Intelligent Monitoring Method and System Based on Online Marking Behavior

Abstract

The present invention relates to the technical field of data processing, and discloses an intelligent monitoring method and system based on online marking behavior, which is used to improve the accuracy of intelligent monitoring based on online marking behavior. The method includes: collecting the behavior data of the marking teachers to obtain marking behavior data; inputting the marking behavior data into a graph neural network for behavior feature extraction to obtain a behavior feature set; inputting the behavior feature set into a variational autoencoder for behavior pattern recognition to obtain a behavior pattern set; inputting the behavior pattern set into a contrastive learning algorithm for abnormal behavior recognition to obtain abnormal behavior data, where the abnormal behavior data includes: rapid scoring behavior, abnormal scoring distribution data, and abnormal operation frequency; analyzing the abnormal behavior data to obtain a marking behavior monitoring report, and transmitting the marking behavior monitoring report to a preset data display terminal.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an intelligent monitoring method and system based on online marking behavior. Background Art

[0002] In the prior art, online marking systems have been widely used in various examinations and evaluations. These systems can improve the efficiency and fairness of marking to a certain extent, and reduce the workload and subjective deviation of manual marking. However, there are still many deficiencies in the existing online marking systems in monitoring the behavior of marking teachers, and it is difficult to effectively identify and manage bad behaviors in the marking process.

[0003] The prior art mainly relies on manual monitoring and post-event review, but this method has the following deficiencies: First, manual monitoring requires a large amount of manpower, and omissions and oversights are prone to occur during the monitoring process; Second, post-event review lacks real-time performance and cannot detect and correct wrong behaviors in the marking process in a timely manner, resulting in the fairness and accuracy of the marking results being affected. In addition, the existing automated monitoring means are mostly based on simple rule and threshold settings, and it is difficult to capture complex behavior patterns and potential anomalies. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide an intelligent monitoring method and system based on online marking behavior, which are used to improve the accuracy of intelligent monitoring based on online marking behavior.

[0005] The present invention provides an intelligent monitoring method based on online marking behavior, including: collecting the behavior data of marking teachers to obtain marking behavior data, where the marking behavior data includes: the number of mouse clicks, the mouse click position, the page stay time, and the scoring submission time;

[0006] Inputting the marking behavior data into a graph neural network for behavior feature extraction to obtain a behavior feature set;

[0007] Inputting the behavior feature set into a variational autoencoder for behavior pattern recognition to obtain a behavior pattern set;

[0008] Inputting the behavior pattern set into a contrastive learning algorithm for abnormal behavior recognition to obtain abnormal behavior data, where the abnormal behavior data includes: fast scoring behavior, abnormal scoring distribution data, and abnormal operation frequency;

[0009] Performing monitoring report analysis on the abnormal behavior data to obtain a marking behavior monitoring report, and transmitting the marking behavior monitoring report to a preset data display terminal.

[0010] In the present invention, the step of inputting the marking behavior data into a graph neural network for behavior feature extraction to obtain a behavior feature set includes:

[0011] Generating a plurality of behavior event nodes according to the number of mouse clicks, the mouse click positions, the page residence time, and the scoring submission time;

[0012] Performing adjacent node extraction on the plurality of behavior event nodes to obtain multiple groups of adjacent nodes;

[0013] Calculating the interval time for each group of adjacent nodes respectively to obtain the interval time of each group of adjacent nodes;

[0014] Calculating the spatial distance for each group of adjacent nodes respectively to obtain the spatial distance of each group of adjacent nodes;

[0015] Performing edge weight analysis on each group of adjacent nodes respectively according to the interval time of each group of adjacent nodes and the spatial distance of each group of adjacent nodes to obtain the edge weight data of each group of adjacent nodes;

[0016] Constructing the edge data of each group of adjacent nodes based on the edge weight data of each group of adjacent nodes, and performing graph structure construction according to the multiple groups of adjacent nodes and the edge data of each group of adjacent nodes to obtain a target graph structure;

[0017] Inputting the target graph structure into the graph neural network for node behavior feature extraction to obtain the behavior feature set.

[0018] In the present invention, the graph neural network includes: an input layer, a message passing layer, and a feature extraction layer. The step of inputting the target graph structure into the graph neural network for node behavior feature extraction to obtain the behavior feature set includes:

[0019] Inputting the target graph structure into the input layer for behavior data vector conversion to obtain a behavior data vector;

[0020] Inputting the behavior data vector into the message passing layer for neighbor node information aggregation to obtain the neighbor node information of each behavior event node;

[0021] Updating the node data of each behavior event node according to the neighbor node information of each behavior event node to obtain each updated behavior event node;

[0022] Inputting each updated behavior event node into the feature extraction layer for node behavior feature extraction to obtain the behavior feature set.

[0023] In the present invention, the variational autoencoder includes an encoder, a latent space layer, and a decoder. The step of inputting the behavior feature set into the variational autoencoder for behavior pattern recognition to obtain a behavior pattern set includes:

[0024] Input the behavior feature set into the encoder for latent variable mapping to obtain the latent variable data corresponding to the behavior feature set, where the latent variable data includes: mean data and standard deviation data;

[0025] Input the latent variable data into the latent space layer for latent variable sampling to obtain sampled latent variable data;

[0026] Input the latent variable data into the decoder for data reconstruction mapping to obtain reconstructed data;

[0027] Perform behavior pattern recognition on the reconstructed data to obtain the behavior pattern set.

[0028] In the present invention, before the step of inputting the behavior pattern set into the contrastive learning algorithm for abnormal behavior recognition to obtain abnormal behavior data, where the abnormal behavior data includes: rapid scoring behavior, abnormal scoring distribution data, and abnormal operation frequency, the following steps are also included:

[0029] Collect a historical behavior pattern set, preprocess the historical behavior pattern set to obtain positive sample pairs and negative sample pairs;

[0030] Input the positive sample pairs and the negative sample pairs into the initial contrastive learning algorithm for feature extraction to obtain the first feature vector corresponding to the positive sample pairs and the second feature vector corresponding to the negative sample pairs;

[0031] Calculate the loss value of the initial contrastive learning algorithm based on the first feature vector and the second feature vector to obtain the target loss value;

[0032] When the target loss value meets the preset threshold, obtain the contrastive learning algorithm.

[0033] In the present invention, the step of performing monitoring report analysis on the abnormal behavior data to obtain a marking behavior monitoring report and transmitting the marking behavior monitoring report to a preset data display terminal includes:

[0034] Classify the abnormal behavior data to obtain abnormal data of multiple abnormal types;

[0035] Perform statistical feature analysis on the abnormal data of each abnormal type to obtain the statistical features of the abnormal data of each abnormal type;

[0036] Extract the behavior details of the abnormal data for each abnormal type to obtain the detailed data of the abnormal data for each abnormal type;

[0037] Analyze the monitoring report based on the statistical characteristics of the abnormal data for each abnormal type and the detailed data of the abnormal data for each abnormal type to obtain a marking behavior monitoring report, and transmit the marking behavior monitoring report to a preset data display terminal.

[0038] In the present invention, the step of analyzing the monitoring report based on the statistical characteristics of the abnormal data for each abnormal type and the detailed data of the abnormal data for each abnormal type to obtain a marking behavior monitoring report, and transmitting the marking behavior monitoring report to a preset data display terminal includes:

[0039] Generate a statistical chart based on the statistical characteristics of the abnormal data for each abnormal type and the detailed data of the abnormal data for each abnormal type;

[0040] Construct data display elements according to the statistical icon and match the rendering parameters of the data display elements;

[0041] Construct a marking behavior monitoring report according to the rendering parameters of the data display elements and the data display elements, and transmit the marking behavior monitoring report to the data display terminal.

