Examinee examination room abnormal behavior analysis method and system for online examination

Abstract

The invention provides an examination room abnormal behavior analysis method and system for examinees in a computer examination, and the method comprises the steps: extracting the facial expression, body posture information, keyboard tapping frequency, mouse movement track and other behavior data of the examinees, and forming a comprehensive behavior feature set; a time sequence prediction algorithm based on a long and short term memory network is combined with an attention mechanism to analyze a behavior pattern in a time dimension, and abnormal behavior information is obtained; creating a target examination room environment graph through a graph neural network algorithm, analyzing abnormal behavior information, and identifying collaborative cheating behaviors; introducing a Q learning reinforcement learning framework, and dynamically adjusting a detection strategy; performing weighted fusion on the abnormal behavior information, the collaborative cheating recognition result and the detection strategy by using an extreme gradient lifting learning algorithm, calculating a comprehensive score, and generating an abnormal behavior judgment report based on the comprehensive score; according to the method, the detection capability and efficiency of the abnormal behaviors and the cheating behaviors in the online examination are improved, and the fairness and fairness of the examination can be maintained.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of data processing technology, and in particular to a method and system for analyzing abnormal behavior of examinees in a computer-based examination. Background Art

[0002] In modern education and examination systems, computer-based examinations have become a common form of examination. In order to ensure the fairness and impartiality of the examinations, it is particularly important to monitor and analyze the behavior of candidates during the examination.

[0003] Traditional invigilation methods mainly rely on manual and camera monitoring. The existing solution uses manual patrols to monitor, which makes it difficult to cover the behavior of all candidates and is prone to missing cheating. The existing solution uses image analysis technology to analyze the examination room video captured by the camera to identify abnormal behavior of candidates. This technology only relies on facial expressions and body posture information and lacks comprehensive analysis of other behavioral data, resulting in low recognition accuracy. The natural movements of candidates may lead to false alarms and missed alarms, affecting the reliability and practicality of recognition. Summary of the invention

[0004] The embodiment of the present invention provides a method and system for analyzing abnormal behavior of examinees in a computer-based examination, so as to solve the problems in the prior art that it is difficult to cover the behaviors of all examinees, cheating behaviors are easily missed, the accuracy of identifying examinees' cheating behaviors is low, and the reliability and practicality are low.

[0005] In a first aspect, an embodiment of the present invention provides a method for analyzing abnormal behavior of examinees in a computer-based examination, comprising: Extracting the examinee's facial expression and body posture information from the monitoring image, and using a behavior recognition algorithm to analyze the keyboard tapping frequency and the mouse movement trajectory to obtain behavior information, integrating the examinee's facial expression, the body posture information and the behavior information to obtain a comprehensive behavior feature set; Applying a time series prediction algorithm based on a long short-term memory network and combining it with an attention mechanism optimization model, the behavior pattern on the time dimension of the comprehensive behavior feature set is analyzed to obtain abnormal behavior information; A target examination room environment map is created based on the candidate seat layout through a graph neural network algorithm, collaborative cheating behavior is identified by analyzing the abnormal behavior information, and collaborative cheating identification results are obtained. A reinforcement learning framework based on Q learning is introduced, and the detection strategy is dynamically adjusted according to the abnormal behavior information and the collaborative cheating identification results to obtain a target detection strategy; An extreme gradient boosting learning algorithm is used to weightedly fuse the abnormal behavior information, the collaborative cheating identification results, and the target detection strategy. In the weighted fusion process, different weight coefficients are set to allocate the contribution of each information. After calculation, a comprehensive score is obtained, and an abnormal behavior determination report is generated based on the comprehensive score.

[0006] Optionally, the target examination room environment map is created according to the candidate seat layout by using a graph neural network algorithm, collaborative cheating behavior is identified by analyzing the abnormal behavior information, and a collaborative cheating identification result is obtained. A reinforcement learning framework based on Q learning is introduced, and the detection strategy is dynamically adjusted according to the abnormal behavior information and the collaborative cheating identification result to obtain a target detection strategy, including: Obtain the candidate seating layout and the examination room video data during the examination to obtain the examination room data information. Based on the examination room data information, use the graph neural network algorithm to construct an initial examination room environment graph, in which each node represents a candidate, and the edge represents the relative position relationship between candidates and the actual degree of interaction between candidates. Dynamically update the edge weights to obtain the target examination room environment graph; Based on the target examination room environment map, the abnormal behavior information is analyzed to identify the collaborative cheating behavior, and a collaborative cheating identification result is obtained. A reinforcement learning framework based on Q learning is introduced, and an adaptive learning rate adjustment mechanism is integrated into the reinforcement learning framework to adjust the learning rate according to the changes in the examination room environment, so as to obtain an optimized collaborative cheating identification model; Based on the optimized collaborative cheating identification model, the detection strategy is dynamically adjusted. The detection strategy includes: increasing the monitoring intensity for areas and candidates with cheating risks higher than a preset threshold in the historical records, shortening the time interval for abnormal behavior detection, and reducing the monitoring frequency for candidates with cheating risks lower than a preset threshold in the historical records, thereby obtaining a target detection strategy.

[0007] Optionally, based on the target examination room environment map, analyzing the abnormal behavior information, identifying the collaborative cheating behavior, obtaining the collaborative cheating identification result, introducing a Q-learning-based reinforcement learning framework, integrating an adaptive learning rate adjustment mechanism in the reinforcement learning framework, adjusting the learning rate according to changes in the examination room environment, and obtaining an optimized collaborative cheating identification model, including: Apply community detection algorithms to identify closely connected groups of examinees and obtain a candidate list of collaborative cheating groups; Based on the candidate list of collaborative cheating groups, an isolation forest detection algorithm is used to screen and determine groups with abnormal behavior patterns to locate collaborative cheating behaviors and obtain collaborative cheating identification results; Based on the collaborative cheating identification results, a Q-learning-based reinforcement learning framework is introduced, and an adaptive learning rate adjustment mechanism is integrated, so that the model can dynamically adjust the learning rate according to the changes in the examination room environment, and the collaborative cheating identification model to be optimized is obtained. The changes in the examination room environment include the examination time period and the distribution density of candidates in the examination room; The cheating behavior data accumulated in historical exams are used as a priori knowledge base, and transfer learning technology is applied to train the collaborative cheating identification model to be optimized to obtain an optimized collaborative cheating identification model.

[0008] Optionally, the cheating behavior data accumulated in historical exams is used as a priori knowledge base, and transfer learning technology is applied to train the collaborative cheating identification model to be optimized to obtain the optimized collaborative cheating identification model, including: Collect cheating behavior data accumulated in historical exams to obtain a priori knowledge base, wherein the cheating behavior data includes confirmed cheating cases, suspected cheating behavior records, and related environmental information; The prior knowledge base is used as source domain data for transfer learning, and the collaborative cheating identification model to be optimized is preliminarily trained by using domain adversarial training technology to optimize the key feature extraction performance of the collaborative cheating identification model to be optimized, thereby obtaining the collaborative cheating identification model being optimized; The fine-tuning technology and the prior knowledge base are used to perform secondary training on the optimized collaborative cheating identification model to accelerate the convergence process of the optimized collaborative cheating identification model. Regularization technology is introduced in the secondary training process to prevent the optimized collaborative cheating identification model from overfitting. The optimized collaborative cheating identification model is obtained after the training.

