A method and system for analyzing user online transaction behavior

By combining dynamic behavioral graphs and an improved EvolveGCN network with a Bayesian uncertainty quantification framework and a reverse game verification mechanism, the problems of missing temporal logical correlations of behaviors and lack of confidence in model predictions in online transactions are solved. This achieves accurate perception, proactive verification, and real-time defense, thereby improving the robustness and defense efficiency of the risk control system.

CN123118260APending Publication Date: 2026-11-17FUJIAN XINAN NETWORK TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611638518.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-22
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing online transaction risk control technologies lack in-depth modeling of the temporal logical relationships of user operations, model predictions lack confidence assessment, and defense strategies are passive with a high false alarm rate, making it difficult to cope with complex and intelligent attack methods.

Method used

By employing dynamic behavior graph construction, an improved EvolveGCN network, a Bayesian uncertainty quantification framework, and a reverse game verification mechanism, a smart risk control system is constructed, which enables dynamic modeling of user habits, measurement of behavioral bias uncertainty, interactive risk verification, and coordinated linkage of defense strategies.

Benefits of technology

It improves the robustness of the model, the accuracy of risk identification, and the speed of system defense response when facing unknown fraud methods, significantly reduces the false alarm rate, and enhances the adaptive evolution capability of the risk control system in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118260A_ABST
    Figure CN123118260A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for analyzing user online transaction behavior, relating to the intersection of network security and artificial intelligence. The method includes the following steps: S1, capturing and cleaning multi-dimensional interaction data to output a standardized behavior dataset; S2, constructing a dynamic behavior graph reflecting temporal and logical dependencies; S3, using an improved EvolveGCN network combined with Bayesian uncertainty quantification to output behavior deviation values ​​and confidence intervals; S4, triggering reverse game verification when the deviation value exceeds a threshold, correcting the deviation value, and outputting a risk assessment result; S5, extracting abnormal behavior patterns and outputting a risk feature encoding set; S6, generating defense rules and executing them on associated user nodes. This invention achieves dynamic modeling and uncertainty measurement of transaction behavior, significantly improves the accuracy of risk identification through active game verification, and enhances the system's ability to defend against complex fraudulent activities.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the intersection of cybersecurity and artificial intelligence, and in particular to a method and system for analyzing user online transaction behavior. Background Technology

[0002] With the rapid development of internet and fintech, online transactions have become an indispensable part of people's daily lives, greatly facilitating payments and consumption. However, the openness and virtuality of the online transaction environment also bring serious security challenges. Account theft, transaction fraud, and machine script attacks are rampant, posing a serious threat to user asset security and platform stability. To address these risks, existing risk control systems typically rely on rule-based detection methods or traditional machine learning models. These traditional methods mainly assess the risk of current transactions by analyzing users' static attribute characteristics or statistical features, such as the transaction amount, the recipient's historical transaction frequency, or the fingerprint information of the device environment.

[0003] While existing technologies offer some degree of transaction security, they still have significant limitations in addressing increasingly complex and sophisticated attack methods. First, traditional detection methods often rely on analyzing independent data samples, lacking in-depth modeling of the logical connections between user actions. In actual online transactions, user actions often exhibit strong temporal dependencies and logical causal relationships, such as the logical transition between login and subsequent transfers. Existing technologies tend to evaluate these actions as isolated points in time, ignoring the spatiotemporal correlations between actions and making it difficult to detect complex fraudulent activities that exploit vulnerabilities in normal operational logic. Second, most existing behavioral analysis models employ deterministic deep learning algorithms. These models output fixed predictions and lack the ability to measure the confidence level of their predictions. When faced with variant fraud or new attack patterns that mimic normal user behavior, deterministic models are prone to incorrect classifications and cannot provide a reliability assessment of the judgment results, leading to a lack of resilience and robustness in risk control systems when facing unknown risks. Furthermore, existing risk control mechanisms mostly employ passive blocking or verification strategies, lacking proactive interactive verification methods. When suspicious behavior is detected, the system often blocks it directly. This approach not only easily harms the user experience of normal users, but also makes it difficult to effectively distinguish between real users and attackers using advanced scripts. Finally, existing defense strategy generation mainly relies on human experience or pre-defined rules, lacking an automated mapping mechanism based on historical successful defense cases, and it is difficult to quickly associate defense strategies with other potentially risky users with similar behavioral characteristics, resulting in lagging defense measures and insufficient coverage.

[0004] Therefore, how to provide a method and system for analyzing user online transaction behavior is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for analyzing user online transaction behavior. This invention fully integrates key steps such as dynamic behavior graph construction, an improved EvolveGCN network, the introduction of a Bayesian uncertainty quantification framework, a reverse game verification mechanism, and associated defense execution. It constructs an intelligent risk control system with dynamic modeling of user habits, measurement of behavioral deviation uncertainty, interactive risk verification, and coordinated linkage of defense strategies. This achieves accurate perception, proactive verification, and real-time defense against abnormal behavior in the complex environment of online transactions. This invention possesses advantages such as comprehensive behavioral logic modeling, accurate risk judgment confidence quantification, strong interactivity of the verification mechanism, and intelligent generation of defense strategies. It can significantly improve model robustness, risk identification accuracy, and system defense response speed when facing unknown fraudulent methods, thereby effectively solving the problems of missing temporal logical correlations in existing technologies, lack of confidence assessment in model predictions, and high false alarm rates in passive defense.

[0006] A method for analyzing user online transaction behavior according to an embodiment of the present invention includes the following steps:

[0007] S1. Capture the click flow, typing rhythm, mouse trajectory and transaction records of users in real time during the online transaction process, calculate the data disorder index of each dimension, remove invalid records and convert the remaining data into matrix form, and output a standardized behavior dataset.

[0008] S2. Treat each action in the standardized behavior dataset as a node, calculate the tightness between nodes according to the order of action occurrence and logical dependencies, connect the nodes into a network structure, and output a dynamic behavior graph.

[0009] S3. Input the dynamic behavior graph into the improved EvolveGCN network, introduce the Bayesian uncertainty quantification framework, construct the weight parameters of the graph convolutional layer as a distribution of random variables that evolves dynamically over time, minimize the lower bound of evidence through variational inference to optimize the distribution parameters, generate the expected operation path and confidence interval, compare the expected operation path with the user's real-time operation path, and output the behavior deviation value.

[0010] S4. When the behavior deviation value exceeds the preset deviation threshold, the reverse game verification mechanism is triggered. The mechanism actively sends a probing command to the user interface, collects the user's response to minor disturbances, corrects the behavior deviation value based on the response, and outputs the risk assessment result.

[0011] S5. Classify the risk assessment results according to the time rhythm of occurrence, extract abnormal behavior patterns with similar time interval characteristics, and output a risk feature code set.

[0012] S6. Generate corresponding defense rules based on the risk feature code set, search for related users in the transaction network whose operating habits match the risk assessment results, and send the defense rules to the related user nodes for execution.

[0013] Optionally, S1 specifically includes:

[0014] S11. Listen to the input interface of the user terminal, capture the coordinate sequence of mouse movement, the timestamp of click operation, the key code and interval of keyboard input according to the set time window, and read the transaction records in the transaction backend database.