[0042] The present invention also provides an intelligent monitoring system based on online marking behavior, including:

[0043] A collection module for collecting the behavior data of marking teachers to obtain marking behavior data, where the marking behavior data includes: the number of mouse clicks, the mouse click position, the page stay time, and the score submission time;

[0044] An extraction module for inputting the marking behavior data into a graph neural network for behavior feature extraction to obtain a behavior feature set;

[0045] An identification module for inputting the behavior feature set into a variational autoencoder for behavior pattern recognition to obtain a behavior pattern set;

[0046] An input module for inputting the behavior pattern set into a contrastive learning algorithm for abnormal behavior recognition to obtain abnormal behavior data, where the abnormal behavior data includes: rapid scoring behavior, abnormal scoring distribution data, and abnormal operation frequency;

[0047] An analysis module for analyzing the monitoring report of the abnormal behavior data to obtain a marking behavior monitoring report, and transmitting the marking behavior monitoring report to a preset data display terminal.

[0048] In the technical solution provided by the present invention, by collecting the behavior data of the marking teachers, including the number of mouse clicks, the mouse click positions, the page residence time, and the scoring submission time, it is possible to comprehensively and meticulously record the operation behaviors of the teachers, providing a rich data basis for subsequent analysis and recognition. These behavior data are transmitted to the data processing center in real time through the data collection module, realizing the timely acquisition and processing of the data, ensuring the real-time performance and accuracy of the monitoring. Secondly, in the solution, a graph neural network is used for behavior feature extraction. By constructing a spatio-temporal graph from the behavior data, taking the operation behaviors at different time points as nodes and the associations between the behaviors as edges, and performing weighted analysis on these nodes and edges, it is possible to effectively capture the spatio-temporal dependence relationships and implicit patterns in the behavior data. This feature extraction method is richer and more accurate than the traditional planar feature extraction method, and can better reflect the operation habits and behavior patterns of the teachers. Further, in the solution, a variational autoencoder is used for behavior pattern recognition. By mapping the high-dimensional behavior feature vectors to a low-dimensional latent space and learning the behavior patterns by maximizing the posterior distribution, it is possible to distinguish normal and abnormal marking behaviors in the latent space. The variational autoencoder can not only effectively reduce the dimension, but also generate new behavior samples, enhancing the robustness and generalization ability of the model, and providing a high-quality set of behavior patterns for subsequent anomaly detection. Next, in the solution, a contrastive learning algorithm is used for abnormal behavior recognition. By constructing positive and negative sample pairs and training the feature encoder, similar behavior features are made closer in the feature space and different behavior features are made farther apart, realizing the accurate recognition of abnormal behaviors. Compared with the traditional supervised learning method, the contrastive learning algorithm not only improves the detection accuracy of the model, but also can effectively train the model using unlabeled data in the absence of a large amount of labeled data. The solution also includes the analysis of the monitoring reports for the abnormal behavior data. By classifying and statistically analyzing the abnormal behavior data, a detailed monitoring report is generated and transmitted to the preset data display terminal. In this way, it is possible to timely provide detailed abnormal behavior reports to the management personnel and the marking teachers, helping them to timely discover and handle abnormal behaviors, improving the transparency of the marking process and the credibility of the scoring results. In addition, the monitoring report also includes specific improvement suggestions and measures to help the teachers improve their operation behaviors and enhance the marking quality. Overall, through the comprehensive application of the graph neural network, the variational autoencoder, and the contrastive learning algorithm, this solution constructs a real-time and intelligent marking behavior monitoring system, which can not only comprehensively and accurately collect and analyze the operation behaviors of the teachers, but also timely and accurately identify and report abnormal behaviors, having high practical value and creativity, effectively solving the deficiencies in the existing technologies, and providing a new technical means for the intelligent monitoring of the online marking system.Especially in the aspects of statistical chart generation and rendering of data display elements, by analyzing the statistical characteristics and detailed data of abnormal data for each abnormal type, corresponding statistical charts are generated, such as the time interval distribution chart and the score deviation bar chart, which can intuitively display the distribution and characteristics of abnormal behaviors, helping managers and teachers better understand and handle abnormal behaviors. These statistical charts ensure the display effect and user experience of the charts by selecting appropriate data display elements and matching corresponding rendering parameters. After generating these display elements, the system can embed them into the marking behavior monitoring report to construct a comprehensive monitoring report document, including an overview of abnormal behaviors, statistical charts, detailed analysis, and improvement suggestions, providing comprehensive reference information for managers and teachers to help them take improvement measures in a timely manner and enhance the transparency of the marking process and the credibility of the grading results. Finally, the monitoring report is securely transmitted to the pre-set data display terminal through the network transmission protocol, ensuring that managers and teachers can conveniently view and process the report content, further enhancing the practicality and effectiveness of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0050] Figure 1 It is a flowchart of an intelligent monitoring method based on online marking behavior in an embodiment of the present invention.

[0051] Figure 2 It is a schematic diagram of an intelligent monitoring system based on online marking behavior in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the drawings. Obviously, the described embodiments are some, but not all, embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0053] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be construed as a limitation to the present invention. In addition, the terms "first", "second", "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0054] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0055] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to Figure 1 , Figure 1 which is a flowchart of an intelligent monitoring method based on online marking behavior according to an embodiment of the present invention. As Figure 1 shown, it includes the following steps:

[0056] S101. Collect the behavior data of the marking teacher to obtain marking behavior data, where the marking behavior data includes: the number of mouse clicks, the mouse click position, the page stay time, and the score submission time;

[0057] S102. Input the marking behavior data into a graph neural network for behavior feature extraction to obtain a behavior feature set;

[0058] S103. Input the behavior feature set into a variational autoencoder for behavior pattern recognition to obtain a behavior pattern set;

[0059] S104. Input the behavior pattern set into a contrastive learning algorithm for abnormal behavior recognition to obtain abnormal behavior data, where the abnormal behavior data includes: rapid scoring behavior, abnormal scoring distribution data, and abnormal operation frequency;

[0060] S105. Analyze the abnormal behavior data to obtain a marking behavior monitoring report, and transmit the marking behavior monitoring report to a preset data display terminal.

[0061] It should be noted that when the teacher starts grading papers, all their operation behaviors will be recorded in real time. These behavior data include the number of mouse clicks, the mouse click positions, the page stay time, and the grading submission time. For example, when the teacher clicks the next page button, the time and position of this click operation will be recorded; when the teacher stays on a certain page for more than a certain time, the stay time will be recorded; when the teacher grades a certain question, the specific time of grading submission will be recorded. After these data are captured by the acquisition module, they are transmitted to the data processing center in real time. The captured grading behavior data will be input into a graph neural network for behavior feature extraction. A graph neural network is a deep learning model that can process graph-structured data and can capture complex relationships and patterns in the data. Here, the behavior data is constructed into a spatio-temporal graph, where the nodes represent operation behaviors at different times, and the edges represent the associations between these operation behaviors. For example, two consecutive mouse click operations can be used as two nodes, and the edge weight between these two nodes can be determined by the time interval and spatial distance between them. Through the message passing mechanism, the graph neural network extracts behavior features from these nodes and edges to form a behavior feature set. This process enables the capture of the teacher's operation habits and behavior patterns at different time points, providing rich and useful feature representations. Then, the extracted behavior feature set is input into a variational autoencoder (VAE) for behavior pattern recognition. A variational autoencoder is a generative model that maps high-dimensional data to a low-dimensional latent space through an encoder and then reconstructs the original data through a decoder. The advantage of VAE is that it can effectively reduce the dimension and generate new samples to enhance the robustness of the model. In this process, the encoder converts the behavior feature set into latent variables, which capture the main features and structures of the data. By maximizing the posterior distribution, VAE learns the latent distribution of the behavior features and reconstructs the input data through the decoder. In the latent space, normal and abnormal grading behaviors will show obvious distribution differences, forming a behavior pattern set. Next, the behavior pattern set is input into a contrastive learning algorithm for abnormal behavior recognition. Contrastive learning is a self-supervised learning method that makes similar samples closer and different samples farther apart in the feature space by constructing positive sample pairs and negative sample pairs. First, a historical behavior pattern set is collected and preprocessed to obtain positive sample pairs and negative sample pairs. The positive sample pairs can be the normal operation behaviors of the same teacher at different time periods, and the negative sample pairs are the comparisons between normal behaviors and abnormal behaviors. These sample pairs are input into the contrastive learning algorithm for training, and the trained feature encoder can effectively distinguish normal and abnormal behavior patterns. Then, the new behavior pattern set is input into the feature encoder. If the position of the encoded feature vector in the feature space is beyond a set threshold compared with the normal behavior features, then this behavior is marked as abnormal. In this way, abnormal behaviors such as rapid grading behaviors, abnormal grading distribution data, and abnormal operation frequencies can be identified. Finally, the identified abnormal behavior data is subjected to monitoring report analysis.Classify and statistically analyze the abnormal behavior data to generate a detailed monitoring report. For example, the number of rapid grading behaviors of a certain teacher within a specific time period can be counted, and a graph can be drawn to show the abnormalities in the grading distribution. At the same time, the specific details of each abnormal behavior, such as the occurrence time, operation details, etc., will be recorded, and improvement suggestions will be generated. These reports are transmitted via the network to a pre-set data display terminal for managers and teachers to view and process. The data display terminal provides a friendly user interface, and users can view detailed abnormal behavior information and statistical results through the interactive interface, helping them discover and handle abnormal behaviors in a timely manner, improving the transparency of the marking process and the credibility of the grading results. For example, when a teacher discovers during the marking process that they have submitted grades rapidly in a short period of time, this behavior may indicate that the teacher is marking carelessly. The operation behavior characteristics of the teacher are extracted through a graph neural network, and this abnormal pattern is identified through a VAE. Subsequently, a contrastive learning algorithm compares this abnormal behavior with normal behaviors and confirms it as a rapid grading behavior. A monitoring report is generated, recording the detailed information of this abnormal behavior, including the number of rapid grading times, specific operation times, etc., and an alarm and report are sent to the data display terminal to remind relevant managers to intervene and handle, thus ensuring the fairness and accuracy of grading.