[0009] Optionally, the extreme gradient boosting learning algorithm is used to weightedly fuse the abnormal behavior information, the collaborative cheating identification result and the target detection strategy, and in the weighted fusion process, different weight coefficients are set to allocate the contribution of each information, and a comprehensive score is obtained after calculation, and an abnormal behavior determination report is generated based on the comprehensive score, including: An extreme gradient boosting learning algorithm is used to perform weighted fusion on the abnormal behavior information, the collaborative cheating identification result, and the target detection strategy. In the weighted fusion process, different weight coefficients are set to allocate the contribution of each information to obtain an initial comprehensive evaluation result; Based on the initial comprehensive evaluation result, the feature importance evaluation function of the extreme gradient boosting learning algorithm is used to adjust the weight coefficients of the abnormal behavior information, the collaborative cheating identification result, and the target detection strategy, and the Shapley value method is introduced to quantify the contribution of each information to obtain an optimized comprehensive evaluation result; Constructing an initial extreme gradient boosting learning model, inputting the optimized comprehensive evaluation result into the initial extreme gradient boosting learning model, and introducing a cross-validation method and a Bayesian optimization technique for training to obtain an optimized extreme gradient boosting learning model; The optimized extreme gradient boosting learning model and multiple learning models are used to predict the cheating behavior data respectively to obtain multiple original prediction results, and the multiple original prediction results are integrated by an ensemble learning method to obtain a fused prediction result. The scoring model is initialized based on the fused prediction result to obtain a preliminary comprehensive scoring model, and an adaptive weighted fusion method is introduced into the preliminary comprehensive scoring model to dynamically adjust the weights of each model according to the performance of the preliminary comprehensive scoring model on different subsets to obtain an optimized comprehensive scoring model. The multiple learning models include a random forest and a support vector machine; A comprehensive score is calculated based on the optimized comprehensive scoring model, and an abnormal behavior determination report is generated according to the comprehensive score. The abnormal behavior determination report includes a description of the abnormal behavior, an assessment of the possibility of cheating, and handling suggestions.

[0010] Optionally, the optimized extreme gradient boosting learning model and multiple learning models are used to predict the cheating behavior data respectively to obtain multiple original prediction results, and the multiple original prediction results are integrated using an integrated learning method to obtain a fused prediction result. The scoring model is initialized based on the fused prediction result to obtain a preliminary comprehensive scoring model, and an adaptive weighted fusion method is introduced into the preliminary comprehensive scoring model to dynamically adjust the weights of each model according to the performance of the preliminary comprehensive scoring model on different subsets to obtain an optimized comprehensive scoring model. The multiple learning models include random forests and support vector machines, including: The optimized extreme gradient boosting learning model and multiple learning models are used as initial base learners; the initial base learners are trained on the cheating behavior data to obtain multiple target base learners, the cheating behavior data is predicted based on the multiple target base learners to obtain multiple prediction results, the multiple prediction results are fused by weighted average to obtain a fused prediction result, and the fused prediction result is used as a meta-feature; Constructing an initial scoring model, using the initial scoring model as a meta-learner, training the meta-learner based on the meta-features and true labels to obtain a preliminary comprehensive scoring model, wherein the true labels include the cheating behavior data and data labels, and the data labels include normal behavior, abnormal behavior, and cheating behavior; The comprehensive behavioral feature set is divided into multiple training sets and validation sets, the performance of the preliminary comprehensive scoring model is evaluated on each validation set to obtain an evaluation performance result, and the prediction errors of the multiple target base learners are recorded, based on the evaluation performance result and the prediction error, the weights of the multiple target base learners are dynamically adjusted using an online gradient descent method to obtain adjusted weights, and the preliminary comprehensive scoring model is trained using the adjusted weights to obtain an optimized comprehensive scoring model.

[0011] Optionally, the application is based on a time series prediction algorithm of a long short-term memory network, and combined with an attention mechanism optimization model, to analyze the behavior pattern on the time dimension of the comprehensive behavior feature set to obtain abnormal behavior information, including: Arranging the data in the comprehensive behavior feature set in chronological order and performing feature extraction to obtain time series data, wherein the time series data includes statistical features within a time window; A time series prediction model based on a long short-term memory network is constructed, and an attention mechanism is introduced into the time series prediction model to optimize the model, thereby obtaining an optimized time series prediction model; An optimized time series prediction model is used to analyze the time series data to obtain analysis results, and the analysis results are processed using an anomaly detection algorithm to obtain behavior anomaly values. A qualified threshold is preset, and when the behavior anomaly value is greater than the qualified threshold, it is judged as abnormal behavior, and the candidate's abnormal behavior information is obtained.

[0012] In a second aspect, an embodiment of the present invention provides a system for analyzing abnormal behavior of examinees in a computer-based examination, comprising: An integration module is used to extract the examinee's facial expression and body posture information from the monitoring image, and use a behavior recognition algorithm to analyze the keyboard tapping frequency and the mouse movement trajectory to obtain behavior information, and integrate the examinee's facial expression, the body posture information and the behavior information to obtain a comprehensive behavior feature set; An analysis module, for applying a time series prediction algorithm based on a long short-term memory network and combining it with an attention mechanism optimization model to analyze the behavior pattern in the time dimension of the comprehensive behavior feature set to obtain abnormal behavior information; The recognition module is used to create a target examination room environment map according to the candidate seat layout through a graph neural network algorithm, identify collaborative cheating behavior by analyzing the abnormal behavior information, obtain collaborative cheating identification results, introduce a Q-learning-based reinforcement learning framework, dynamically adjust the detection strategy according to the abnormal behavior information and the collaborative cheating identification results, and obtain a target detection strategy; The generation module is used to adopt an extreme gradient boosting learning algorithm to weightedly fuse the abnormal behavior information, the collaborative cheating identification result and the target detection strategy, and in the weighted fusion process, the contribution of each information is allocated by setting different weight coefficients, and a comprehensive score is obtained after calculation, and an abnormal behavior determination report is generated based on the comprehensive score.

[0013] In a third aspect, an embodiment of the present invention provides a computing device, comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a method for analyzing abnormal behavior of examinees in a computer-based examination as described in any one of the first aspects.

[0014] In a fourth aspect, an embodiment of the present invention provides a computer storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement a method for analyzing abnormal behavior of examinees in a computer-based examination as described in any one of the first aspects.