[0015] S12. Calculate the spatial location dispersion using the geometric center position of the coordinate sequence, calculate the information entropy of the click timestamp as the temporal disorder, and calculate the variance of the keyboard interval duration as the rhythm fluctuation. The spatial location dispersion, temporal disorder, and rhythm fluctuation are collectively referred to as the data disorder index of each dimension.

[0016] S13. Compare the data disorder index of each dimension with the preset normal threshold range, determine the records that exceed the normal threshold range as invalid records and delete them, and retain the remaining data that are within the normal threshold range.

[0017] S14. Perform numerical mapping encoding on the character action labels in the remaining data, normalize and scale the timestamp data, concatenate the processed data into a matrix by rows and columns, and output the standardized behavior dataset.

[0018] Optionally, S2 specifically includes:

[0019] S21. Read the records in the standardized behavior dataset line by line, define each mouse click event, keyboard key event and transaction confirmation event as an action node, and assign a unique identity number to each action node.

[0020] S22. According to the chronological order of the actions, select two adjacent action nodes in turn, calculate the absolute value of the timestamp difference between the two action nodes, and determine the degree of closeness in the time dimension by the reciprocal of the absolute value of the timestamp difference.

[0021] S23. Identify the page jump relationship and data input / output relationship between the current action node and other action nodes. If there is a page jump or data transfer, it is determined that there is a logical dependency. Set the tightness between the logically dependent node pairs to a preset logical association strength value.

[0022] S24. The tightness in the time dimension and the logical association strength value are weighted and summed. The result is used as the tightness between nodes. If there is no logical dependency between node pairs, only the tightness in the time dimension is used. Node pairs with tightness greater than zero are connected by edges.

[0023] S25. Summarize all action nodes and connecting edges to construct a mesh structure, using the identity number of the action node as the vertex and the density as the edge weight, and output a dynamic behavior graph.

[0024] Optionally, S3 specifically includes:

[0025] S31. Read the node feature matrix and adjacency matrix in the dynamic behavior graph, input the node feature matrix into the improved EvolveGCN network, introduce a long short-term memory network in the graph convolutional layer, concatenate the weight matrix of the previous time step with the input of the current time step, calculate the output of the forget gate and the input gate through the long short-term memory network, calculate the candidate hidden state at the current time step, update the hidden state of the weight matrix, and output the weight parameters that evolve dynamically over time.

[0026] S32. Introduce a Bayesian uncertainty quantification framework, construct the weight parameters that evolve dynamically over time as random variables that follow a normal distribution, calculate the KL divergence and the log-likelihood probability of the prediction results, and output the lower bound of evidence.

[0027] S33. Maximize the lower bound of evidence using the stochastic gradient descent algorithm. Iteratively update the mean and variance parameters of the normal distribution through backpropagation. Stop iterating when the growth rate of the lower bound of evidence is less than the preset convergence threshold, and output the optimized distribution parameters.

[0028] S34. Perform a preset number of Monte Carlo samplings on the optimized distribution parameters. In each sampling, a set of weight values ​​are randomly selected from the normal distribution and assigned to the graph convolutional layer. The adjacency matrix is ​​used to perform forward propagation to calculate the output operation probability. The mean of all sampling results is used as the expected operation path, and the variance range of all sampling results is used as the confidence interval.

[0029] S35. Read the user's real-time operation path, calculate the Euclidean distance between the expected operation path and the user's real-time operation path, normalize the Euclidean distance by dividing it by the half-width of the confidence interval, and output the behavior deviation value.

[0030] Optionally, S32 specifically includes:

[0031] S321. Introduce the Bayesian uncertainty quantification framework, establish two learnable parameter matrices with the same dimensions as the weight parameters as the mean vector and log-variance vector of the weight parameter distribution, respectively. Input the log-variance vector into the exponential function to obtain the standard deviation. Randomly sample from the standard normal distribution to generate a random noise matrix with the same dimensions as the weight parameters. Multiply the random noise matrix with the standard deviation element by element, and add the product to the mean vector element by element to obtain the weight random variable that follows a normal distribution.

[0032] S322. Calculate the distribution distance between the normal distribution formed by the mean vector and the standard deviation and the standard normal distribution. Substitute the weighted random variables that follow the normal distribution into the graph convolutional network for forward propagation to obtain the predicted probability vector. Perform a dot product operation between the predicted probability vector and the true label vector. Take the logarithm of the operation result and take the negative sign to obtain the cross-entropy loss function value. Use the negative of the cross-entropy loss function value as the log-likelihood probability.

[0033] Optionally, S4 specifically includes:

[0034] S41. Read the behavior deviation value, compare the behavior deviation value containing uncertainty measurement with the preset deviation threshold, and trigger the reverse game verification mechanism when the behavior deviation value is greater than the preset deviation threshold.

[0035] S42. Actively send probing commands to the user interface, listen to the user terminal's input interface, collect the user's response to the probing commands, and calculate the response time and operation trajectory offset.

[0036] S43. Correct the behavior deviation value according to the response action, compare the corrected behavior deviation value with the preset deviation threshold. If the corrected behavior deviation value is greater than or equal to the preset threshold, output the risk judgment result as high risk. If the corrected behavior deviation value is less than the preset threshold, output the risk judgment result as low risk.

[0037] Optionally, S5 specifically includes:

[0038] S51. Read the risk assessment results one by one, extract the timestamp information of the behavioral data corresponding to each risk assessment result, and arrange the risk assessment results into a time series according to the order of the timestamp information.

[0039] S52. Calculate the time interval between two adjacent risk assessment results in the time series, compare all time interval values ​​with the preset time interval threshold, and classify the risk assessment results with time interval values ​​less than the preset time interval threshold into the same time group.

[0040] S53. Within the same time group, extract the user operation action types included in the statistical risk judgment results, and combine them to form an abnormal behavior pattern.

[0041] S54. Assign a unique numerical code to each type of operation action in the abnormal behavior pattern. Concatenate the numerical codes into a numerical string according to the order in which the operation action types occur. Output the numerical string as the risk feature code. Summarize the risk feature codes of all time groups to form a risk feature code set.

[0042] Optionally, S6 specifically includes:

[0043] S61. Read the risk feature code set, traverse each risk feature code in the risk feature code set, calculate the similarity value between the risk feature code and the historical feature code stored in the preset code defense rule mapping table, and encapsulate the defense strategy text corresponding to the historical feature code with a similarity value greater than the preset matching threshold into a defense rule.

[0044] S62. Traverse the user nodes in the dynamic behavior graph, calculate the cosine similarity value of the node feature matrix between the current user node and other user nodes, and determine other user nodes whose cosine similarity value is greater than the preset similarity threshold as related users.

[0045] S63. Find the node identity number of the associated user in the dynamic behavior graph, and send the defense rules to the terminal node where the associated user is located according to the node identity number to perform the defense operation.

[0046] A user online transaction behavior analysis system according to an embodiment of the present invention includes:

[0047] The data acquisition and standardization module is used to capture click streams, typing rhythms, mouse trajectories and transaction records during users' online transactions in real time, calculate data disorder indicators in various dimensions, remove invalid records and convert the remaining data into matrix form, and output a standardized behavior dataset.