[0062] By performing the above steps, by collecting the behavior data of the marking teachers, including the number of mouse clicks, the mouse click positions, the page stay time, and the scoring submission time, it is possible to comprehensively and meticulously record the operation behaviors of the teachers, providing a rich data basis for subsequent analysis and recognition. These behavior data are transmitted to the data processing center in real time through the data collection module, realizing the timely acquisition and processing of the data, ensuring the real-time nature and accuracy of the monitoring. Secondly, the graph neural network is adopted in the solution for behavior feature extraction. By constructing a spatio-temporal graph from the behavior data, taking the operation behaviors at different time points as nodes and the associations between the behaviors as edges, and performing weighted analysis on these nodes and edges, it is possible to effectively capture the spatio-temporal dependence relationships and implicit patterns in the behavior data. This feature extraction method is richer and more accurate than the traditional planar feature extraction method and can better reflect the operation habits and behavior patterns of the teachers. Further, the variational autoencoder is used in the solution for behavior pattern recognition. By mapping the high-dimensional behavior feature vectors to a low-dimensional latent space and learning the behavior patterns by maximizing the posterior distribution, it is possible to distinguish normal and abnormal marking behaviors in the latent space. The variational autoencoder can not only effectively reduce the dimension but also generate new behavior samples, enhancing the robustness and generalization ability of the model and providing a high-quality set of behavior patterns for subsequent anomaly detection. Next, the contrastive learning algorithm is adopted in the solution for abnormal behavior recognition. By constructing positive sample pairs and negative sample pairs and training the feature encoder, similar behavior features are made closer in the feature space and different behavior features are made farther apart, realizing the accurate recognition of abnormal behaviors. Compared with the traditional supervised learning method, the contrastive learning algorithm not only improves the detection accuracy of the model but also can effectively train the model using unlabeled data in the absence of a large amount of labeled data. The solution also includes the analysis of the monitoring reports for the abnormal behavior data. By classifying and statistically analyzing the abnormal behavior data, a detailed monitoring report is generated and transmitted to the preset data display terminal. In this way, it is possible to timely provide detailed abnormal behavior reports to the management personnel and the marking teachers, helping them discover and handle abnormal behaviors in a timely manner, improving the transparency of the marking process and the credibility of the scoring results. In addition, the monitoring report also includes specific improvement suggestions and measures to help teachers improve their operation behaviors and enhance the marking quality. Overall, this solution constructs a real-time and intelligent marking behavior monitoring system by comprehensively applying the graph neural network, the variational autoencoder, and the contrastive learning algorithm. It can not only comprehensively and accurately collect and analyze the operation behaviors of the teachers but also timely and precisely identify and report abnormal behaviors. It has high practical value and creativity, effectively solves the deficiencies in the existing technologies, and provides a new technical means for the intelligent monitoring of the online marking system.Especially in the aspects of statistical chart generation and the rendering of data display elements, by analyzing the statistical characteristics and detailed data of abnormal data for each abnormal type, corresponding statistical charts are generated, such as time interval distribution charts and score deviation bar charts, which can intuitively display the distribution and characteristics of abnormal behaviors, helping managers and teachers better understand and handle abnormal behaviors. These statistical charts ensure the display effect and user experience of the charts by selecting appropriate data display elements and matching corresponding rendering parameters. After generating these display elements, the system can embed them into the marking behavior monitoring report to construct a comprehensive monitoring report document, including an overview of abnormal behaviors, statistical charts, detailed analysis, and improvement suggestions, providing comprehensive reference information for managers and teachers to help them take improvement measures in a timely manner and enhance the transparency of the marking process and the credibility of the scoring results. Finally, the monitoring report is securely transmitted to the pre-set data display terminal through the network transmission protocol, ensuring that managers and teachers can conveniently view and process the report content, further enhancing the practicality and effectiveness of the system.

[0063] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0064] (1) Generate multiple behavior event nodes based on the number of mouse clicks, mouse click positions, page stay times, and score submission times;

[0065] (2) Extract adjacent nodes from the multiple behavior event nodes to obtain multiple groups of adjacent nodes;

[0066] (3) Calculate the interval time for each group of adjacent nodes respectively to obtain the interval time of each group of adjacent nodes;

[0067] (4) Calculate the spatial distance for each group of adjacent nodes respectively to obtain the spatial distance of each group of adjacent nodes;

[0068] (5) Based on the interval time of each group of adjacent nodes and the spatial distance of each group of adjacent nodes, perform edge weight analysis on each group of adjacent nodes respectively to obtain the edge weight data of each group of adjacent nodes;

[0069] (6) Construct the edge data of each group of adjacent nodes based on the edge weight data of each group of adjacent nodes, and construct a graph structure according to the multiple groups of adjacent nodes and the edge data of each group of adjacent nodes to obtain the target graph structure;

[0070] (7) Input the target graph structure into a graph neural network for node behavior feature extraction to obtain a behavior feature set.

[0071] Specifically, when the teacher clicks on an option during the marking process, record the time and click position of this operation and take it as a behavioral event node. Then, when the teacher stays on a certain page for more than a certain period of time, record this stay time and generate a corresponding node. The marking submission time is also a key data point. Each time the marking is submitted, record the specific submission time and generate a node. After generating multiple behavioral event nodes, the next step is to extract adjacent nodes from these nodes. The purpose of adjacent node extraction is to pair nodes that are close in time to form multiple groups of adjacent nodes. For example, if the teacher clicks on multiple options continuously within a short period of time, the nodes of these click operations will be extracted as a group of adjacent nodes. Similarly, the stay times of the teacher between different pages can also form adjacent nodes. Then, for each group of adjacent nodes, it is necessary to calculate the time interval between them. The time interval refers to the time difference between two adjacent nodes. For example, if node A and node B represent the teacher's click operations at two different time points, calculate the time difference between the two adjacent nodes. This time interval reflects the time pattern of the teacher's operations and helps to identify abnormal behaviors. Next, it is necessary to calculate the spatial distance between each group of adjacent nodes. The spatial distance refers to the position difference between two adjacent nodes on the screen. After obtaining the time interval and spatial distance of each group of adjacent nodes, it is necessary to perform edge weight analysis on these nodes. The purpose of edge weight analysis is to combine the time interval and spatial distance to calculate the edge weight data of each group of adjacent nodes. The calculation formula of the edge weight can be a weighted combination of the time interval and spatial distance. Based on the edge weight data of each group of adjacent nodes, the edge data of each group of adjacent nodes can be constructed. The edge data contains the association information between each pair of adjacent nodes, including the time interval, spatial distance, and edge weight. For example, if the teacher clicks on multiple options continuously within a short period of time and the distance between the click positions is small, the edge weight of these click operations will be high, indicating that these operations are highly correlated. Then, according to multiple groups of adjacent nodes and their edge data, construct the graph structure of the entire behavioral event. The graph structure consists of multiple nodes and edges. The nodes represent the operation behaviors at different moments, and the edges represent the associations between these operation behaviors. The process of constructing the graph structure includes adding edges for each pair of adjacent nodes and assigning the calculated weights to each edge. Finally, input the constructed target graph structure into the graph neural network for node behavior feature extraction. The graph neural network (GNN) is a deep learning model for processing graph-structured data that can capture complex relationships and patterns in the data. The input layer converts the node and edge data in the target graph structure into behavioral data vectors. The message passing layer aggregates the information of the neighbor nodes of each node and updates the node data. The feature extraction layer extracts the behavior features from the updated node data to form a set of behavior features. For example, assume that during a certain marking process, it is recorded that the teacher clicks on multiple options within a short period of time and frequently switches between pages.Through the above steps, multiple behavioral event nodes are generated from these click operations and page stays, and adjacent nodes are extracted from these nodes. The time interval and spatial distance between them are calculated to obtain edge weight data. Then, a target graph structure is constructed and input into a graph neural network for feature extraction to form a behavioral feature set. These behavioral feature sets can accurately reflect the operation habits and behavioral patterns of teachers, providing an important basis for subsequent abnormal behavior recognition. Through further analysis of these feature sets, rapid grading behaviors, abnormal grading distributions, and abnormal operation frequencies of teachers can be identified, thereby generating a detailed monitoring report and transmitting it to a preset data display terminal to help managers and teachers detect and handle abnormal behaviors in a timely manner, improving the transparency of the marking process and the credibility of grading results.