[0015] In an embodiment of the present invention, facial expression and body posture information of examinees are extracted from monitoring images, and a behavior recognition algorithm is used to analyze keyboard tapping frequency and mouse movement trajectory to obtain behavior information, and the facial expression of the examinee, the body posture information and the behavior information are integrated to obtain a comprehensive behavior feature set; a time series prediction algorithm based on a long short-term memory network is applied, and in combination with an attention mechanism optimization model, the behavior pattern in the time dimension of the comprehensive behavior feature set is analyzed to obtain abnormal behavior information; a target examination room environment map is created according to the seat layout of the examinees through a graph neural network algorithm, and collaborative cheating behavior is identified by analyzing the abnormal behavior information to obtain a collaborative cheating identification result, a reinforcement learning framework based on Q learning is introduced, and the detection strategy is dynamically adjusted according to the abnormal behavior information and the collaborative cheating identification result to obtain a target detection strategy; an extreme gradient boosting learning algorithm is used to weightedly fuse the abnormal behavior information, the collaborative cheating identification result and the target detection strategy, and in the weighted fusion process, different weight coefficients are set to allocate the contribution of each information, and a comprehensive score is obtained after calculation, and an abnormal behavior determination report is generated based on the comprehensive score. The technical solution provided by the present invention improves the detection capability and efficiency of abnormal behavior and cheating behavior in computer-based examinations, and helps maintain the fairness and impartiality of the examination. Among them, transfer learning and domain adversarial training techniques optimize the key feature extraction capability of the model, making the model perform better in new environments; the application of fine-tuning technology and regularization technology accelerates the convergence process of the model, reduces training time and resource consumption, and improves training efficiency; regularization technology prevents model overfitting, improves the generalization ability and robustness of the model, and ensures that the model still performs well when facing new data; the optimized collaborative cheating identification model can comprehensively and accurately monitor the behavior of candidates, effectively prevent cheating, and ensure the fairness and impartiality of the examination.

[0016] These and other aspects of the present invention will become more apparent from the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A flowchart of a method for analyzing abnormal behavior of examinees in a computer-based examination provided by an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a system for analyzing abnormal behavior of examinees in a computer-based examination provided by an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a computing device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0020] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0022] Figure 1 A flowchart of a method for analyzing abnormal behavior of examinees in a computer-based examination is provided in an embodiment of the present invention, such as Figure 1 As shown, the method includes: In modern education and examination systems, computer-based examinations are widely used due to their convenience and efficiency. However, how to ensure the fairness and impartiality of the examinations, especially how to effectively monitor and identify the abnormal behaviors and cheating behaviors of examinees, has become an important issue. Based on this, the present invention provides a method for analyzing the abnormal behaviors of examinees in computer-based examinations, such as Figure 1 ,include: Step 101: extracting the examinee's facial expression and body posture information from the monitoring image, and using a behavior recognition algorithm to analyze the keyboard tapping frequency and the mouse movement trajectory to obtain behavior information, and integrating the examinee's facial expression, the body posture information and the behavior information to obtain a comprehensive behavior feature set; In this step, facial expressions refer to changes in the candidate's facial expressions, such as smiling, frowning, etc., which can reflect the candidate's emotional state; body posture refers to the candidate's body posture and movements, such as head rotation, hand movement, etc., which can reflect the candidate's behavior pattern; behavior recognition algorithm refers to identifying the candidate's behavior pattern and abnormal behavior by analyzing the keyboard tapping frequency and mouse movement trajectory.

[0023] This step extracts the examinee's facial expression and body posture information from the examinee's face and body images captured by the camera in the examination room in real time. For example, facial expression features are extracted using a facial recognition algorithm, and body posture features are extracted using a posture estimation algorithm; the extracted facial expression and body posture information are stored in a database for subsequent analysis; behavior recognition algorithms, such as hidden Markov models and support vector machines, are used to analyze the keyboard tapping frequency and mouse movement trajectory to identify the examinee's behavior pattern; the identified behavior pattern is stored in a database; multimodal data such as facial expression, body posture information, keyboard tapping frequency, and mouse movement trajectory are fused to form a comprehensive behavior feature set; the comprehensive behavior feature set is standardized to ensure that each feature is compared under the same dimension.

[0024] Step 102: applying a time series prediction algorithm based on a long short-term memory network and combining it with an attention mechanism optimization model to analyze the behavior pattern in the time dimension of the comprehensive behavior feature set to obtain abnormal behavior information; In this step, the time series prediction algorithm refers to predicting future behavior patterns by modeling data in the time dimension. Commonly used methods include the long short-term memory network (LSTM) attention mechanism, which refers to an optimization model that enables the model to focus on the key parts of the time series and improve the model's capture ability.

[0025] This step builds a time series prediction model based on the long short-term memory network (LSTM) and introduces an attention mechanism into the model so that the model can focus on the key parts of the time series; uses a comprehensive behavioral feature set and real labels to train the LSTM model to ensure that the model can learn the behavior patterns of the examinees; uses the LSTM model to analyze the behavior patterns in the time dimension of the comprehensive behavioral feature set and identify the examinees' abnormal behavior patterns; uses anomaly detection algorithms (such as Isolation Forest, One-Class SVM, etc.) to analyze the prediction results and identify abnormal behaviors. Set a suitable threshold, and when the probability of abnormal behavior predicted by the model exceeds the threshold, it is judged as abnormal behavior.

[0026] Step 103: Create a target examination room environment map according to the candidate seat layout through a graph neural network algorithm, identify collaborative cheating behavior by analyzing the abnormal behavior information, obtain collaborative cheating identification results, introduce a Q-learning-based reinforcement learning framework, dynamically adjust the detection strategy according to the abnormal behavior information and the collaborative cheating identification results, and obtain a target detection strategy; In this step, the graph neural network algorithm refers to the analysis of the relationship between nodes (candidates) by constructing a graph structure to identify collaborative cheating behavior; the target examination environment graph refers to a graph structure created based on the candidate seat layout, where each node represents a candidate and the edges represent the relative position relationship between candidates.

[0027] This step creates a graph structure based on the seat layout of the candidates, where each node represents a candidate and the edge represents the relative position relationship between the candidates; uses the graph neural network (GNN) algorithm to analyze the nodes and edges in the graph structure and identify the interactions between the candidates; by analyzing the abnormal behavior information of the candidates, possible collaborative cheating behaviors are identified and collaborative cheating identification results are generated; introduces a reinforcement learning framework based on Q-learning to dynamically adjust the detection strategy based on abnormal behavior information and collaborative cheating identification results; through continuous learning and optimization, the detection strategy can better adapt to different examination scenarios and improve the accuracy and efficiency of detection.

[0028] Step 104: using an extreme gradient boosting learning algorithm to weightedly fuse the abnormal behavior information, the collaborative cheating identification result, and the target detection strategy, in the weighted fusion process, different weight coefficients are set to allocate the contribution of each information, a comprehensive score is obtained after calculation, and an abnormal behavior determination report is generated based on the comprehensive score; In this step, the extreme gradient boosting learning algorithm (XGBoost) refers to an ensemble learning method that improves the accuracy and robustness of the model by weighted fusion of the prediction results of multiple models.

[0029] This step performs weighted fusion of abnormal behavior information, collaborative cheating identification results, and target detection strategies, and allocates the contribution of each information by setting different weight coefficients; uses the extreme gradient boosting learning algorithm to build a fusion model to ensure that the model can comprehensively consider multiple information; calculates a comprehensive score through the fusion model to reflect the severity of the examinee's abnormal behavior and cheating behavior; generates an abnormal behavior determination report based on the comprehensive score, recording the examinee's abnormal behavior and cheating behavior in detail.