[0048] The dynamic behavior graph construction module is used to treat each action in the standardized behavior dataset as a node, calculate the tightness between nodes according to the chronological order of the actions and logical dependencies, connect the nodes into a network structure, and output a dynamic behavior graph.

[0049] The uncertainty behavior analysis module is used to input dynamic behavior graphs into the improved EvolveGCN network, introduce a Bayesian uncertainty quantification framework, construct the weight parameters of the graph convolutional layer as a distribution of random variables that evolves dynamically over time, minimize the lower bound of evidence through variational inference to optimize the distribution parameters, generate expected operation paths and confidence intervals, compare the expected operation paths with the user's real-time operation paths, and output the behavior deviation value.

[0050] The reverse game verification module is used to trigger the reverse game verification mechanism when the behavior deviation value exceeds the preset deviation threshold. It actively sends a probing command to the user interface, collects the user's response to minor disturbances, corrects the behavior deviation value based on the response, and outputs the risk assessment result.

[0051] The abnormal pattern coding module is used to classify the risk assessment results according to the time rhythm of occurrence, extract abnormal behavior patterns with similar time interval characteristics, and output a risk feature coding set.

[0052] The associated defense execution module is used to generate corresponding defense rules based on the risk feature encoding set, search for associated users in the transaction network whose operating habits match the risk assessment results, and send the defense rules to the associated user nodes for execution.

[0053] The beneficial effects of this invention are:

[0054] This invention addresses the challenges of high-dimensional, heterogeneous, and noisy transaction data by capturing user network transaction behavior data in real time and performing disorder index calculation and standardization. It constructs a dynamic behavior graph reflecting user transaction habits. By inputting this dynamic behavior graph into an improved EvolveGCN network and introducing a Bayesian uncertainty quantification framework, the weight parameters of the graph convolutional layers are constructed as a dynamically evolving random variable distribution over time. Monte Carlo sampling is used to generate expected operation paths and confidence intervals, outputting a behavior deviation value that includes uncertainty measurement. During the risk assessment phase, when the behavior deviation value exceeds a threshold, a reverse game-theoretic verification mechanism is triggered, proactively sending minor interference probes. The deviation value is corrected based on the user's response, and the risk assessment result is output. By extracting abnormal behavior patterns, a risk feature encoding set is output. Defense rules are generated based on similarity matching, and related user nodes are searched in the transaction network to execute defense operations. Ultimately, this achieves closed-loop risk management through dynamic spatiotemporal modeling of network transaction behavior, accurate measurement of uncertainty, proactive interactive verification, and associated collaborative defense, effectively improving the accuracy of identifying complex fraudulent behaviors, the robustness of risk assessment, and the response speed of defense strategies. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is a flowchart of a user network transaction behavior analysis method proposed in this invention;

[0057] Figure 2 This is a structural diagram of a user network transaction behavior analysis system proposed in this invention;

[0058] Figure 3 This is a flowchart of the behavioral bias calculation based on the improved EvolveGCN network and Bayesian uncertainty quantification proposed in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] refer to Figures 1-3 A method for analyzing user online transaction behavior includes the following steps:

[0061] S1. Capture the click flow, typing rhythm, mouse trajectory and transaction records of users in real time during the online transaction process, calculate the data disorder index of each dimension, remove invalid records and convert the remaining data into matrix form, and output a standardized behavior dataset.

[0062] S2. Treat each action in the standardized behavior dataset as a node, calculate the tightness between nodes based on the chronological order of the actions and logical dependencies, connect the nodes into a network structure, and output a dynamic behavior graph that reflects the user's transaction habits.

[0063] S3. Input the dynamic behavior graph into the improved EvolveGCN network, introduce the Bayesian uncertainty quantification framework, construct the weight parameters of the graph convolutional layer as a distribution of random variables that evolves dynamically over time, minimize the lower bound of evidence through variational inference to optimize the distribution parameters, generate the expected operation path and confidence interval based on Monte Carlo sampling, compare the expected operation path with the user's real-time operation path, and output the behavior deviation value containing uncertainty measurement.

[0064] S4. When the behavioral deviation value exceeds the preset deviation threshold, the reverse game verification mechanism is triggered. The mechanism actively sends a probing command containing minor interference to the user interface, collects the user's response to the minor interference, corrects the behavioral deviation value based on the response, and outputs the risk assessment result.

[0065] S5. Classify the risk assessment results according to the time rhythm of occurrence, extract abnormal behavior patterns with similar time interval characteristics, and output a risk feature code set.

[0066] S6. Generate corresponding defense rules based on the risk feature code set, search for related users in the transaction network whose operating habits match the risk assessment results, and send the defense rules to the related user nodes for execution.

[0067] This invention significantly improves the accuracy, robustness, and proactive defense capabilities of online transaction risk identification. By capturing multi-dimensional interaction data such as clickstreams and typing rhythms in real time and constructing a dynamic behavioral graph, it achieves deep spatiotemporal modeling of user trading habits, effectively compensating for the shortcomings of traditional methods in capturing behavioral logical correlation features. Utilizing an improved EvolveGCN network combined with a Bayesian uncertainty quantification framework, the model weight parameters are dynamically evolved into a random distribution. This not only accurately predicts expected operation paths but also outputs behavioral bias values ​​containing confidence intervals, solving the problems of rigid model outputs and lack of confidence assessment in the face of new fraud methods in existing technologies. In particular, when the behavioral bias value exceeds the threshold, an innovative reverse game verification mechanism is triggered. By actively sending small interference probe commands and analyzing user responses, it achieves accurate differentiation between machine scripts and real users, greatly reducing the false positive rate. Furthermore, by extracting abnormal patterns to generate a risk feature encoding set and executing associated user defense, a closed-loop response system of "single-point identification and network-wide joint defense" is constructed, effectively enhancing the adaptive evolution capability and overall security protection level of the risk control system in complex online transaction environments.

[0068] In this embodiment, S1 specifically includes:

[0069] S11. Listen to the input interface of the user terminal, capture the coordinate sequence of mouse movement, the timestamp of click operation, the key code and interval of keyboard input according to the set time window, and read the transaction records in the transaction backend database.

[0070] S12. Calculate the spatial location dispersion using the geometric center position of the coordinate sequence, calculate the information entropy of the click timestamp as the temporal disorder, and calculate the variance of the keyboard interval duration as the rhythm fluctuation. The spatial location dispersion, temporal disorder, and rhythm fluctuation are collectively referred to as the data disorder index of each dimension. The calculation of spatial location dispersion is to calculate the Euclidean distance from each coordinate point to the geometric center position, and the ratio of the standard deviation to the mean of the Euclidean distance is used as the spatial location dispersion. The calculation of the information entropy of the click timestamp is to statistically analyze the probability distribution of the click timestamp and calculate the negative expected value.

[0071] S13. Compare the data disorder index of each dimension with the preset normal threshold range, determine the records that exceed the normal threshold range as invalid records and delete them, and retain the remaining data within the normal threshold range. The preset normal threshold range is a confidence interval calculated based on the average value of historical normal user behavior data plus or minus three times the standard deviation.