[0072] In a specific embodiment, the process of performing the step of inputting the target graph structure into a graph neural network for node behavioral feature extraction may specifically include the following steps:

[0073] (1) Input the target graph structure into the input layer for behavioral data vector conversion to obtain a behavioral data vector;

[0074] (2) Input the behavioral data vector into the message passing layer for neighbor node information aggregation to obtain neighbor node information for each behavioral event node;

[0075] (3) Update the node data for each behavioral event node according to the neighbor node information of each behavioral event node to obtain each updated behavioral event node;

[0076] (4) Input each updated behavioral event node into the feature extraction layer for node behavioral feature extraction to obtain a behavioral feature set.

[0077] Specifically, the constructed target graph structure is input into the input layer of a graph neural network (GNN). This graph structure consists of multiple nodes and edges. The nodes represent operation behaviors at different times, and the edges represent the associations between these operation behaviors. When the target graph structure is input into the input layer of the graph neural network, the task of the input layer is to convert the data of these nodes and edges into a behavioral data vector. The initial feature vector of each node includes the number of mouse clicks, the mouse click position, the page stay time, and the grading submission time, etc. For example, if a certain node represents a click operation of a teacher at a specific time point, then the initial feature vector of this node may be [number of clicks, click position X, click position Y, page stay time, grading submission time]. In this way, the input layer converts the node data in the graph structure into a behavioral data vector, preparing for subsequent graph convolution operations.

[0078] Next, the behavioral data vector is input into the message passing layer for neighbor node information aggregation. The task of the message passing layer is to collect information from the neighbor nodes of each node and aggregate this information to the node itself. The core of this step lies in combining the information of each node with the information of its neighbor nodes through the message passing mechanism, thereby updating the feature representation of the node. Specifically, each node receives information from its neighbor nodes, including the feature vectors of the neighbor nodes and the weights of the edges. Then, the node performs weighted summation or other aggregation operations on this information to obtain a new feature representation. After the neighbor node information aggregation is completed, the node data of each behavioral event node is updated according to the neighbor node information of each behavioral event node. The process of updating node data is similar to the forward propagation in a neural network. By performing weighted summation on the feature vectors of the node and its neighbor nodes and then performing a non-linear transformation through an activation function, a new node feature representation is obtained. Finally, each updated behavioral event node is input into the feature extraction layer for node behavioral feature extraction. The task of the feature extraction layer is to extract high-level behavioral features from the updated node feature vectors to form a set of behavioral features. This process is similar to the pooling layer or fully connected layer in a neural network. By reducing the dimension and aggregating the node features, a more abstract and generalized feature representation is obtained. For example, through global pooling operations, the node features in the entire graph can be aggregated into a global feature vector, or through a multi-layer perceptron (MLP), further non-linear transformations can be performed on the node features to extract key behavioral features.

[0079] For example, assume that during a certain marking process, the system records that the teacher clicks on the same option multiple times within a short period and frequently switches between multiple pages. The system first generates multiple behavioral event nodes from these click operations and page stays and constructs a target graph structure containing these nodes and edges. When the target graph structure is input into the graph neural network, the input layer converts the initial data of each node into a behavioral data vector, such as [number of clicks, click position X, click position Y, page stay time, score submission time]. Then, the message passing layer collects the neighbor node information of each node and performs weighted summation and activation operations to update the feature representation of the node. Through this process, the system can combine the operation behaviors of each node with the behaviors of its neighbor nodes to capture more complex behavioral patterns. Finally, the feature extraction layer extracts high-level behavioral features from the updated node feature vectors to form a set of behavioral features. These sets of behavioral features can accurately reflect the teacher's operation habits and behavioral patterns, providing an important basis for subsequent behavioral pattern recognition and abnormal behavior detection.

[0080] In a specific embodiment, the process of performing step S103 may specifically include the following steps:

[0081] (1) Input the behavior feature set into the encoder for latent variable mapping to obtain the latent variable data corresponding to the behavior feature set, where the latent variable data includes: mean data and standard deviation data;

[0082] (2) Input the latent variable data into the latent space layer for latent variable sampling to obtain the sampled latent variable data;

[0083] (3) Input the latent variable data into the decoder for data reconstruction mapping to obtain the reconstructed data;

[0084] (4) Perform behavior pattern recognition on the reconstructed data to obtain the behavior pattern set.

[0085] Specifically, input the behavior feature set extracted from the graph neural network into the encoder. The encoder is a neural network module responsible for mapping the high-dimensional behavior feature set to the low-dimensional latent space. In the encoder, each behavior feature vector is processed and mapped into a latent variable, which includes two parts: mean and standard deviation. For example, assume there is a feature vector x. The encoder, through a series of fully connected layers and activation functions, finally outputs the mean μ and standard deviation σ of the latent variable. Then, input the latent variable data into the latent space layer for latent variable sampling. Since it is desired that the latent variable can reflect the main distribution law of the behavior features, the reparameterization trick is usually used to achieve differentiable sampling.

[0086] Next, input the sampled latent variable into the decoder for data reconstruction mapping. The decoder is another neural network module responsible for mapping the low-dimensional latent variable back to the high-dimensional feature space and reconstructing data similar to the original input feature vector. The decoder, through a series of fully connected layers and activation functions, converts the latent variable into the reconstructed data. After obtaining the reconstructed data, perform behavior pattern recognition on these reconstructed data to identify normal and abnormal behavior patterns. During the behavior pattern recognition process, further analysis and classification of the reconstructed data are carried out to find out the abnormal behaviors. For example, if the scoring submission time and page stay time corresponding to a certain reconstructed data are significantly abnormal, it will be marked as an abnormal behavior. In this way, a behavior pattern set can be constructed, which includes all the identified normal and abnormal behavior patterns.

[0087] For example, assume that during a certain marking process, it is recorded that the teacher submits grades quickly multiple times within a short period and frequently switches between multiple pages. First, extract the feature set from these behavior data through a graph neural network and input it into the encoder. The encoder maps these high-dimensional behavior features to the mean and standard deviation of latent variables, and then, through the reparameterization trick, samples latent variables from these Gaussian distributions. Next, input the latent variables into the decoder, and the decoder reconstructs the latent variables into data similar to the original features. Through further analysis of the reconstructed data, it is found that some of the behavior patterns are significantly different from normal operations, such as the grading submission time being abnormally short and the page stay time being abnormally long. These abnormal behaviors are identified and marked to form a behavior pattern set.