[0030] The embodiments of the present invention can achieve the following beneficial effects through the above steps: By integrating the behavioral feature set, the accuracy and comprehensiveness of behavior recognition are improved; The time series prediction algorithm based on the long short-term memory network (LSTM) combined with the attention mechanism can effectively capture the changing trends and patterns of examinees' behaviors over time, and improve the ability to identify abnormal behaviors; The target examination room environment map constructed by the graph neural network (GNN) algorithm can simulate the spatial relationship between candidates in the examination room. By analyzing these relationships and the abnormal behavior information of candidates, collaborative cheating behavior can be effectively identified; The reinforcement learning framework can dynamically adjust the detection strategy, improving the adaptability and accuracy of the detection strategy; The comprehensive score can objectively reflect the behavior of the examinees and provide a solid foundation for generating abnormal behavior judgment reports.

[0031] In order to effectively identify and prevent collaborative cheating in examinations and improve the recognition accuracy of cheating, the present invention provides a specific embodiment. In step 103, a target examination room environment map is created according to the seat layout of the examinees through a graph neural network algorithm, and collaborative cheating behavior is identified by analyzing the abnormal behavior information to obtain a collaborative cheating recognition result. A reinforcement learning framework based on Q learning is introduced, and the detection strategy is dynamically adjusted according to the abnormal behavior information and the collaborative cheating recognition result to obtain a target detection strategy. The specific steps include: Step 310: Obtain the candidate seat layout and the examination room video data during the examination to obtain the examination room data information, and based on the examination room data information, use the graph neural network algorithm to construct an initial examination room environment graph, in which each node represents a candidate, and the edge represents the relative position relationship between the candidates and the actual degree of interaction between the candidates, and dynamically update the edge weight to obtain the target examination room environment graph; In this step, the graph neural network algorithm refers to a deep learning algorithm used to process graph structured data and can analyze the relationship between nodes.

[0032] This step obtains the candidate seating layout and the examination room video materials taken by the camera in the examination room during the examination; integrates the candidate seating layout and the examination room video materials during the examination to form the examination room data information; uses the graph neural network algorithm to construct the initial examination room environment map based on the candidate seating layout, where each node represents a candidate and the edge represents the relative position relationship between candidates; calculates the actual degree of interaction between candidates by analyzing the examination room video materials. For example, the interaction between candidates can be judged by information such as facial orientation and line of sight; dynamically updates the edge weights based on the actual degree of interaction of the candidates to obtain the target examination room environment map.

[0033] Step 311: Based on the target examination room environment map, the abnormal behavior information is analyzed to identify the collaborative cheating behavior, and a collaborative cheating identification result is obtained. A reinforcement learning framework based on Q learning is introduced, and an adaptive learning rate adjustment mechanism is integrated into the reinforcement learning framework. The learning rate is adjusted according to the changes in the examination room environment to obtain an optimized collaborative cheating identification model. In this step, the Q-learning-based reinforcement learning framework refers to a reinforcement learning method that enables the model to self-optimize according to the current abnormal behavior information and collaborative cheating identification results by dynamically adjusting the strategy.

[0034] This step is based on the target examination room environment map, and uses the graph neural network algorithm to analyze the interactive relationship between candidates, identify collaborative cheating behavior, and obtain collaborative cheating identification results; introduces a reinforcement learning framework based on Q-learning to dynamically adjust the detection strategy according to abnormal behavior information and collaborative cheating identification results; integrates an adaptive learning rate adjustment mechanism in the framework to adjust the learning rate according to changes in the examination room environment; through the adaptive learning rate adjustment mechanism, an optimized collaborative cheating identification model is obtained to improve the adaptability and accuracy of the model.

[0035] Step 312: Based on the optimized collaborative cheating identification model, dynamically adjust the detection strategy, the detection strategy includes: increasing the monitoring intensity for the areas and examinees whose cheating risks in the historical records are higher than the preset threshold, shortening the time interval for abnormal behavior detection, and reducing the monitoring frequency for examinees whose cheating risks in the historical records are lower than the preset threshold, to obtain a target detection strategy; In this step, the preset threshold refers to a pre-set standard value used to determine the cheating risk of the examinee.

[0036] This step assesses the cheating risk of each candidate and region based on historical records. A threshold is preset , when the cheating risk of candidates or regions is higher than When the test is detected, it is marked as high risk; for high-risk areas and candidates, the monitoring intensity is increased and the time interval for abnormal behavior detection is shortened. For low-risk areas and candidates, the monitoring frequency is reduced to save resources; based on the above evaluation results, a target detection strategy is generated to ensure the effective use of invigilation resources.

[0037] The embodiments of the present invention can achieve the following beneficial effects through the above steps: By building a target examination room environment map through a graph neural network algorithm, we can more accurately analyze the interactive relationship between candidates and improve the accuracy of collaborative cheating identification. The introduction of a Q-learning-based reinforcement learning framework and an adaptive learning rate adjustment mechanism enables the model to dynamically adjust the detection strategy according to changes in the examination environment, improving the adaptability and robustness of the model. By dynamically adjusting the detection strategy, we increased monitoring efforts for high-risk areas and candidates, and reduced monitoring frequency for low-risk areas and candidates, effectively utilizing invigilation resources and improving invigilation efficiency; Through a variety of methods, the behavior of candidates is monitored comprehensively and accurately, effectively preventing cheating and ensuring the fairness and impartiality of the examination.

[0038] In order to improve the recognition accuracy and adaptive ability of the collaborative cheating identification model and ensure the fairness and impartiality of the examination, the present invention provides a specific embodiment, in which the step 311 analyzes the abnormal behavior information based on the target examination room environment map, identifies the collaborative cheating behavior, obtains the collaborative cheating identification result, introduces a Q-learning-based reinforcement learning framework, integrates an adaptive learning rate adjustment mechanism in the reinforcement learning framework, adjusts the learning rate according to the changes in the examination room environment, and obtains an optimized collaborative cheating identification model, which specifically includes the following steps: Step 321: applying a community detection algorithm to identify closely connected groups among examinees, and obtaining a candidate list of collaborative cheating groups; In this step, community detection algorithm refers to an algorithm used to identify groups of closely connected nodes in a graph, which is often used in social network analysis.

[0039] This step uses community detection algorithms, such as the Louvain algorithm and the Girvan-Newman algorithm, to identify closely connected groups among candidates, which may be potential collaborative cheating groups; the identified closely connected groups are included in the candidate list of collaborative cheating groups.

[0040] Step 322: Based on the candidate list of collaborative cheating groups, an isolation forest detection algorithm is used to screen and determine the groups with abnormal behavior patterns to locate the collaborative cheating behavior and obtain collaborative cheating identification results; In this step, the isolation forest detection algorithm refers to an unsupervised anomaly detection algorithm that identifies anomalies by building an isolation tree.

[0041] This step extracts the behavior pattern data of each group in the candidate list of collaborative cheating groups; uses the isolation forest detection algorithm to analyze the behavior pattern of each group and identify groups with abnormal behavior patterns; marks groups with abnormal behavior patterns as collaborative cheating groups and generates collaborative cheating identification results.