[0072] S14. Perform numerical mapping encoding on the character action labels in the remaining data, normalize and scale the timestamp data, concatenate the processed data into a matrix by rows and columns, and output the standardized behavior dataset.

[0073] In this embodiment, S2 specifically includes:

[0074] S21. Read the records in the standardized behavior dataset line by line, define each mouse click event, keyboard key event and transaction confirmation event as an action node, and assign a unique identity number to each action node.

[0075] S22. According to the chronological order of the actions, select two adjacent action nodes in turn, calculate the absolute value of the timestamp difference between the two action nodes, and determine the degree of closeness in the time dimension by the reciprocal of the absolute value of the timestamp difference.

[0076] S23. Identify the page jump relationship and data input / output relationship between the current action node and other action nodes. If there is a page jump or data transfer, it is determined that there is a logical dependency. Set the tightness between the logically dependent node pairs to a preset logical association strength value. The preset logical association strength value is a connection strength coefficient defined based on expert experience, and the value range is 0.5 to 1.0.

[0077] S24. The tightness in the time dimension and the logical association strength value are weighted and summed. The result is used as the tightness between nodes. If there is no logical dependency between node pairs, only the tightness in the time dimension is used. Node pairs with tightness greater than zero are connected by edges.

[0078] S25. Summarize all action nodes and connecting edges to construct a mesh structure, using the identity number of the action node as the vertex and the density as the edge weight, and output a dynamic behavior graph that reflects the user's trading habits.

[0079] In this embodiment, S3 specifically includes:

[0080] S31. Read the node feature matrix and adjacency matrix from the dynamic behavior graph, input the node feature matrix into the improved EvolveGCN network, introduce a long short-term memory network into the graph convolutional layer, concatenate the weight matrix of the previous time step with the input of the current time step, calculate the output of the forget gate and input gate through the long short-term memory network, calculate the candidate hidden state at the current time step, update the hidden state of the weight matrix, and output the weight parameters that evolve dynamically over time. The calculation of the output of the forget gate and input gate through the long short-term memory network is to multiply the concatenated vector with the weight matrix corresponding to the long short-term memory network and add a bias term, and input the result into the Sigmoid activation function to calculate the forget probability value and the input probability value. The calculation of the candidate hidden state at the current time step is to perform a linear transformation on the concatenated vector and input the result into the Tanh activation function to map it to the interval from -1 to 1. The update of the hidden state of the weight matrix is ​​to calculate the product of the forget probability value and the hidden state at the previous time step, calculate the product of the input probability value and the candidate hidden state at the current time step, and add the two products to obtain the new hidden state value.

[0081] S32. Introduce a Bayesian uncertainty quantification framework, construct the weight parameters that evolve dynamically over time as random variables that follow a normal distribution, calculate the KL divergence and the log-likelihood probability of the prediction results, and output the lower bound of evidence.

[0082] S33. Maximize the lower bound of evidence using the stochastic gradient descent algorithm. Iteratively update the mean and variance parameters of the normal distribution through backpropagation. Stop iterating when the increase of the lower bound of evidence is less than a preset convergence threshold, and output the optimized distribution parameters. The iterative update of the mean and variance parameters of the normal distribution through backpropagation involves calculating the partial derivatives of the lower bound of evidence with respect to the mean and log-variance parameters, and subtracting the product of the learning rate and the partial derivatives from the mean and log-variance parameters to obtain the updated parameter values. Stopping iteration when the increase of the lower bound of evidence is less than the preset convergence threshold involves calculating the absolute value of the difference between the lower bound of evidence in the current iteration and the lower bound of evidence in the previous iteration. When the absolute value of the difference is less than a set value between 0.001 and 0.01 for three consecutive times, the model is determined to have converged and the iteration process is terminated.

[0083] S34. Perform a preset number of Monte Carlo samplings on the optimized distribution parameters. Each sampling randomly selects a set of weight values ​​from the normal distribution and assigns them to the graph convolutional layer. Use the adjacency matrix to perform forward propagation to calculate the output operation probability. Calculate the mean of all sampling results as the expected operation path and the variance range of all sampling results as the confidence interval. The preset number of samplings is calculated based on the central limit theorem to ensure that the standard error of the sampling mean is less than the preset accuracy requirement. The specific steps are as follows: set the allowable standard error of the sampling mean, divide the square of the standard error value as the denominator and the sample variance of the output operation probability as the numerator, and round the result up to obtain the minimum sampling number. Calculating the mean of all sampling results as the expected operation path involves summing the corresponding elements of the operation probability vector obtained by the preset number of samplings and dividing the sum by the number of samplings to obtain the average operation probability vector. Calculating the variance range of all sampling results as the confidence interval involves calculating the square of the difference between each sampling result and the average operation probability vector, summing all the squared differences and dividing by the number of samplings minus one to obtain the variance value.

[0084] S35. Read the user's real-time operation path, calculate the Euclidean distance between the expected operation path and the user's real-time operation path, normalize the Euclidean distance by dividing it by the half-width value of the confidence interval, and output the behavior deviation value containing uncertainty measurement. The half-width value is half of the difference between the maximum and minimum values ​​among all sampled results. Divide half of the difference by the average operation probability vector to obtain the relative deviation value.

[0085] In this embodiment, S32 specifically includes:

[0086] S321. Introduce the Bayesian uncertainty quantification framework, establish two learnable parameter matrices with the same dimensions as the weight parameters as the mean vector and log-variance vector of the weight parameter distribution, respectively. Input the log-variance vector into the exponential function to obtain the standard deviation. Randomly sample from the standard normal distribution to generate a random noise matrix with the same dimensions as the weight parameters. Multiply the random noise matrix with the standard deviation element by element, and add the product to the mean vector element by element to obtain the weight random variable that follows a normal distribution.

[0087] S322. Calculate the distribution distance between the normal distribution formed by the mean vector and the standard deviation and the standard normal distribution. Substitute the weighted random variables that follow the normal distribution into the graph convolutional network for forward propagation to obtain the predicted probability vector. Perform a dot product operation between the predicted probability vector and the true label vector. Take the logarithm of the operation result and take the negative sign to obtain the cross-entropy loss function value. Use the negative of the cross-entropy loss function value as the log-likelihood probability.

[0088] This invention utilizes an improved EvolveGCN network combined with a Bayesian uncertainty quantification framework to achieve dynamic evolution modeling and accurate risk measurement of user transaction behavior sequences. The node features and adjacency matrix of the dynamic behavior graph are input into the network, and a Long Short-Term Memory (LSTM) network is used to update the graph convolution weights, capturing the patterns of behavioral changes over time. By constructing normally distributed random variables for the weight parameters, variational inference is used to calculate the lower bound of evidence and iteratively optimize the distribution parameters, effectively quantifying the uncertainty of the model's predictions. Monte Carlo sampling generates the expected operation path and confidence interval. By calculating and normalizing the Euclidean distance between the real-time path and the expected path, a behavioral bias value containing uncertainty measurement is output. This invention maintains high robustness in behavioral predictions in complex scenarios with variable transaction behavior patterns, noisy data, or logical gaps, accurately identifies subtle behavioral deviations, and provides confidence level references, significantly improving the model's ability to perceive unknown fraudulent methods and the reliability of risk assessment.