[0088] In this way, the variational autoencoder can effectively map high-dimensional behavior features to a low-dimensional latent space, and then reconstruct the data through the decoder, ultimately realizing the recognition of behavior patterns. The advantage of the variational autoencoder is that it can capture the main distribution laws of behavior features, generate new samples to enhance the robustness of the model, and at the same time distinguish normal and abnormal behavior patterns through the representation in the latent space. This method not only improves the accuracy of behavior pattern recognition but also enhances the processing ability in the face of complex behavior data, thus effectively improving the transparency of the marking process and the credibility of the grading results.

[0089] In a specific embodiment, the process before performing step S104 may specifically include the following steps:

[0090] (1) Collect the historical behavior pattern set, preprocess the historical behavior pattern set to obtain positive sample pairs and negative sample pairs;

[0091] (2) Input the positive sample pairs and negative sample pairs into the initial contrast learning algorithm for feature extraction to obtain the first feature vector corresponding to the positive sample pairs and the second feature vector corresponding to the negative sample pairs;

[0092] (3) Calculate the loss value of the initial contrast learning algorithm based on the first feature vector and the second feature vector to obtain the target loss value;

[0093] (4) When the target loss value meets the preset threshold, obtain the contrast learning algorithm.

[0094] Specifically, a large amount of historical behavior pattern data is extracted, which can include all the operation records of past teachers, such as the number of mouse clicks, click positions, page residence times, and grading submission times. In this way, we can obtain a historical behavior pattern set containing rich operation behaviors. Next, the historical behavior pattern set is preprocessed to divide it into positive sample pairs and negative sample pairs. Positive sample pairs refer to two sets of similar behavior patterns, such as the normal operation behaviors of the same teacher at different time periods; negative sample pairs refer to two sets of different behavior patterns, such as the comparison between normal behavior and abnormal behavior. To construct effective sample pairs, we need to analyze and annotate the historical behavior pattern data. For example, annotate which operation behaviors are normal and which operation behaviors may be abnormal. Then, pair similar normal behaviors as positive sample pairs and pair normal behaviors with abnormal behaviors as negative sample pairs.

[0095] After constructing the positive sample pairs and negative sample pairs, these sample pairs are input into the initial contrast learning algorithm for feature extraction. The contrast learning algorithm is a self-supervised learning method that can train the model by comparing between sample pairs. Specifically, the goal of positive sample pairs is to make their feature vectors closer in the feature space, while the goal of negative sample pairs is to make their feature vectors farther apart in the feature space. The contrast learning algorithm inputs the behavior features of each sample pair through a feature encoder and converts them into feature vectors.

[0096] After calculating the loss value, the target loss value is obtained. If the target loss value meets the preset threshold, the training of the contrast learning algorithm reaches the requirement, and it can be considered that the contrast learning algorithm has completed training and has the ability to identify positive and negative sample pairs. In this way, the contrast learning algorithm can effectively distinguish normal and abnormal behaviors in the feature space.

[0097] For example, assume that during a certain marking process, a large amount of historical behavior data is extracted and these data are annotated. Through analysis and annotation, it is determined which operation behaviors are normal and which operation behaviors may be abnormal. For example, the grading submission time of a teacher is usually between 20 and 30 seconds under normal circumstances, while the grading submission time in abnormal circumstances may be less than 5 seconds. Then, these normal behaviors are paired as positive sample pairs, and normal behaviors are paired with abnormal behaviors as negative sample pairs.

[0098] After constructing the positive sample pairs and negative sample pairs, these sample pairs are input into the initial contrast learning algorithm for feature extraction. For example, for the positive sample pair (x1, x2), x1 and x2 are respectively converted into feature vectors h1 and h2. For the negative sample pair (x1, x3), x1 and x3 are respectively converted into feature vectors h1 and h3. In this way, the first feature vector pair of positive sample pairs and the second feature vector pair of negative sample pairs are obtained.

[0099] Next, based on these feature vector pairs, the loss value is calculated. Using the contrastive loss function, the similarity of the positive sample pairs and the similarity of the negative sample pairs are calculated. For example, the feature vectors h1 and h2 of the positive sample pairs should be closer in the feature space, while the feature vectors h1 and h3 of the negative sample pairs should be farther apart in the feature space. In this way, the target loss value is calculated.

[0100] Finally, it is checked whether the target loss value meets the preset threshold. If the target loss value meets the preset threshold, the training of the contrastive learning algorithm meets the requirements, and it is considered that the contrastive learning algorithm already has the ability to identify positive and negative sample pairs. For example, the preset threshold is set to 0.1. If the calculated target loss value is less than 0.1, the training of the contrastive learning algorithm is successful, and it can effectively distinguish normal and abnormal behaviors in the feature space.

[0101] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0102] (1) Classify the abnormal behavior data to obtain abnormal data of multiple abnormal types;

[0103] (2) Conduct statistical feature analysis on the abnormal data of each abnormal type to obtain the statistical features of the abnormal data of each abnormal type;

[0104] (3) Extract the behavior details of the abnormal data of each abnormal type to obtain the detail data of the abnormal data of each abnormal type;

[0105] (4) Analyze the monitoring report according to the statistical features of the abnormal data of each abnormal type and the detail data of the abnormal data of each abnormal type to obtain the marking behavior monitoring report, and transmit the marking behavior monitoring report to the preset data display terminal.

[0106] Specifically, the identified abnormal behavior data is classified. The abnormal behavior data may include rapid scoring behavior, abnormal scoring distribution data, and abnormal operation frequency, etc. During the classification process, the abnormal behavior data will be classified into different abnormal types according to the characteristics of the abnormal behavior. For example, if a certain teacher submits scores rapidly multiple times in a short period, these behaviors will be classified as rapid scoring behavior; if the scoring distribution significantly deviates from the normal range, these behaviors will be classified as abnormal scoring distribution data.

[0107] Next, perform statistical feature analysis on the abnormal data of each abnormal type. The purpose of statistical feature analysis is to extract the overall features of the abnormal data, facilitating subsequent extraction of behavioral details and analysis of monitoring reports. For example, for the rapid scoring behavior, the time interval between each scoring can be calculated to identify behaviors with abnormally short scoring times. Specific statistical features can include the number of scorings, scoring time intervals, scoring distributions, etc. Through these statistical features, the overall situation of abnormal behaviors can be better understood.

[0108] Then, extract the behavioral details of the abnormal data for each abnormal type. The purpose of extracting behavioral details is to deeply analyze the specific operation details of each abnormal behavior, facilitating the provision of detailed monitoring reports. For example, for the rapid scoring behavior, the specific time of each scoring, the specific content of the scoring, and the page location where the scoring is submitted can be recorded. Through these detailed data, a more detailed description of the abnormal behavior can be provided, offering valuable reference information for management personnel.

[0109] After obtaining the statistical features and behavioral details of each abnormal type, conduct monitoring report analysis based on these data. The purpose of monitoring report analysis is to integrate all the statistical features and detailed data to generate a comprehensive monitoring report. This report should include an overview of abnormal behaviors, statistical charts, specific details, and improvement suggestions. For example, the monitoring report can include the occurrence times, time distribution, involved teachers and test questions of each abnormal type. At the same time, the report can also contain statistical charts, such as the time interval distribution chart of the rapid scoring behavior and the bar chart of the abnormal scoring distribution. These charts can intuitively display the patterns and characteristics of abnormal behaviors.

[0110] Finally, transmit the generated monitoring report of the marking behavior to the pre-set data display terminal. The data display terminal is a user-friendly interface through which management personnel and teachers can view the detailed content of the monitoring report. The report can be transmitted via network transmission protocols such as HTTP or HTTPS to ensure the secure transmission of report data. The data display terminal should have good interactivity, allowing users to view different abnormal behavior categories and detailed information through clicking and filtering. For example, management personnel can choose to view the detailed report of the rapid scoring behavior, check the specific operation time and scoring content of each rapid scoring, and thus take corresponding improvement measures.

