximum number of iterations, and to determine and output the group label of each user based on the user affiliation probability of the current iteration.

[0099] This invention provides a clear convergence criterion for the variational inference process by designing an iteration termination condition based on the maximum change in user affiliation probability. This ensures that the algorithm automatically stops when it reaches sufficient accuracy and protects against non-convergence by using the maximum number of iterations.

[0100] Optionally, in this embodiment of the invention, after the parameter update submodule, an anomaly tolerance update submodule is further included; The anomaly tolerance update submodule is used to iteratively solve the fixed-point equation constructed from the anomaly tolerance parameters using the Newton-Raphson algorithm to obtain the updated anomaly tolerance parameters.

[0101] The embodiments of the present invention use the Newton-Raphson algorithm to adaptively update the anomaly tolerance parameter after the main iteration, which can automatically learn the optimal robustness based on the noise level of the current data.

[0102] Optionally, in this embodiment of the invention, the behavior status module 204 includes: a tag comparison submodule and an operation execution submodule; The tag comparison submodule is used to store each group tag in a database and compare it with the historical group tags stored in the database; The operation execution submodule is used to determine that the user's behavior status has changed if the user's group tag is different from the historical group tag; wherein, when the user's behavior status changes, the operation interface of the game server is called to execute a preset action.

[0103] This invention enables an automated closed loop from user behavior analysis to operational strategy execution by comparing the grouping results in a database and automatically calling the operation interface when changes in user groups are detected.

[0104] It is understood that the above system item embodiments correspond to the method item embodiments of the present invention, and can implement the game user behavior intelligent analysis method provided by any of the above method item embodiments of the present invention.

[0105] This invention employs a data acquisition module 201 to retrieve behavioral logs within a specified time period from a message queue and performs feature engineering, transforming scattered raw events into a structured user feature matrix. A probability modeling module 202 constructs a probability density model based on the Student-t distribution, establishing a probability density model with heavy-tailed distribution characteristics for game user behavior, thereby improving robustness to outlier data during subsequent clustering calculations. A variational inference module 203 iteratively updates the variational factors of the variational distribution family based on the user feature matrix and the probability density model, enabling efficient solution of model parameters through variational inference. Compared to existing technologies that are sensitive to outliers and suffer from high-dimensional data computation bottlenecks, this application improves the accuracy and efficiency of game user behavior analysis.

[0106] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0107] Based on the above-described embodiment of the intelligent analysis method for game user behavior, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements an intelligent analysis method for game user behavior according to any embodiment of the present invention.

[0108] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0109] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0110] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0111] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the intelligent analysis method for game user behavior described in any of the above-described method embodiments of the present invention.

[0112] The modules / units integrated into the system / terminal device, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0113] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for intelligent analysis of game user behavior, characterized in that, include: Retrieve the behavior logs of each user within a set time period from the message queue, and perform feature engineering on each behavior log to obtain a user feature matrix; wherein, each row of the user feature matrix corresponds to a user feature vector; Construct a probability density model based on the Student-t distribution, and set the variational distribution family corresponding to the probability density model; wherein, the variational factors of the variational distribution family include mixed weights, group parameters, user affiliation probabilities and scale variables, and the group parameters include group center vectors; Based on the user feature matrix and the probability density model, each variational factor is iteratively updated until a preset iteration termination condition is met. The group label of each user is determined and output based on the user affiliation probability of the current iteration. In each iteration, the scaling variable, user affiliation probability, mixture weight and group parameter are updated sequentially based on the distance between each user feature vector and the group center vector of the current iteration. Each group label is compared with the corresponding user's historical group labels, and the changes in the user's behavioral status are determined based on the comparison results.

2. The intelligent analysis method for game user behavior as described in claim 1, characterized in that, Before retrieving the behavior logs of each user within a set time period from the message queue, the method further includes: By deploying a tracking SDK on the game server, the behavioral events of each user are monitored in real time. The behavioral events of each user are aggregated into behavioral logs based on the user ID and stored in the message queue.