[0089] The improved EvolveGCN network of this invention shares similarities with the original EvolveGCN network in that both retain the core architecture of spatiotemporal graph neural networks. Specifically, they handle the spatial dimension of node topology through graph convolutional layers and utilize recurrent neural networks to capture the dynamic evolutionary features of the temporal dimension. Both treat the weight parameters of the graph convolutional layers as the hidden states of the system, model the temporal changes of the weights using a Long Short-Term Memory (LSTM) network, and employ gating mechanisms such as forget gates and input gates to process the concatenated vector of the weights from the previous time step and the input at the current time step. This outputs weight parameters that dynamically evolve over time, enabling the graph model to adapt to dynamic data.

[0090] The difference lies in that this invention breaks the limitation of the original EvolveGCN network, which only outputs deterministic weight parameters for unidirectional inference, and introduces a Bayesian uncertainty quantification framework to construct a probabilistic weight expression mechanism. While the original model only outputs a deterministic weight matrix through LSTM and directly uses it for graph convolution calculation, this invention, in steps S32 and S321, constructs the dynamically evolving weight parameters as random variables following a normal distribution, establishing two learnable parameter matrices: a mean vector and a log-variance vector, and generating random weight variables through reparameterization techniques. Furthermore, in steps S33 and S322, variational inference is used to calculate the KL divergence and log-likelihood probability, constructing an evidence lower bound (ELBO) as the optimization objective, and iteratively updating the distribution parameters by maximizing the evidence lower bound. In step S34, unlike the original model's single forward propagation, this invention performs Monte Carlo sampling, generating the expected operational path and confidence interval through multiple samplings, thereby outputting a behavioral bias value that includes an uncertainty measure.

[0091] Based on the aforementioned improvements, the beneficial effect of this invention lies in the fact that by introducing a Bayesian framework, the improved EvolveGCN network can quantify the confidence level of the model's prediction results from a probabilistic perspective. This breaks the rigid mode of the original model being unable to perceive its own predictive limitations, achieving a leap from "point prediction" to "interval prediction." This design enables the model to effectively distinguish between random fluctuations and abnormal deviations by using the confidence interval width when facing uncertain scenarios such as drifting user trading habits or data noise interference. This significantly improves the sensitivity to new types of fraudulent behavior and the reliability of risk assessment, enhancing the robustness of the risk control system in complex and ever-changing trading environments.

[0092] In this embodiment, S4 specifically includes:

[0093] S41. Read the behavior deviation value, compare the behavior deviation value containing uncertainty measurement with the preset deviation threshold, and trigger the reverse game verification mechanism when the behavior deviation value is greater than the preset deviation threshold. The preset deviation threshold is the statistical average value of the behavior deviation values ​​of the historical normal user group in the same transaction scenario. The judgment threshold value is obtained by adding twice the standard deviation to the statistical average value.

[0094] S42. Actively send a probing command containing minor interference to the user interface, listen to the user terminal's input interface, collect the user's response to the probing command, and calculate the reaction time and operation trajectory offset of the response action. The probing command containing minor interference is to randomly select a coordinate position in the non-operation area of ​​the user's current transaction interface, generate a semi-transparent mask block with a size of 10 pixels by 10 pixels at the coordinate position, and send the interface display command containing the semi-transparent mask block to the user interface. The reaction time of the response action is calculated as the difference in seconds between the timestamp of the semi-transparent mask block display and the timestamp of the user's next click operation. The operation trajectory offset is calculated by collecting the mouse trajectory, calculating the minimum Euclidean distance from each point on the mouse trajectory to the preset standard movement trajectory, and extracting the maximum value among all Euclidean distances as the operation trajectory offset. The preset standard movement trajectory is a baseline trajectory curve formed by connecting the average coordinates corresponding to each timestamp of the mouse coordinate sequence of historical normal users performing the same operation at the same interface position in the standardized behavior dataset.

[0095] S43. Correct the behavior deviation value based on the response action. The corrected behavior deviation value is obtained by subtracting the reaction time from a preset normal reaction time threshold, dividing the result by the preset normal reaction time threshold to obtain the reaction time deviation rate, subtracting the operation trajectory offset from a preset normal offset threshold, dividing the result by the preset normal offset threshold to obtain the trajectory deviation rate, and calculating the weighted sum of the reaction time deviation rate and the trajectory deviation rate to obtain the correction coefficient. The correction coefficient is multiplied by the behavior deviation value containing uncertainty measurement to obtain the corrected behavior deviation value. The preset normal reaction time threshold is the average reaction time of a historical normal user group when encountering interference. The preset normal offset threshold is the average operation trajectory offset of a historical normal user group. The corrected behavior deviation value is compared with the preset deviation threshold. If the corrected behavior deviation value is greater than or equal to the preset threshold, the risk judgment result is output as high risk. If the corrected behavior deviation value is less than the preset threshold, the risk judgment result is output as low risk.

[0096] This invention introduces a reverse game-theoretic verification mechanism combined with a dynamic interactive probing strategy to accurately identify and eliminate false alarms regarding user transaction risk status. When the behavioral deviation value exceeds a preset deviation threshold based on the statistical average and twice the standard deviation, the reverse game-theoretic verification mechanism is triggered. This mechanism proactively sends a probing command containing a small, semi-transparent masking block to the non-operational area of ​​the user interface, while simultaneously collecting the user's reaction time and operational trajectory offset in response to the interference. A correction coefficient is obtained by calculating the weighted sum of the reaction time deviation rate and the trajectory deviation rate, dynamically correcting behavioral deviation values ​​that include uncertainty metrics. This invention effectively distinguishes between the mechanical, seamless navigation of machine scripts and the natural avoidance behavior of real users, breaking the limitations of traditional single static feature judgment modes. While ensuring the accuracy of high-risk interception, it significantly reduces the false alarm rate caused by accidental user operation drift, enhancing the risk control system's adaptive judgment capability and response accuracy in complex human-computer interaction scenarios.

[0097] In this embodiment, S5 specifically includes:

[0098] S51. Read the risk assessment results one by one, extract the timestamp information of the behavioral data corresponding to each risk assessment result, and arrange the risk assessment results into a time series according to the order of the timestamp information.

[0099] S52. Calculate the time interval between two adjacent risk assessment results in the time series, compare all time interval values ​​with a preset time interval threshold, and classify risk assessment results with time interval values ​​less than the preset time interval threshold into the same time group. The preset time interval threshold is twice the average value of the time interval between adjacent operations of historical normal users.

[0100] S53. Within the same time group, the user operation action types included in the statistical risk judgment results are extracted, and the operation action types that occur more frequently than a preset frequency threshold are combined to form an abnormal behavior pattern. The preset frequency threshold is the average frequency of abnormal actions based on historical risk data.