[0111] For example, assume that during a certain marking process, it is identified that a teacher submits marks multiple times quickly within a short period, and the mark distribution is abnormal. First, classify the data of these abnormal behaviors. Classify the behavior of quickly submitting marks as quick marking behavior, and classify the data with abnormal mark distribution as abnormal mark distribution data. Then, conduct statistical feature analysis on these classified abnormal data. For the quick marking behavior, calculate the time interval between each mark submission, and it is found that there are multiple cases where the time interval between marks is less than 5 seconds, which is significantly lower than the normal mark submission time interval. For the abnormal mark distribution data, calculate the distribution of each mark, and it is found that the mark distribution significantly deviates from the normal range and there is a large deviation.

[0112] Next, extract the behavior details of the abnormal data for each abnormal type. For the quick marking behavior, record the specific time of each mark, the content of the mark, and the page position where the mark is submitted. For example, it is recorded that a certain mark was submitted at 10:05 on May 26, 2023, the mark content is 3 points, and the submission position is at the lower right corner of the page. For the abnormal mark distribution data, record the distribution of each mark and the specific score. For example, it is recorded that a certain mark is 2 points, but according to the marking criteria, this mark should be 4 points.

[0113] After obtaining the statistical features and behavior details of each abnormal type, conduct a monitoring report analysis based on this data. The monitoring report includes an overview of the abnormal behaviors, such as the number of occurrences of quick marking behavior is 10 times, and the number of occurrences of abnormal mark distribution data is 5 times; statistical charts, such as the time interval distribution chart of quick marking behavior shows that most mark time intervals are less than 5 seconds, and the bar chart of abnormal mark distribution shows a large mark deviation; specific details, such as the specific operation time and mark content of each quick marking, and the specific marks and marking criteria of abnormal mark distribution. The monitoring report also includes improvement suggestions, such as suggesting that teachers pay attention to the marking time and avoid quickly submitting marks, and suggesting that managers check the marking criteria to ensure a reasonable mark distribution.

[0114] Finally, transmit the generated marking behavior monitoring report to the pre-set data display terminal through the HTTP protocol. Managers and teachers can view the detailed monitoring report through the data display terminal to understand the specific situation and improvement suggestions of each abnormal behavior. For example, managers can view the detailed report of quick marking behavior and find that a certain teacher submitted marks quickly multiple times within a certain time period, and thus take corresponding measures to remind the teacher to pay attention to the marking time. In this way, it is possible to effectively monitor and analyze abnormal behaviors, improve the transparency of the marking process and the credibility of the marking results, and provide valuable reference information for managers and teachers.

[0115] In a specific embodiment, the process of performing the monitoring report analysis step based on the statistical characteristics of the abnormal data of each abnormal type and the detailed data of the abnormal data of each abnormal type may specifically include the following steps:

[0116] (1) Generate a statistical chart based on the statistical characteristics of the abnormal data of each abnormal type and the detailed data of the abnormal data of each abnormal type;

[0117] (2) Construct data display elements according to the statistical icon and match the rendering parameters of the data display elements;

[0118] (3) Construct a marking behavior monitoring report according to the rendering parameters of the data display elements and the data display elements, and transmit the marking behavior monitoring report to the data display terminal.

[0119] It should be noted that according to the statistical characteristics and detailed data of the abnormal data of each abnormal type obtained in the previous steps, corresponding statistical charts are generated. These statistical charts can intuitively display the distribution and characteristics of abnormal behaviors, helping managers and teachers better understand the specific situation of abnormal behaviors. For example, assume there are two main types of abnormal behaviors: rapid grading behavior and abnormal grading distribution data. For rapid grading behavior, the time interval between each grading has been calculated, and those records with abnormally short grading times have been identified. To generate a statistical chart, these time interval data can be used to draw a time interval distribution chart. For example, the X-axis represents the grading time interval, and the Y-axis represents the number of gradings. In this way, managers can intuitively see which time periods have shorter grading time intervals, thus identifying the time points where rapid grading behavior may exist. Similarly, for abnormal grading distribution data, the deviation of the grading distribution has been calculated. To generate a statistical chart, the grading deviation can be plotted as a bar chart or a box plot. For example, the X-axis represents the specific scores of the grading distribution, and the Y-axis represents the degree of grading deviation. In this way, managers can intuitively see the specific distribution of the grading deviation, thus identifying the specific data points with abnormal grading distributions. After generating these statistical charts, the next step is to construct data display elements based on the statistical charts and match the rendering parameters of the data display elements. Data display elements refer to the specific components used to display statistical charts in the graphical user interface, such as line charts, bar charts, pie charts, etc. These display elements need to be matched and configured according to the data characteristics of the statistical charts to ensure the display effect and user experience of the charts. For example, for the time interval distribution chart, a line chart can be selected as the display element, and the labels, scales, and colors of the X-axis and Y-axis can be configured as rendering parameters; for the bar chart of grading deviation, a bar chart can be selected as the display element, and the colors, heights, and widths of the bars can be configured as rendering parameters. In this way, data display elements that conform to the characteristics of the statistical charts can be constructed, and appropriate rendering parameters can be matched for each display element. Next, based on these data display elements and rendering parameters, a monitoring report on the grading behavior will be constructed. The monitoring report is a comprehensive document that contains all the statistical charts and detailed analysis results of abnormal behaviors. These statistical charts and analysis results will be embedded in the monitoring report as part of the data display elements to display the specific situation of abnormal behaviors in a vivid and graphic way. For example, assume that during a certain grading process, it is identified that a teacher submitted grades rapidly multiple times in a short period, and the grading distribution is abnormal. First, two statistical charts are generated: one is the time interval distribution chart of rapid grading behavior, and the other is the bar chart of grading deviation. Next, a line chart is selected as the data display element for the time interval distribution chart, and the labels of the X-axis and Y-axis are configured as "grading time interval (seconds)" and "number of gradings", and the color of the line is configured as blue.For the bar chart of scoring deviation, select the bar chart as the data display element, configure the label of the X-axis as "Scoring Score" and the label of the Y-axis as "Scoring Deviation", and at the same time configure the color of the bars to be red. After configuring the data display elements and rendering parameters, embed these display elements into the marking behavior monitoring report. The monitoring report includes multiple parts, including an overview of abnormal behaviors, statistical charts, detailed analysis, and improvement suggestions. For example, the first part of the report is an overview of abnormal behaviors, listing the occurrence times and time distributions of rapid scoring behaviors and abnormal scoring distribution data; the second part is statistical charts, showing the time interval distribution chart and the bar chart of scoring deviation; the third part is detailed analysis, describing the specific operation time, scoring content, scoring location, etc. of each abnormal behavior; the fourth part is improvement suggestions, providing specific improvement measures for teachers, such as avoiding rapid submission of scores to ensure the rationality and fairness of scoring. Finally, transmit the generated marking behavior monitoring report to the preset data display terminal through the network transmission protocol. The data display terminal is a user-friendly interface, and managers and teachers can view the detailed content of the monitoring report through this terminal. The report can be transmitted through the HTTP protocol or other data transmission protocols to ensure the secure transmission of report data. The data display terminal should have good interactivity, and users can view different categories of abnormal behaviors and detailed information by clicking and filtering. For example, managers can choose to view the detailed report of rapid scoring behaviors, view the specific operation time and scoring content of each rapid scoring, and thus take corresponding improvement measures.

[0120] An embodiment of the present invention also provides an intelligent monitoring system based on online marking behavior, as Figure 2 shown. The intelligent monitoring system based on online marking behavior specifically includes:

[0121] A collection module 201, configured to collect the behavior data of the marking teacher to obtain marking behavior data, where the marking behavior data includes: the number of mouse clicks, the mouse click position, the page stay time, and the scoring submission time;

[0122] An extraction module 202, configured to input the marking behavior data into a graph neural network for behavior feature extraction to obtain a behavior feature set;

[0123] An identification module 203, configured to input the behavior feature set into a variational autoencoder for behavior pattern identification to obtain a behavior pattern set;

[0124] An input module 204, configured to input the behavior pattern set into a contrastive learning algorithm for abnormal behavior identification to obtain abnormal behavior data, where the abnormal behavior data includes: rapid scoring behavior, abnormal scoring distribution data, abnormal operation frequency;

[0125] The analysis module 205 is configured to monitor and report the analysis of the abnormal behavior data to obtain a marking behavior monitoring report, and transmit the marking behavior monitoring report to a preset data display terminal.