[0042] Step 323: Based on the collaborative cheating identification result, a Q-learning-based reinforcement learning framework is introduced, and an adaptive learning rate adjustment mechanism is integrated to enable the model to dynamically adjust the learning rate according to changes in the examination room environment, thereby obtaining a collaborative cheating identification model to be optimized. The changes in the examination room environment include the examination time period and the distribution density of examinees in the examination room. This step uses the collaborative cheating identification results as input to build a reinforcement learning framework based on Q-learning; integrates an adaptive learning rate adjustment mechanism into the reinforcement learning framework to dynamically adjust the learning rate according to changes in the examination environment; uses the reinforcement learning framework to train the model, dynamically adjusts the learning strategy according to changes in the examination environment, such as the examination time period and the distribution density of candidates, and obtains the collaborative cheating identification model to be optimized.

[0043] Step 324: using the cheating behavior data accumulated in the past exams as a priori knowledge base, applying transfer learning technology, training the collaborative cheating identification model to be optimized, and obtaining an optimized collaborative cheating identification model; In this step, transfer learning technology refers to a machine learning method that improves the performance of the target domain model by leveraging source domain data.

[0044] This step collects cheating behavior data accumulated in historical exams, including confirmed cheating cases, suspected cheating behavior records and related environmental information, to build a prior knowledge base; the prior knowledge base is used as source domain data, and transfer learning techniques, such as domain adversarial training and feature alignment, are applied to preliminarily train the collaborative cheating identification model to be optimized; through transfer learning technology, the rich information in the prior knowledge base is utilized to optimize the key feature extraction performance of the model, improve the accuracy and robustness of the model, and obtain the optimized collaborative cheating identification model.

[0045] The embodiments of the present invention can achieve the following beneficial effects through the above steps: The community detection algorithm and the isolated forest detection algorithm can more accurately identify closely connected groups of examinees and groups with abnormal behavior patterns, effectively locate collaborative cheating behaviors, and improve the accuracy of collaborative cheating identification; The introduction of a Q-learning-based reinforcement learning framework and an adaptive learning rate adjustment mechanism enables the model to dynamically adjust its learning strategy according to changes in the examination environment, thereby improving the adaptability and robustness of the model.

[0046] Through transfer learning technology, the cheating behavior data accumulated in historical exams is used as a priori knowledge base to optimize the key feature extraction performance of the model, thereby improving the accuracy and generalization ability of the model.

[0047] Through a variety of advanced technologies, it is possible to comprehensively and accurately monitor the behavior of candidates, effectively prevent cheating, and ensure the fairness and impartiality of the examination.

[0048] In order to further optimize the collaborative cheating identification model and improve the model performance, based on this, the present invention provides a specific embodiment, in which the step 324 uses the cheating behavior data accumulated in historical exams as a priori knowledge base, applies transfer learning technology, trains the collaborative cheating identification model to be optimized, and obtains the optimized collaborative cheating identification model, which specifically includes the following steps: Step 331: Collect cheating behavior data accumulated in historical exams to obtain a priori knowledge base, wherein the cheating behavior data includes confirmed cheating cases, suspected cheating behavior records, and related environmental information; This step collects confirmed cheating cases, suspected cheating behavior records and related environmental information from historical exams. These data should include multimodal information such as the examinee's behavioral characteristics, exam environment, and surveillance video. The collected data is organized into a structured format to build a priori knowledge base. For example, the data can be stored in a database, where each record includes fields such as examinee ID, behavior type, behavior description, and environmental information.

[0049] Step 332: using the prior knowledge base as source domain data for transfer learning, and using domain adversarial training technology to preliminarily train the collaborative cheating identification model to be optimized, so as to optimize the key feature extraction performance of the collaborative cheating identification model to be optimized, and obtain the collaborative cheating identification model being optimized; In this step, domain adversarial training technology refers to a transfer learning method that uses adversarial training to enable the model to better adapt to the target domain data and improve the generalization ability of the model.

[0050] In this step, the data in the prior knowledge base is used as the source domain data, and the target domain data is the data of the current exam; the domain adversarial training technology is used to preliminarily train the collaborative cheating identification model to be optimized, and the feature extractor G and the classifier C are optimized by minimizing the classification loss and maximizing the domain discrimination loss to obtain the collaborative cheating identification model being optimized.

[0051] More specifically, the embodiment of the present invention provides a calculation formula for the classification loss function and the domain discrimination loss function in adversarial training. The two functions are combined through adversarial training technology to form a total adversarial loss function to optimize the key feature extraction performance of the collaborative cheating identification model to be optimized. The specific calculation formula is as follows: ; ; ; ; in, is the total adversarial loss function; is a classification loss function, which is used to measure the classification performance of the collaborative cheating identification model to be optimized on the source domain data; It is a domain discrimination loss function, which is used to measure the performance of the model in distinguishing source domain and target domain data; is a feature extractor; is a classifier; is a domain discriminator; is the source domain dataset; is the target domain dataset; is the weight of adversarial training; Representation feature extractor For input samples Output after feature extraction; is the input sample The true label of is a loss function, which is a cross entropy loss function in this embodiment; is the field label; It is the influencing parameter of the domain discrimination loss function.

[0052] Through the combination of classification loss function, domain discrimination loss function and total adversarial loss function, the collaborative cheating identification model to be optimized can more accurately identify and prevent cheating in a changing examination environment, ensuring the fairness and impartiality of the examination. By minimizing the classification loss, it can be ensured that the model can accurately identify and classify cheating in the source domain; by maximizing the domain discrimination loss, the feature extractor can be optimized. The extracted features have similar distribution between the source domain data and the target domain data, thereby improving the generalization ability of the model; the total adversarial loss function combines the classification loss function and the domain discrimination loss function. Through adversarial training, the model can maintain the source domain classification performance while improving its adaptability to the target domain data.

[0053] Step 333: Apply the fine-tuning technology and the prior knowledge base to perform secondary training on the optimized collaborative cheating identification model to accelerate the convergence process of the optimized collaborative cheating identification model. In the secondary training process, a regularization technology is introduced to prevent the optimized collaborative cheating identification model from overfitting. After the training, the optimized collaborative cheating identification model is obtained. In this step, fine-tuning technology refers to further training the model using target domain data based on preliminary training to improve the performance of the model; regularization technology refers to preventing overfitting of the model and improving the generalization ability of the model by introducing regularization terms.

[0054] This step uses the source domain data as training data and the target domain data as verification data to fine-tune the collaborative cheating identification model being optimized. Regularization technology is used to prevent model overfitting and accelerate the model convergence process, ultimately obtaining the optimized collaborative cheating identification model.

[0055] More specifically, the embodiment of the present invention provides an expression of a regularization term and a classification loss function of the collaborative cheating identification model being optimized, and combines the two to form an optimization objective function to fine-tune the collaborative cheating identification model being optimized, prevent overfitting, and accelerate the convergence process of the model. The specific calculation formula is as follows: ; ; ; in, is the classification loss function used to measure the collaborative cheating identification model in optimization Classification performance on the source domain dataset; is a regularization term used to prevent collaborative cheating identification models in optimization Overfitting; is the optimization objective function in the fine-tuning process; represents the model after preliminary training, which is the collaborative cheating identification model being optimized in this embodiment; is the source domain dataset; It is a collaborative cheating identification model in optimization For input samples The predicted output of is the input sample The true label of is a loss function, which is a cross entropy loss function in this embodiment; It is a collaborative cheating identification model in optimization The weight in All weights The square norm of ; is the regularization coefficient.