[0101] S54. Assign a unique numerical code to each type of operation action in the abnormal behavior pattern. Concatenate the numerical codes into a numerical string according to the order in which the operation action types occur. Output the numerical string as the risk feature code. Summarize the risk feature codes of all time groups to form a risk feature code set.

[0102] In this embodiment, S6 specifically includes:

[0103] S61. Read the risk feature code set, traverse each risk feature code in the risk feature code set, and calculate the similarity value between the risk feature code and the historical feature codes stored in the preset code defense rule mapping table. The calculation of the similarity value between the risk feature code and the historical feature codes stored in the preset code defense rule mapping table involves converting the risk feature code and the historical feature code into vector representations of the same dimension, calculating the sum of the products of corresponding elements of the two vectors as the numerator, calculating the product of the magnitudes of the two vectors as the denominator, dividing the numerator by the denominator to obtain the cosine similarity value, and setting the similarity value greater than a preset matching threshold. The historical feature codes corresponding to the defense strategy text are encapsulated as defense rules. The preset code defense rule mapping table is obtained by reading historical transaction risk logs, extracting the historical risk feature codes corresponding to historical risk events with true risk judgment results in the logs, querying the defense operation records of each historical risk event that were ultimately successfully executed without causing financial loss, establishing key-value pair relationships between the historical risk feature codes and the defense strategy text in the defense operation records, and summarizing all key-value pair relationships to generate a mapping dictionary. The preset matching threshold is the minimum average similarity between historical risk feature codes of the same type, with a value range of 0.85 to 0.95.

[0104] S62. Traverse the user nodes in the dynamic behavior graph, calculate the cosine similarity value of the node feature matrix between the current user node and other user nodes, and determine other user nodes whose cosine similarity value is greater than a preset similarity threshold as associated users. The preset similarity threshold is 0.8 times the average value of the cosine similarity between the feature matrix of risk users and normal user nodes corresponding to the historical risk judgment results.

[0105] S63. Find the node identity number of the associated user in the dynamic behavior graph, and send the defense rules to the terminal node where the associated user is located according to the node identity number to perform the defense operation.

[0106] A user online transaction behavior analysis system, comprising:

[0107] The data acquisition and standardization module is used to capture click streams, typing rhythms, mouse trajectories and transaction records during users' online transactions in real time, calculate data disorder indicators in various dimensions, remove invalid records and convert the remaining data into matrix form, and output a standardized behavior dataset.

[0108] The dynamic behavior graph construction module treats each action in the standardized behavior dataset as a node, calculates the density between nodes based on the chronological order of the actions and logical dependencies, connects the nodes into a network structure, and outputs a dynamic behavior graph that reflects the user's transaction habits.

[0109] The uncertainty behavior analysis module is used to input dynamic behavior graphs into the improved EvolveGCN network, introduce a Bayesian uncertainty quantification framework, construct the weight parameters of the graph convolutional layer as a distribution of random variables that evolves dynamically over time, minimize the lower bound of evidence through variational inference to optimize the distribution parameters, generate expected operation paths and confidence intervals based on Monte Carlo sampling, compare the expected operation paths with the user's real-time operation paths, and output a behavior deviation value containing uncertainty measurement.

[0110] The reverse game verification module is used to trigger the reverse game verification mechanism when the behavior deviation value exceeds the preset deviation threshold. It actively sends a probing command containing minor interference to the user interface, collects the user's response to the minor interference, corrects the behavior deviation value based on the response, and outputs the risk judgment result.

[0111] The abnormal pattern coding module is used to classify the risk assessment results according to the time rhythm of occurrence, extract abnormal behavior patterns with similar time interval characteristics, and output a risk feature coding set.

[0112] The associated defense execution module is used to generate corresponding defense rules based on the risk feature encoding set, search for associated users in the transaction network whose operating habits match the risk assessment results, and send the defense rules to the associated user nodes for execution.

[0113] Example 1: To verify the practical application feasibility and technical superiority of this invention in the field of online transaction behavior analysis and risk control, the invention was deployed in the online banking business risk control system of a large domestic joint-stock commercial bank for field operation testing. This banking system covers individual and corporate customers nationwide, with over 8.5 million daily active users and processing 120 million transaction requests daily, involving a massive scale of fund transfers. The transaction scenarios are extremely diverse, covering the entire business process including account login, fund transfers, wealth management purchases, utility bill payments, and loan applications. In such a complex transaction environment with high concurrency and high frequency, traditional risk control methods face enormous challenges. Traditional systems mainly rely on static threshold judgments based on expert rules, such as simple "daily transfer limits," "login blocking from different locations," or "key IP blacklist" matching. While these methods can intercept conventional attacks, they are inadequate when facing increasingly intelligent and covert new types of fraud. For example, attackers use proxy IP pools to simulate normal regional logins, use automated scripts to mimic human operation rhythms for "account farming," and bypass frequency limits through "slow attacks." Traditional methods often only detect anomalies through post-event audits, or lead to a large number of false alarms due to overly strict threshold settings, which seriously affect user experience and cause awkward situations such as normal customer transfers being blocked and the need for manual telephone verification, which increases operating costs and reduces customer satisfaction.

[0114] In this practical application scenario, the method of this invention constructs a comprehensive digital mirror of user behavior by real-time access to raw interaction data from user terminals and backend transaction log data. The system first captures in real-time user clickstream data, keyboard keystrokes and intervals, mouse movement trajectory coordinates, and specific transaction record content during online transactions. Using the aforementioned data cleaning algorithm, it calculates data disorder indices for each dimension, automatically eliminating duplicate records caused by network lag and invalid data from the test environment, generating a high-quality, standardized behavioral dataset. Subsequently, the system treats each atomic operation as a graph node, calculating the density between nodes based on the chronological order of actions and the logical dependencies of page navigation and data input / output, constructing a dynamic behavioral graph reflecting user transaction habits. This process transforms the originally discrete log data into a structured spatiotemporal relational network, enabling the system not only to see "what the user did" but also to understand "the logical connections between operations," thereby effectively identifying risky behaviors with logical breaks or temporal anomalies.

[0115] In the core risk identification stage, this invention inputs a dynamic behavioral graph into an improved EvolveGCN network, introducing a Bayesian uncertainty quantification framework to construct the weight parameters of the graph convolutional layer as a distribution of random variables that dynamically evolves over time. Through this innovative mechanism, the system no longer outputs a rigid "yes / no" judgment, but instead provides a behavioral bias value that includes an uncertainty measure. During testing, the system successfully identified a new type of fraud: fraudsters use scripts to slowly simulate the browsing behavior of normal users in the early morning hours, followed by large-sum transfers. Traditional systems consider this operation rhythm within a threshold range and judge it as normal. However, the method of this invention, through Monte Carlo sampling, found that although the operation path conforms to conventional characteristics, its confidence interval for behavioral bias values ​​is extremely narrow, and the uncertainty measure is extremely low, exhibiting significant "mechanical" characteristics, thus triggering an alert. Subsequently, the system triggers a reverse game-theoretic verification mechanism, proactively sending a probing command containing minor interference to the user interface, i.e., randomly generating a semi-transparent mask block in the non-operational area of ​​the transaction interface. Normal users will naturally avoid it or click cancel, while the script program exhibits the abnormal characteristic of completely ignoring interference and traversing the mask block in a straight line. After the system detects this response, it quickly corrects the behavioral deviation value, determines the risk level to be high, and immediately executes a blocking operation.