[0126] Through the collaborative work of the above modules, by collecting the behavioral data of the marking teachers, including the number of mouse clicks, mouse click positions, page dwell time, and score submission time, the teachers' operation behaviors can be recorded comprehensively and meticulously, providing a rich data basis for subsequent analysis and identification. These behavioral data are transmitted to the data processing center in real time through the data acquisition module, realizing the timely acquisition and processing of data, and ensuring the real-time and accuracy of monitoring. Secondly, the scheme uses graph neural networks for behavioral feature extraction. By constructing the behavioral data into a spatiotemporal graph, taking the operation behaviors at different time points as nodes, and the associations between behaviors as edges, and performing weighted analysis on these nodes and edges, the spatiotemporal dependencies and implicit patterns in the behavioral data can be effectively captured. This feature extraction method is richer and more accurate than the traditional planar feature extraction method, and can better reflect the teachers' operating habits and behavioral patterns. Furthermore, the scheme uses variational autoencoders for behavioral pattern recognition. By mapping the high-dimensional behavioral feature vector to the low-dimensional latent space and learning the behavioral pattern by maximizing the posterior distribution, normal and abnormal marking behaviors can be distinguished in the latent space. The variational autoencoder can not only effectively reduce the dimension, but also generate new behavior samples, enhance the robustness and generalization ability of the model, and provide a high-quality behavior pattern set for subsequent anomaly detection. Next, the scheme uses a contrastive learning algorithm to identify abnormal behaviors. By constructing positive and negative sample pairs and training feature encoders, similar behavior features are closer in the feature space and different behavior features are farther away, thus achieving accurate identification of abnormal behaviors. Compared with traditional supervised learning methods, the contrastive learning algorithm not only improves the detection accuracy of the model, but also can use unlabeled data for effective model training without a large amount of labeled data. The scheme also includes monitoring report analysis of abnormal behavior data. By classifying and statistically analyzing abnormal behavior data, a detailed monitoring report is generated and transmitted to the preset data display terminal. In this way, detailed abnormal behavior reports can be provided to management personnel and marking teachers in a timely manner to help them discover and deal with abnormal behaviors in a timely manner, improve the transparency of the marking process and the credibility of the scoring results. In addition, the monitoring report also includes specific improvement suggestions and measures to help teachers improve their operating behaviors and improve the quality of marking. Overall, this solution builds a real-time, intelligent marking behavior monitoring system by comprehensively using graph neural networks, variational autoencoders and contrastive learning algorithms. It can not only comprehensively and accurately collect and analyze teachers' operating behaviors, but also identify and report abnormal behaviors in a timely and accurate manner. It has high practical value and creativity, effectively solves the shortcomings of existing technologies, and provides new technical means for the intelligent monitoring of online marking systems.Especially in the aspects of statistical chart generation and the rendering of data display elements, by analyzing the statistical characteristics and detailed data of abnormal data for each abnormal type, corresponding statistical charts are generated, such as the time interval distribution chart and the score deviation bar chart, which can intuitively display the distribution and characteristics of abnormal behaviors, helping managers and teachers better understand and handle abnormal behaviors. These statistical charts ensure the display effect and user experience of the charts by selecting appropriate data display elements and matching corresponding rendering parameters. After generating these display elements, the system can embed them into the marking behavior monitoring report to construct a comprehensive monitoring report document, including an overview of abnormal behaviors, statistical charts, detailed analysis, and improvement suggestions, providing comprehensive reference information for managers and teachers to help them take improvement measures in a timely manner and enhance the transparency of the marking process and the credibility of the scoring results. Finally, the monitoring report is securely transmitted to the pre-set data display terminal through the network transmission protocol, ensuring that managers and teachers can conveniently view and process the report content, further enhancing the practicality and effectiveness of the system.

[0127] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that: still modifications or equivalent substitutions can be made to the specific implementation manners of the present invention, and any modification or equivalent substitution without departing from the spirit and scope of the present invention shall be covered by the scope of the claims of the present invention.

Claims

1. An intelligent monitoring method based on online marking behavior, characterized in that, Including: Collecting the behavior data of the marking teachers to obtain marking behavior data, where the marking behavior data includes: the number of mouse clicks, the mouse click position, the page stay time, and the score submission time; Inputting the marking behavior data into a graph neural network for behavior feature extraction to obtain a behavior feature set, including: generating a plurality of behavior event nodes according to the number of mouse clicks, the mouse click position, the page stay time, and the score submission time; extracting adjacent nodes from the plurality of behavior event nodes to obtain multiple groups of adjacent nodes; calculating the interval time for each group of adjacent nodes respectively to obtain the interval time of each group of adjacent nodes; calculating the spatial distance for each group of adjacent nodes respectively to obtain the spatial distance of each group of adjacent nodes; analyzing the edge weights for each group of adjacent nodes respectively according to the interval time of each group of adjacent nodes and the spatial distance of each group of adjacent nodes to obtain the edge weight data of each group of adjacent nodes; constructing the edge data of each group of adjacent nodes based on the edge weight data of each group of adjacent nodes, and constructing a graph structure according to the multiple groups of adjacent nodes and the edge data of each group of adjacent nodes to obtain a target graph structure; Inputting the target graph structure into the graph neural network for node behavior feature extraction to obtain the behavior feature set; the graph neural network includes: an input layer, a message passing layer, and a feature extraction layer, and the step of inputting the target graph structure into the graph neural network for node behavior feature extraction to obtain the behavior feature set includes: inputting the target graph structure into the input layer for behavior data vector conversion to obtain a behavior data vector; inputting the behavior data vector into the message passing layer for neighbor node information aggregation to obtain the neighbor node information of each behavior event node; updating the node data of each behavior event node according to the neighbor node information of each behavior event node to obtain each updated behavior event node; inputting each updated behavior event node into the feature extraction layer for node behavior feature extraction to obtain the behavior feature set; Among them, when clicking on an option during the marking process, record the time and click position of this operation, and use it as a behavioral event node; when staying on a certain page for more than a certain period of time, record the stay time and generate a corresponding node; the scoring submission time is a key data point, and each time the score is submitted, record the specific submission time and generate a node; after generating multiple behavioral event nodes, extract adjacent nodes; the purpose of adjacent node extraction is to pair nodes that are close in time to form multiple groups of adjacent nodes; if multiple options are continuously clicked within a short period of time, the nodes of these click operations will be extracted as a group of adjacent nodes; the stay times between different pages can also form adjacent nodes; then, for each group of adjacent nodes, calculate the time interval between them; the time interval refers to the time difference between two adjacent nodes; if node A and node B represent click operations at two different time points, calculate the time difference between the two adjacent nodes; calculate the spatial distance between each group of adjacent nodes; the spatial distance refers to the position difference between two adjacent nodes on the screen; after obtaining the time interval and spatial distance of each group of adjacent nodes, perform edge weight analysis on these nodes; the purpose of edge weight analysis is to combine the time interval and spatial distance to calculate the edge weight data of each group of adjacent nodes; the calculation formula of the edge weight is a weighted combination of the time interval and spatial distance, and based on the edge weight data of each group of adjacent nodes, construct the edge data of each group of adjacent nodes; the edge data contains the association information between each pair of adjacent nodes, including the time interval, spatial distance, and edge weight; if multiple options are continuously clicked within a short period of time and the distance between the click positions is small, the edge weight of these click operations will be high, indicating that these operations are highly correlated; then, according to multiple groups of adjacent nodes and their edge data, construct the graph structure of the entire behavioral event; the graph structure consists of multiple nodes and edges, where the nodes represent operation behaviors at different moments and the edges represent the associations between these operation behaviors; the process of constructing the graph structure includes adding edges for each pair of adjacent nodes and assigning the calculated weights to each edge; input the constructed target graph structure into a graph neural network for node behavior feature extraction; the input layer converts the node and edge data in the target graph structure into behavioral data vectors, the message passing layer aggregates the neighbor node information of each node, updates the node data, and the feature extraction layer extracts behavioral features from the updated node data to form a behavioral feature set; Input the behavioral feature set into a variational autoencoder for behavioral pattern recognition to obtain a behavioral pattern set; Input the behavioral pattern set into a contrastive learning algorithm for abnormal behavior recognition to obtain abnormal behavior data, where the abnormal behavior data includes: rapid scoring behavior, abnormal scoring distribution data, and abnormal operation frequency; Conduct a monitoring report analysis on the abnormal behavior data to obtain a marking behavior monitoring report, and transmit the marking behavior monitoring report to a preset data display terminal.