[0056] Through the objective optimization function, the classification performance and model complexity are balanced during the fine-tuning process to ensure that the model has high accuracy and generalization ability on the target domain data; The classification performance of the collaborative cheating identification model in the optimization is ensured on the source domain data; the regularization term controls the complexity of the model through the regularization coefficient to prevent overfitting; ,balance The relationship between the regularization term and the classification term ensures that the model maintains classification performance and improves generalization ability during the optimization process.

[0057] The embodiments of the present invention can achieve the following beneficial effects through the above steps: Through transfer learning and domain adversarial training technology, we use the cheating behavior data accumulated in historical exams as a priori knowledge base to optimize the key feature extraction performance of the model, so that the model can better adapt to the new exam environment; Apply fine-tuning and regularization techniques to accelerate the convergence process of the model, reduce training time and resource consumption, and improve the training efficiency of the model; By introducing regularization technology, the model is prevented from overfitting during training, thus improving the generalization ability and robustness of the model. The collaborative cheating identification model, which has been optimized through various technical training, can comprehensively and accurately monitor the behavior of candidates, effectively prevent cheating, and ensure the fairness and impartiality of the examination.

[0058] In order to comprehensively and accurately identify and evaluate abnormal behaviors in the examination and improve the accuracy and reliability of identification, based on this, the present invention provides a specific embodiment, in which the step 104 adopts an extreme gradient boosting learning algorithm to weightedly fuse the abnormal behavior information, the collaborative cheating identification result and the target detection strategy, and in the weighted fusion process, the contribution of each information is allocated by setting different weight coefficients, and a comprehensive score is obtained after calculation, and an abnormal behavior determination report is generated based on the comprehensive score, which specifically includes the following steps: Step 401: using an extreme gradient boosting learning algorithm to perform weighted fusion on the abnormal behavior information, the collaborative cheating identification result, and the target detection strategy, and in the weighted fusion process, different weight coefficients are set to allocate the contribution of each information to obtain an initial comprehensive evaluation result; This step uses the extreme gradient boosting algorithm to perform weighted fusion of abnormal behavior information, collaborative cheating identification results, and target detection strategies. For example, different weight coefficients H1, H2, and H3 are set to represent the contribution of abnormal behavior information, collaborative cheating identification results, and target detection strategies, respectively. These information are fused to obtain the initial comprehensive evaluation results.

[0059] Step 402: Based on the initial comprehensive evaluation result, the feature importance evaluation function of the extreme gradient boosting learning algorithm is used to adjust the weight coefficients of the abnormal behavior information, the collaborative cheating identification result, and the target detection strategy, and the Shapley value method is introduced to quantify the contribution of each information to obtain an optimized comprehensive evaluation result; In this step, feature importance evaluation refers to a method provided by the extreme gradient boosting learning algorithm to evaluate the contribution of features to the model prediction performance; the Shapley value method refers to a method for quantifying feature contributions by calculating the marginal contribution of each feature in all possible feature combinations.

[0060] This step uses the feature importance evaluation function of the extreme gradient boosting learning algorithm to calculate the importance score of each feature, namely abnormal behavior information, collaborative cheating identification results, and target detection strategy; adjust the weight coefficients H1, H2, and H3 according to the feature importance scores; introduce the Shapley value method to quantify the contribution of each information to the comprehensive evaluation results; the Shapley value method calculates the marginal contribution of each feature in all possible feature combinations to obtain the final contribution value of each feature; recalculate the comprehensive evaluation results based on the adjusted weight coefficients and Shapley values ​​to obtain optimized comprehensive evaluation results.

[0061] Step 403: constructing an initial extreme gradient boosting learning model, inputting the optimized comprehensive evaluation result into the initial extreme gradient boosting learning model, and introducing a cross-validation method and a Bayesian optimization technique for training to obtain an optimized extreme gradient boosting learning model; In this step, the cross-validation method refers to dividing the dataset into multiple subsets, training and validating the model multiple times to evaluate the stability and generalization ability of the model; Bayesian optimization technology refers to a method for optimizing hyperparameters, which gradually approaches the optimal solution by constructing a proxy model of the objective function.

[0062] This step builds the initial extreme gradient boosting learning model, takes the optimization comprehensive evaluation results as input, and trains the model; uses the K-fold cross-validation method to divide the data set into K subsets, uses K-1 subsets for training each time, and the remaining subset for validation, repeating K times to ensure the stability and generalization ability of the model; uses Bayesian optimization technology to automatically search for the best hyperparameter combination, including learning rate, maximum depth, number of trees, etc.; uses Bayesian optimization to find the optimal hyperparameter settings and improve the performance of the model; and obtains the optimized XGBoost model based on the results of cross-validation and Bayesian optimization.

[0063] Step 404: using the optimized extreme gradient boosting learning model and multiple learning models to predict the cheating behavior data respectively, obtaining multiple original prediction results, integrating the multiple original prediction results using an ensemble learning method to obtain a fused prediction result, training an initial scoring model based on the fused prediction result to obtain a preliminary comprehensive scoring model, introducing an adaptive weighted fusion method into the preliminary comprehensive scoring model, dynamically adjusting the weights of each model according to the performance of the preliminary comprehensive scoring model on different subsets, and obtaining an optimized comprehensive scoring model, wherein the multiple learning models include a random forest and a support vector machine; In this step, the integrated learning method refers to improving the accuracy and stability of the overall prediction by combining the prediction results of multiple learning models; the adaptive weighted fusion method refers to dynamically adjusting the weights of each model according to the performance of the model on different subsets to obtain an optimized comprehensive scoring model.

[0064] In this step, a preliminary comprehensive scoring model is obtained through a variety of techniques, and the model is further optimized to integrate the advantages of different models to obtain an optimized comprehensive scoring model.

[0065] Step 405: Calculate a comprehensive score based on the optimized comprehensive scoring model, and generate an abnormal behavior determination report according to the comprehensive score, wherein the abnormal behavior determination report includes a description of the abnormal behavior, an assessment of the possibility of cheating, and handling suggestions; This step uses an optimized comprehensive scoring model to calculate the comprehensive score for each candidate and generate a report based on it that contains a description of abnormal behavior, an assessment of the possibility of cheating, and handling recommendations.

[0066] The embodiments of the present invention can achieve the following beneficial effects through the above steps: The key feature extraction performance of the model is optimized through the extreme gradient boosting learning algorithm; The Shapley value method is introduced to quantify the contribution of each piece of information, ensuring that the model's decision is more reasonable and reliable; Cross-validation and Bayesian optimization techniques are used to ensure that the model performs stably on different subsets, thus improving the robustness of the model. The adaptive weighted fusion method dynamically adjusts the weights of each model to further optimize the performance of the model; Through the integrated learning method of multiple learning models, the advantages of different models are integrated to improve the accuracy of cheating behavior identification; Generates a detailed abnormal behavior determination report, providing a comprehensive description of abnormal behavior, cheating possibility assessment and handling suggestions, which helps the examination management agency take timely measures to maintain the fairness and impartiality of the examination.