[0116] After a three-month trial run, the method of this invention has achieved a qualitative leap in key indicators such as risk identification accuracy, response speed, and false alarm rate control. In particular, it demonstrates strong adaptability and robustness in dealing with unknown types of fraud attacks. Furthermore, the system automatically evolves its defense rule base by extracting risk feature codes generated from abnormal behavior patterns, achieving joint prevention and control of related risks.

[0117] Table 1 below details the comparison data of the core performance indicators of the method of the present invention and the bank's original traditional risk control system in typical online transaction risk scenarios during the three-month operation period:

[0118] Table 1. Performance Comparison Data Between the Invention and Traditional Risk Control Systems in Typical Online Transaction Risk Scenarios

[0119] Based on the comparative data shown in Table 1, it can be seen that the user network transaction behavior analysis method proposed in this invention has significant performance advantages over traditional rule systems in the identification of complex transaction risks, especially in key indicators such as early warning accuracy, response speed, risk level matching degree, and control of false alarms and false alarms.

[0120] Regarding early warning accuracy, this invention maintains a high level of over 93% in all five typical online transaction risk scenarios, far exceeding the average accuracy of traditional systems. For example, in the "machine script attack" scenario, traditional systems rely on simple frequency thresholds and static feature matching, which almost fail when faced with intelligent scripts simulating normal operating rhythms, achieving only 44.9% accuracy. In contrast, this invention, through an improved EvolveGCN network to capture the temporal dependency features of operating paths and introducing a Bayesian uncertainty quantification framework to perceive the "mechanized" low uncertainty features of script behavior, achieves an accuracy of 97.6%, effectively improving the ability to identify highly concealed automated attacks. Similarly, in the "new vulnerability attack" scenario, the accuracy of traditional methods is only 30.0%, while this invention reaches 93.6%, demonstrating the method's strong generalization ability against unknown threats.

[0121] Regarding response timeliness, this invention significantly shortens the average response time through the efficient computation of graph neural networks and the rapid intervention of a reverse game verification mechanism. In various behavioral scenarios, the average response time is drastically reduced from over 450 milliseconds in traditional methods to approximately 130 milliseconds, a nearly fourfold increase in speed. For example, the traditional response to "new vulnerability attacks" typically takes 800 milliseconds, often missing the opportunity to intercept, while this invention only requires 130 milliseconds, ensuring that risk assessment and blocking are completed instantly before the user's transaction ends, greatly enhancing the security of bank funds.

[0122] Regarding risk level matching accuracy, this invention maintains a accuracy rate above 88% across all scenarios, while traditional methods typically hover around 60%. Traditional methods often use fixed thresholds to classify risk levels, which can easily lead to normal large transfers being misjudged as high-risk, or fraudulent activities involving phased transfers being underestimated. This invention, however, utilizes the evolutionary trend comparison of dynamic behavioral graphs to make the risk level assessment more refined and realistic, ensuring the accuracy and differentiation of subsequent response strategies. For example, in the "account theft and fraudulent use" scenario, the matching accuracy increased from 62.3% to 92.5%.

[0123] Significant advantages are also demonstrated in controlling false positive and false negative rates. The average false positive rate of this invention is controlled below 5%, and the false negative rate is around 2%, significantly reducing redundant warnings and omissions compared to traditional methods, which have a maximum false positive rate of 28.5% and a false negative rate of 35.0%. Particularly in the case of "machine script attacks," the false negative rate of traditional methods reaches as high as 35.2%, while that of this invention is only 0.5%, effectively reducing the cost of manual intervention and improving system stability. Overall, this invention achieves efficient, accurate, and intelligent risk identification by constructing a dynamic behavior graph and combining uncertainty measurement and reverse game theory verification, possessing extremely high practical application value.

[0124] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for analyzing user online transaction behavior, characterized in that, Includes the following steps: S1. Capture the click flow, typing rhythm, mouse trajectory and transaction records of users in real time during the online transaction process, calculate the data disorder index of each dimension, remove invalid records and convert the remaining data into matrix form, and output a standardized behavior dataset. S2. Treat each action in the standardized behavior dataset as a node, calculate the tightness between nodes according to the order of action occurrence and logical dependencies, connect the nodes into a network structure, and output a dynamic behavior graph. S3. Input the dynamic behavior graph into the improved EvolveGCN network, introduce the Bayesian uncertainty quantification framework, construct the weight parameters of the graph convolutional layer as a distribution of random variables that evolves dynamically over time, minimize the lower bound of evidence through variational inference to optimize the distribution parameters, generate the expected operation path and confidence interval, compare the expected operation path with the user's real-time operation path, and output the behavior deviation value. S4. When the behavior deviation value exceeds the preset deviation threshold, the reverse game verification mechanism is triggered. The mechanism actively sends a probing command to the user interface, collects the user's response to minor disturbances, corrects the behavior deviation value based on the response, and outputs the risk assessment result. S5. Classify the risk assessment results according to the time rhythm of occurrence, extract abnormal behavior patterns with similar time interval characteristics, and output a risk feature code set. S6. Generate corresponding defense rules based on the risk feature code set, search for related users in the transaction network whose operating habits match the risk assessment results, and send the defense rules to the related user nodes for execution.

2. The method for analyzing user online transaction behavior according to claim 1, characterized in that, S1 specifically includes: S11. Listen to the input interface of the user terminal, capture the coordinate sequence of mouse movement, the timestamp of click operation, the key code and interval of keyboard input according to the set time window, and read the transaction records in the transaction backend database. S12. Calculate the spatial location dispersion using the geometric center position of the coordinate sequence, calculate the information entropy of the click timestamp as the temporal disorder, and calculate the variance of the keyboard interval duration as the rhythm fluctuation. The spatial location dispersion, temporal disorder, and rhythm fluctuation are collectively referred to as the data disorder index of each dimension. S13. Compare the data disorder index of each dimension with the preset normal threshold range, determine the records that exceed the normal threshold range as invalid records and delete them, and retain the remaining data that are within the normal threshold range. S14. Perform numerical mapping encoding on the character action labels in the remaining data, normalize and scale the timestamp data, concatenate the processed data into a matrix by rows and columns, and output the standardized behavior dataset.

3. The method for analyzing user online transaction behavior according to claim 1, characterized in that, S2 specifically includes: S21. Read the records in the standardized behavior dataset line by line, define each mouse click event, keyboard key event and transaction confirmation event as an action node, and assign a unique identity number to each action node. S22. According to the chronological order of the actions, select two adjacent action nodes in turn, calculate the absolute value of the timestamp difference between the two action nodes, and determine the degree of closeness in the time dimension by the reciprocal of the absolute value of the timestamp difference. S23. Identify the page jump relationship and data input / output relationship between the current action node and other action nodes. If there is a page jump or data transfer, it is determined that there is a logical dependency. Set the tightness between the logically dependent node pairs to a preset logical association strength value. S24. The tightness in the time dimension and the logical association strength value are weighted and summed. The result is used as the tightness between nodes. If there is no logical dependency between node pairs, only the tightness in the time dimension is used. Node pairs with tightness greater than zero are connected by edges. S25. Summarize all action nodes and connecting edges to construct a mesh structure, using the identity number of the action node as the vertex and the density as the edge weight, and output a dynamic behavior graph.