2. The intelligent monitoring method based on online marking behavior according to claim 1, wherein The variational autoencoder includes an encoder, a latent space layer, and a decoder. The step of inputting the set of behavioral features into the variational autoencoder for behavioral pattern recognition to obtain a set of behavioral patterns includes: Inputting the set of behavioral features into the encoder for latent variable mapping to obtain latent variable data corresponding to the set of behavioral features, where the latent variable data includes: mean data and standard deviation data; Inputting the latent variable data into the latent space layer for latent variable sampling to obtain sampled latent variable data; Inputting the latent variable data into the decoder for data reconstruction mapping to obtain reconstructed data; Performing behavioral pattern recognition on the reconstructed data to obtain the set of behavioral patterns.

3. The intelligent monitoring method based on online marking behavior according to claim 1, wherein, Before the step of inputting the set of behavioral patterns into a contrastive learning algorithm for abnormal behavior recognition to obtain abnormal behavior data, where the abnormal behavior data includes: rapid scoring behavior, abnormal scoring distribution data, and abnormal operation frequency, the following steps are also included: Collecting a set of historical behavioral patterns, preprocessing the set of historical behavioral patterns to obtain positive sample pairs and negative sample pairs; Inputting the positive sample pairs and the negative sample pairs into an initial contrastive learning algorithm for feature extraction to obtain a first feature vector corresponding to the positive sample pairs and a second feature vector corresponding to the negative sample pairs; Calculating a loss value for the initial contrastive learning algorithm based on the first feature vector and the second feature vector to obtain a target loss value; When the target loss value meets a preset threshold, obtaining the contrastive learning algorithm.

4. The intelligent monitoring method based on online marking behavior according to claim 1, wherein The step of performing monitoring report analysis on the abnormal behavior data to obtain a marking behavior monitoring report and transmitting the marking behavior monitoring report to a preset data display terminal includes: Classifying the abnormal behavior data to obtain abnormal data of multiple abnormal types; Respectively performing statistical feature analysis on the abnormal data of each abnormal type to obtain statistical features of the abnormal data of each abnormal type; Respectively extracting behavioral details from the abnormal data of each abnormal type to obtain detail data of the abnormal data of each abnormal type; Performing monitoring report analysis according to the statistical features of the abnormal data of each abnormal type and the detail data of the abnormal data of each abnormal type to obtain a marking behavior monitoring report, and transmitting the marking behavior monitoring report to a preset data display terminal.

5. The intelligent monitoring method based on online marking behavior according to claim 4, wherein The step of performing monitoring report analysis according to the statistical features of the abnormal data of each abnormal type and the detail data of the abnormal data of each abnormal type to obtain a marking behavior monitoring report and transmitting the marking behavior monitoring report to a preset data display terminal includes: Generating a statistical chart according to the statistical features of the abnormal data of each abnormal type and the detail data of the abnormal data of each abnormal type; Constructing data display elements according to the statistical chart and matching rendering parameters of the data display elements; Constructing a marking behavior monitoring report according to the rendering parameters of the data display elements and the data display elements, and transmitting the marking behavior monitoring report to the data display terminal.

6. An intelligent monitoring system based on online marking behavior, which is used to execute the intelligent monitoring method based on online marking behavior according to any one of claims 1 to 5, characterized in that, Including: A collection module, configured to collect the behavior data of the marking teachers to obtain marking behavior data, where the marking behavior data includes: the number of mouse clicks, the mouse click position, the page stay time, and the score submission time; An extraction module, configured to input the marking behavior data into a graph neural network for behavior feature extraction to obtain a behavior feature set, including: generating multiple behavior event nodes according to the number of mouse clicks, the mouse click position, the page stay time, and the score submission time; extracting adjacent nodes from the multiple behavior event nodes to obtain multiple groups of adjacent nodes; calculating the interval time for each group of adjacent nodes to obtain the interval time of each group of adjacent nodes; calculating the spatial distance for each group of adjacent nodes to obtain the spatial distance of each group of adjacent nodes; performing edge weight analysis on each group of adjacent nodes according to the interval time of each group of adjacent nodes and the spatial distance of each group of adjacent nodes to obtain the edge weight data of each group of adjacent nodes; constructing the edge data of each group of adjacent nodes based on the edge weight data of each group of adjacent nodes, and constructing a graph structure according to the multiple groups of adjacent nodes and the edge data of each group of adjacent nodes to obtain a target graph structure; Inputting the target graph structure into the graph neural network for node behavior feature extraction to obtain the behavior feature set; the graph neural network includes: an input layer, a message passing layer, and a feature extraction layer, and the step of inputting the target graph structure into the graph neural network for node behavior feature extraction to obtain the behavior feature set includes: inputting the target graph structure into the input layer for behavior data vector conversion to obtain a behavior data vector; inputting the behavior data vector into the message passing layer for neighbor node information aggregation to obtain the neighbor node information of each behavior event node; updating the node data of each behavior event node according to the neighbor node information of each behavior event node to obtain each updated behavior event node; inputting each updated behavior event node into the feature extraction layer for node behavior feature extraction to obtain the behavior feature set; Among them, when clicking on an option during the marking process, record the time and click position of this operation, and use it as a behavioral event node; when staying on a certain page for more than a certain time, record the stay time and generate a corresponding node; the scoring submission time is a key data point, and each time the scoring is submitted, record the specific submission time and generate a node; after generating multiple behavioral event nodes, extract adjacent nodes from the nodes; the purpose of adjacent node extraction is to pair nodes that are close in time to form multiple groups of adjacent nodes; if multiple options are continuously clicked within a short period of time, the nodes of these click operations will be extracted as a group of adjacent nodes; the stay times between different pages can also form adjacent nodes; then, for each group of adjacent nodes, calculate the time interval between them; the time interval refers to the time difference between two adjacent nodes; if node A and node B represent click operations at two different time points, calculate the time difference between the two adjacent nodes; calculate the spatial distance between each group of adjacent nodes; the spatial distance refers to the position difference between two adjacent nodes on the screen; after obtaining the time interval and spatial distance of each group of adjacent nodes, perform edge weight analysis on these nodes; the purpose of edge weight analysis is to combine the time interval and spatial distance to calculate the edge weight data of each group of adjacent nodes; the calculation formula of the edge weight is the weighted combination of the time interval and spatial distance, and based on the edge weight data of each group of adjacent nodes, construct the edge data of each group of adjacent nodes; the edge data contains the association information between each pair of adjacent nodes, including the time interval, spatial distance, and edge weight; if multiple options are continuously clicked within a short period of time and the distance between the click positions is small, the edge weight of these click operations will be very high, indicating that these operations are highly correlated; then, according to multiple groups of adjacent nodes and their edge data, construct the graph structure of the entire behavioral event; the graph structure consists of multiple nodes and edges, the nodes represent the operation behaviors at different times, and the edges represent the associations between these operation behaviors; the process of constructing the graph structure includes adding edges for each pair of adjacent nodes and assigning the calculated weights to each edge; input the constructed target graph structure into a graph neural network for node behavior feature extraction; the input layer converts the node and edge data in the target graph structure into behavioral data vectors, the message passing layer aggregates the neighbor node information of each node, updates the node data, and the feature extraction layer extracts behavioral features from the updated node data to form a set of behavioral features; An identification module, configured to input the set of behavioral features into a variational autoencoder for behavioral pattern recognition to obtain a set of behavioral patterns; An input module, configured to input the set of behavioral patterns into a contrastive learning algorithm for abnormal behavior recognition to obtain abnormal behavior data, where the abnormal behavior data includes: rapid scoring behavior, abnormal scoring distribution data, and abnormal operation frequency; An analysis module, configured to perform monitoring report analysis on the abnormal behavior data to obtain a marking behavior monitoring report, and transmit the marking behavior monitoring report to a preset data display terminal.

Images