[0067] In order to improve the accuracy and reliability of cheating behavior identification, based on this, the present invention provides a specific embodiment, in which the step 404 uses the optimized extreme gradient boosting learning model and multiple learning models to predict the cheating behavior data respectively to obtain multiple original prediction results, uses an integrated learning method to integrate the multiple original prediction results to obtain a fused prediction result, trains an initial scoring model based on the fused prediction result to obtain a preliminary comprehensive scoring model, introduces an adaptive weighted fusion method into the preliminary comprehensive scoring model, dynamically adjusts the weights of each model according to the performance of the preliminary comprehensive scoring model on different subsets, and obtains an optimized comprehensive scoring model, wherein the multiple learning models include a random forest and a support vector machine, and specifically includes the following steps: Step 411: using the optimized extreme gradient boosting learning model and multiple learning models as initial base learners; respectively training the initial base learners on the cheating behavior data to obtain multiple target base learners, predicting the cheating behavior data based on the multiple target base learners to obtain multiple prediction results, fusing the multiple prediction results in a weighted average manner to obtain a fused prediction result, and using the fused prediction result as a meta-feature; In this step, meta features refer to the fused prediction results, which are used as higher-level feature inputs to train the meta learner and improve the overall performance of the model; meta learner refers to a model used to learn and integrate the prediction results of multiple base learners.

[0068] This step will optimize the extreme gradient boosting learning model, random forest and support vector machine as the initial base learners, and ensure that these models have been preliminarily optimized, including hyperparameter adjustment and cross-validation; use the cheating behavior data to train the initial base learners respectively; through training, these models become target base learners and have the ability to predict the cheating behavior data; use the trained target base learners to predict the cheating behavior data and generate multiple prediction results; each base learner generates a prediction result to form a set of multiple prediction results; multiple prediction results are fused by weighted averaging; set initial weights, such as equal weights or different weights according to the performance of the model; the fused results are used as meta-features for subsequent meta-learner training.

[0069] Step 412: constructing an initial scoring model, using the initial scoring model as a meta-learner, training the meta-learner based on the meta-features and true labels to obtain a preliminary comprehensive scoring model, wherein the true labels include the cheating behavior data and data labels, and the data labels include normal behavior, abnormal behavior, and cheating behavior; In this step, we construct an initial scoring model and select a suitable algorithm, such as linear regression, logistic regression, etc. We use the fused prediction results as meta-features and the true labels of the cheating behavior data as target variables. We use the meta-features and true labels to train the initial scoring model and obtain a preliminary comprehensive scoring model.

[0070] Step 413: Divide the comprehensive behavior feature set into multiple training sets and validation sets, evaluate the performance of the preliminary comprehensive scoring model on each validation set to obtain an evaluation performance result, and record the prediction errors of the multiple target base learners, and based on the evaluation performance result and the prediction error, dynamically adjust the weights of the multiple target base learners using an online gradient descent method to obtain adjusted weights, and use the adjusted weights to train the preliminary comprehensive scoring model to obtain an optimized comprehensive scoring model; In this step, Online Gradient Descent refers to an optimization algorithm that iteratively updates model parameters to minimize the loss function.

[0071] In this step, the comprehensive behavioral feature set is divided into multiple training sets and validation sets, for example, using K-fold cross validation; the performance of the preliminary comprehensive scoring model is evaluated on each validation set, and evaluation indicators such as accuracy, recall, F1 score, etc. are recorded to obtain the evaluation performance results; the prediction error of each target base learner on the validation set is recorded; according to the evaluation performance results and prediction error, the weights of each base learner are dynamically adjusted using the online gradient descent method; the loss function of the model is minimized by iteratively updating the weights to improve the comprehensive performance of the model; the preliminary comprehensive scoring model is retrained using the adjusted weights to obtain an optimized comprehensive scoring model.

[0072] The embodiments of the present invention can achieve the following beneficial effects through the above steps: Through the ensemble learning method of multiple base learners, the advantages of different models are integrated to improve the prediction accuracy of the model; Dynamically adjust the weights of the base learners to ensure that the model performs more stably and reliably on different subsets; Cross-validation and online gradient descent methods are used to ensure the stability of the model's performance on different subsets, thus improving the robustness of the model. By dynamically adjusting weights, the performance of the model is further optimized and the risk of overfitting is reduced; Through the learning and integration of the meta-learner, the model's recognition accuracy of cheating behavior is improved.

[0073] In order to effectively identify and analyze abnormal behaviors in the examination, based on this, the present invention provides a specific embodiment, in which the step 102 applies a time series prediction algorithm based on a long short-term memory network and combines it with an attention mechanism optimization model to analyze the behavior pattern on the time dimension of the comprehensive behavior feature set to obtain abnormal behavior information: Step 201: Arrange the data in the comprehensive behavior feature set in chronological order and perform feature extraction to obtain time series data, wherein the time series data includes statistical features within a time window; In this step, time series data refers to data arranged in chronological order. Each data point contains statistical features within a certain time window and is used as the input of the time series prediction model to capture the temporal variation pattern of behavior.

[0074] This step arranges the data in the comprehensive behavior feature set in chronological order, extracts features from the data in each time window, and the extracted features include statistical features such as mean, variance, maximum value, minimum value, etc.; the extracted features are organized into time series data.

[0075] Step 202: constructing a time series prediction model based on a long short-term memory network, introducing an attention mechanism into the time series prediction model to optimize the model, and obtaining an optimized time series prediction model; This step builds a time series prediction model based on the long short-term memory network. The attention mechanism is introduced into the model so that the model can focus on the key parts of the input sequence. Through training, the parameters of the model are optimized to improve the model's prediction and generalization capabilities.

[0076] Step 203: Analyze the time series data using the optimized time series prediction model to obtain analysis results, process the analysis results using an anomaly detection algorithm to obtain a behavior anomaly value, preset a qualified threshold, and when the behavior anomaly value is greater than the qualified threshold, determine it as abnormal behavior, and obtain the examinee's abnormal behavior information; This step uses an optimized time series prediction model to analyze the time series data and generate analysis results, which reflect the temporal changes in the examinee's behavior; selects a suitable anomaly detection algorithm, such as isolation forest, LOF, etc., to process the analysis results and identify behavioral anomalies; presets a qualified threshold, which is determined based on historical data and experience; when the behavioral anomaly value is greater than the qualified threshold, it is judged as abnormal behavior; records the examinee's abnormal behavior information, including abnormal behavior descriptions and related information.

[0077] The embodiments of the present invention can achieve the following beneficial effects through the above steps: Through feature extraction, the original data is converted into a form that can be processed by the model, which improves the input quality of the model; The optimized time series prediction model with the introduction of attention mechanism can better capture the key information in time series data and improve the prediction and generalization capabilities of the model; Using time series data and optimizing time series forecasting models can effectively capture the temporal variation of behaviors and improve the robustness of the model; Anomaly detection algorithms can accurately identify abnormal behaviors, reducing false positives and false negatives; Improved the accuracy of identifying cheating behaviors through optimized time series prediction models and anomaly detection algorithms; Preset qualified thresholds ensure that only significant abnormal behaviors are marked, reducing the possibility of misjudgment.