3. The intelligent analysis method for game user behavior as described in claim 1, characterized in that, The feature engineering of each of the aforementioned behavior logs to obtain a user feature matrix includes: The first user feature vector of each user is obtained by aggregating and calculating the behavior logs through a sliding window; wherein, if there are missing values ​​in the first user feature vector, the missing values ​​are filled with the median of the feature values ​​corresponding to the non-missing samples. Each of the first user feature vectors is z-score normalized to obtain multiple user feature vectors, and a user feature matrix is ​​constructed based on each of the user feature vectors.

4. The intelligent analysis method for game user behavior as described in claim 1, characterized in that, The construction of the probability density model based on the Student-t distribution includes: User behavior is modeled using the Student-t distribution. By introducing a scaling variable, the Student-t distribution is decomposed into a scaling mixture of Gaussian and Gamma distributions to obtain a probability density model. Specifically, the probability density model is as follows: ； in, Represents the user feature vector; Indicates the first The group center vector of each group; Indicates the first Behavioral concentration parameters for each group; This represents the anomaly tolerance parameter; Represents a scale variable; express The identity matrix.

5. The intelligent analysis method for game user behavior as described in claim 4, characterized in that, The variational distribution family is set as follows: ； in, Indicate the posterior distribution; Indicates mixed weights; Indicates the first Group parameters for each group; Represents a set of group tags; Represents a set of scale variables.

6. The intelligent analysis method for game user behavior as described in claim 1, characterized in that, Before iteratively updating each variational factor based on the user feature matrix and the probability density model, the method further includes: The K-means++ algorithm was used to calculate multiple initial group center vectors, and prior parameters were set.

7. The intelligent analysis method for game user behavior as described in claim 1, characterized in that, The process of iteratively updating each variational factor based on the user feature matrix and the probability density model until a preset iteration termination condition is met, and determining and outputting the group label for each user based on the user affiliation probability of the current iteration, includes: During each iteration, the scaling variable, user affiliation probability, mixture weights, and group parameters are updated sequentially based on the distance between each user feature vector and the group center vector of the current iteration. For each user affiliation probability, calculate the change between the value of the current iteration and the value of the previous iteration. If the maximum value of each change is less than a preset threshold, then the current iteration is determined to meet the iteration termination sub-condition. If the iterations for a preset number of consecutive iterations all satisfy the iteration termination sub-condition or the total number of iterations reaches the preset maximum number of iterations, then the iteration termination condition is determined to be satisfied, and the group label of each user is determined and output based on the user affiliation probability of the current iteration.

8. The intelligent analysis method for game user behavior as described in claim 1, characterized in that, In each iteration, after sequentially updating the scale variable, user affiliation probability, mixture weights, and group parameters based on the distance between each user feature vector and the group center vector of the current iteration, the process further includes: The Newton-Raphson algorithm is used to iteratively solve the fixed-point equation constructed from the anomaly tolerance parameters to obtain the updated anomaly tolerance parameters.

9. The intelligent analysis method for game user behavior as described in claim 1, characterized in that, The step of comparing each group tag with the corresponding user's historical group tags and determining the change in each user's behavioral status based on the comparison results includes: Each group tag is stored in a database and compared with the historical group tags stored in the database; If a user's group tag is different from their historical group tag, it is determined that the user's behavior status has changed; when the user's behavior status changes, the game server's operation interface is invoked to execute a preset action.

10. A game user behavior intelligent analysis system, characterized in that, include: Data acquisition module, probability modeling module, variational inference module, and behavior state module; The data acquisition module is used to acquire the behavior logs of each user within a set time period from the message queue, and perform feature engineering on each behavior log to obtain a user feature matrix; wherein, each row of the user feature matrix corresponds to a user feature vector. The probability modeling module is used to construct a probability density model based on the Student-t distribution and set the variational distribution family corresponding to the probability density model; wherein, the variational factors of the variational distribution family include mixed weights, group parameters, user affiliation probabilities and scale variables, and the group parameters include group center vectors; The variational inference module is used to iteratively update each variational factor based on the user feature matrix and the probability density model until a preset iteration termination condition is met, and to determine and output the group label of each user based on the user affiliation probability of the current iteration; wherein, in each iteration, the scaling variable, user affiliation probability, mixture weight and group parameter are updated sequentially based on the distance between each user feature vector and the group center vector of the current iteration. The behavior status module is used to compare each group tag with the corresponding user's historical group tags, and determine the change in the behavior status of each user based on the comparison results.

Images