4. The method for analyzing user online transaction behavior according to claim 1, characterized in that, S3 specifically includes: S31. Read the node feature matrix and adjacency matrix in the dynamic behavior graph, input the node feature matrix into the improved EvolveGCN network, introduce a long short-term memory network in the graph convolutional layer, concatenate the weight matrix of the previous time step with the input of the current time step, calculate the output of the forget gate and the input gate through the long short-term memory network, calculate the candidate hidden state at the current time step, update the hidden state of the weight matrix, and output the weight parameters that evolve dynamically over time. S32. Introduce a Bayesian uncertainty quantification framework, construct the weight parameters that evolve dynamically over time as random variables that follow a normal distribution, calculate the KL divergence and the log-likelihood probability of the prediction results, and output the lower bound of evidence. S33. Maximize the lower bound of evidence using the stochastic gradient descent algorithm. Iteratively update the mean and variance parameters of the normal distribution through backpropagation. Stop iterating when the growth rate of the lower bound of evidence is less than the preset convergence threshold, and output the optimized distribution parameters. S34. Perform a preset number of Monte Carlo samplings on the optimized distribution parameters. In each sampling, a set of weight values ​​are randomly selected from the normal distribution and assigned to the graph convolutional layer. The adjacency matrix is ​​used to perform forward propagation to calculate the output operation probability. The mean of all sampling results is used as the expected operation path, and the variance range of all sampling results is used as the confidence interval. S35. Read the user's real-time operation path, calculate the Euclidean distance between the expected operation path and the user's real-time operation path, normalize the Euclidean distance by dividing it by the half-width of the confidence interval, and output the behavior deviation value.

5. The method for analyzing user online transaction behavior according to claim 4, characterized in that, Specifically, S32 includes: S321. Introduce the Bayesian uncertainty quantification framework, establish two learnable parameter matrices with the same dimensions as the weight parameters as the mean vector and log-variance vector of the weight parameter distribution, respectively. Input the log-variance vector into the exponential function to obtain the standard deviation. Randomly sample from the standard normal distribution to generate a random noise matrix with the same dimensions as the weight parameters. Multiply the random noise matrix with the standard deviation element by element, and add the product to the mean vector element by element to obtain the weight random variable that follows a normal distribution. S322. Calculate the distribution distance between the normal distribution formed by the mean vector and the standard deviation and the standard normal distribution. Substitute the weighted random variables that follow the normal distribution into the graph convolutional network for forward propagation to obtain the predicted probability vector. Perform a dot product operation between the predicted probability vector and the true label vector. Take the logarithm of the operation result and take the negative sign to obtain the cross-entropy loss function value. Use the negative of the cross-entropy loss function value as the log-likelihood probability.

6. The method for analyzing user online transaction behavior according to claim 1, characterized in that, S4 specifically includes: S41. Read the behavior deviation value, compare the behavior deviation value containing uncertainty measurement with the preset deviation threshold, and trigger the reverse game verification mechanism when the behavior deviation value is greater than the preset deviation threshold. S42. Actively send probing commands to the user interface, listen to the user terminal's input interface, collect the user's response to the probing commands, and calculate the response time and operation trajectory offset. S43. Correct the behavior deviation value according to the response action, compare the corrected behavior deviation value with the preset deviation threshold. If the corrected behavior deviation value is greater than or equal to the preset threshold, output the risk judgment result as high risk. If the corrected behavior deviation value is less than the preset threshold, output the risk judgment result as low risk.

7. The method for analyzing user online transaction behavior according to claim 1, characterized in that, S5 specifically includes: S51. Read the risk assessment results one by one, extract the timestamp information of the behavioral data corresponding to each risk assessment result, and arrange the risk assessment results into a time series according to the order of the timestamp information. S52. Calculate the time interval between two adjacent risk assessment results in the time series, compare all time interval values ​​with the preset time interval threshold, and classify the risk assessment results with time interval values ​​less than the preset time interval threshold into the same time group. S53. Within the same time group, extract the user operation action types included in the statistical risk judgment results, and combine them to form an abnormal behavior pattern. S54. Assign a unique numerical code to each type of operation action in the abnormal behavior pattern. Concatenate the numerical codes into a numerical string according to the order in which the operation action types occur. Output the numerical string as the risk feature code. Summarize the risk feature codes of all time groups to form a risk feature code set.

8. The method for analyzing user online transaction behavior according to claim 1, characterized in that, S6 specifically includes: S61. Read the risk feature code set, traverse each risk feature code in the risk feature code set, calculate the similarity value between the risk feature code and the historical feature code stored in the preset code defense rule mapping table, and encapsulate the defense strategy text corresponding to the historical feature code with a similarity value greater than the preset matching threshold into a defense rule. S62. Traverse the user nodes in the dynamic behavior graph, calculate the cosine similarity value of the node feature matrix between the current user node and other user nodes, and determine other user nodes whose cosine similarity value is greater than the preset similarity threshold as related users. S63. Find the node identity number of the associated user in the dynamic behavior graph, and send the defense rules to the terminal node where the associated user is located according to the node identity number to perform the defense operation.

9. A user online transaction behavior analysis system, executing the user online transaction behavior analysis method according to any one of claims 1 to 8, characterized in that, include: The data acquisition and standardization module is used to capture click streams, typing rhythms, mouse trajectories and transaction records during users' online transactions in real time, calculate data disorder indicators in various dimensions, remove invalid records and convert the remaining data into matrix form, and output a standardized behavior dataset. The dynamic behavior graph construction module is used to treat each action in the standardized behavior dataset as a node, calculate the tightness between nodes according to the chronological order of the actions and logical dependencies, connect the nodes into a network structure, and output a dynamic behavior graph. The uncertainty behavior analysis module is used to input dynamic behavior graphs into the improved EvolveGCN network, introduce a Bayesian uncertainty quantification framework, construct the weight parameters of the graph convolutional layer as a distribution of random variables that evolves dynamically over time, minimize the lower bound of evidence through variational inference to optimize the distribution parameters, generate expected operation paths and confidence intervals, compare the expected operation paths with the user's real-time operation paths, and output the behavior deviation value. The reverse game verification module is used to trigger the reverse game verification mechanism when the behavior deviation value exceeds the preset deviation threshold. It actively sends a probing command to the user interface, collects the user's response to minor disturbances, corrects the behavior deviation value based on the response, and outputs the risk assessment result. The abnormal pattern coding module is used to classify the risk assessment results according to the time rhythm of occurrence, extract abnormal behavior patterns with similar time interval characteristics, and output a risk feature coding set. The associated defense execution module is used to generate corresponding defense rules based on the risk feature encoding set, search for associated users in the transaction network whose operating habits match the risk assessment results, and send the defense rules to the associated user nodes for execution.