[0078] Figure 2A schematic diagram of a big data processing system for implementing hybrid data analysis is provided for an embodiment of the present invention. Figure 2 As shown, the system includes: The integration module 21 is used to extract the examinee's facial expression and body posture information from the monitoring image, and use a behavior recognition algorithm to analyze the keyboard tapping frequency and the mouse movement trajectory to obtain behavior information, and integrate the examinee's facial expression, the body posture information and the behavior information to obtain a comprehensive behavior feature set; An analysis module 22 is used to apply a time series prediction algorithm based on a long short-term memory network and combine it with an attention mechanism optimization model to analyze the behavior pattern in the time dimension of the comprehensive behavior feature set to obtain abnormal behavior information; The identification module 23 is used to create a target examination room environment map according to the candidate seat layout through a graph neural network algorithm, identify the collaborative cheating behavior by analyzing the abnormal behavior information, obtain the collaborative cheating identification result, introduce a reinforcement learning framework based on Q learning, dynamically adjust the detection strategy according to the abnormal behavior information and the collaborative cheating identification result, and obtain the target detection strategy; The generation module 24 is used to adopt an extreme gradient boosting learning algorithm to weightedly fuse the abnormal behavior information, the collaborative cheating identification result and the target detection strategy, and in the weighted fusion process, the contribution of each information is allocated by setting different weight coefficients, and a comprehensive score is obtained after calculation, and an abnormal behavior determination report is generated based on the comprehensive score.

[0079] Figure 2 The system for analyzing abnormal behavior of examinees in a computer-based examination can be executed Figure 1 The implementation principle and technical effect of the method for analyzing abnormal behavior of candidates in a computer-based test described in the embodiment are not described in detail. The specific way in which each module and unit performs operations in the system for analyzing abnormal behavior of candidates in a computer-based test in the above embodiment has been described in detail in the embodiment of the method, and will not be described in detail here.

[0080] In one possible design, Figure 2 The system for analyzing abnormal behavior of examinees in a computer-based examination in the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32; The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0081] The processing component 32 is used to: extract the examinee's facial expression and body posture information from the monitoring image, and use the behavior recognition algorithm to analyze the keyboard tapping frequency and the mouse movement trajectory to obtain the behavior information, integrate the examinee's facial expression, the body posture information and the behavior information to obtain a comprehensive behavior feature set; apply the time series prediction algorithm based on the long short-term memory network, and combine it with the attention mechanism optimization model to analyze the behavior pattern in the time dimension of the comprehensive behavior feature set to obtain abnormal behavior information; create a target examination room environment map according to the examinee's seat layout through the graph neural network algorithm, identify the collaborative cheating behavior by analyzing the abnormal behavior information, obtain the collaborative cheating identification result, introduce the reinforcement learning framework based on Q learning, dynamically adjust the detection strategy according to the abnormal behavior information and the collaborative cheating identification result, and obtain the target detection strategy; use the extreme gradient boosting learning algorithm to weightedly fuse the abnormal behavior information, the collaborative cheating identification result and the target detection strategy, and in the weighted fusion process, set different weight coefficients to allocate the contribution of each information, obtain a comprehensive score after calculation, and generate an abnormal behavior determination report based on the comprehensive score.

[0082] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0083] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0084] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0085] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0086] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0087] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0088] The embodiment of the present invention further provides a computer storage medium storing a computer program, which can achieve the above-mentioned Figure 1 The illustrated embodiment is a method for analyzing abnormal behavior of examinees in a computer-based examination.

[0089] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0090] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0091] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by software plus a necessary general hardware platform, or by hardware. Based on this understanding, the above technical solution, in essence or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for analyzing abnormal behavior of examinees in a computer-based examination, characterized in that: include: Extracting the examinee's facial expression and body posture information from the monitoring image, and using a behavior recognition algorithm to analyze the keyboard tapping frequency and the mouse movement trajectory to obtain behavior information, integrating the examinee's facial expression, the body posture information and the behavior information to obtain a comprehensive behavior feature set; Applying a time series prediction algorithm based on a long short-term memory network and combining it with an attention mechanism optimization model, the behavior pattern on the time dimension of the comprehensive behavior feature set is analyzed to obtain abnormal behavior information; A target examination room environment map is created based on the candidate seat layout through a graph neural network algorithm, collaborative cheating behavior is identified by analyzing the abnormal behavior information, and collaborative cheating identification results are obtained. A reinforcement learning framework based on Q learning is introduced, and the detection strategy is dynamically adjusted according to the abnormal behavior information and the collaborative cheating identification results to obtain a target detection strategy; An extreme gradient boosting learning algorithm is used to weightedly fuse the abnormal behavior information, the collaborative cheating identification results, and the target detection strategy. In the weighted fusion process, different weight coefficients are set to allocate the contribution of each information. After calculation, a comprehensive score is obtained, and an abnormal behavior determination report is generated based on the comprehensive score.

2. The method according to claim 1, characterized in that: The target examination room environment map is created according to the candidate seat layout through the graph neural network algorithm, the collaborative cheating behavior is identified by analyzing the abnormal behavior information, and the collaborative cheating identification result is obtained. The reinforcement learning framework based on Q learning is introduced, and the detection strategy is dynamically adjusted according to the abnormal behavior information and the collaborative cheating identification result to obtain the target detection strategy, including: Obtain the candidate seating layout and the examination room video data during the examination to obtain the examination room data information. Based on the examination room data information, use the graph neural network algorithm to construct an initial examination room environment graph, in which each node represents a candidate, and the edge represents the relative position relationship between candidates and the actual degree of interaction between candidates. Dynamically update the edge weights to obtain the target examination room environment graph; Based on the target examination room environment map, the abnormal behavior information is analyzed to identify the collaborative cheating behavior, and a collaborative cheating identification result is obtained. A reinforcement learning framework based on Q learning is introduced, and an adaptive learning rate adjustment mechanism is integrated into the reinforcement learning framework to adjust the learning rate according to the changes in the examination room environment, so as to obtain an optimized collaborative cheating identification model; Based on the optimized collaborative cheating identification model, the detection strategy is dynamically adjusted. The detection strategy includes: increasing the monitoring intensity for areas and candidates with cheating risks higher than a preset threshold in the historical records, shortening the time interval for abnormal behavior detection, and reducing the monitoring frequency for candidates with cheating risks lower than a preset threshold in the historical records, thereby obtaining a target detection strategy.

3. The method according to claim 2, characterized in that Based on the target examination room environment map, the abnormal behavior information is analyzed, the collaborative cheating behavior is identified, and the collaborative cheating identification result is obtained. A reinforcement learning framework based on Q learning is introduced, and an adaptive learning rate adjustment mechanism is integrated in the reinforcement learning framework. The learning rate is adjusted according to the changes in the examination room environment to obtain an optimized collaborative cheating identification model, including: Apply community detection algorithms to identify closely connected groups of examinees and obtain a candidate list of collaborative cheating groups; Based on the candidate list of collaborative cheating groups, an isolation forest detection algorithm is used to screen and determine groups with abnormal behavior patterns to locate collaborative cheating behaviors and obtain collaborative cheating identification results; Based on the collaborative cheating identification results, a Q-learning-based reinforcement learning framework is introduced, and an adaptive learning rate adjustment mechanism is integrated, so that the model can dynamically adjust the learning rate according to the changes in the examination room environment, and the collaborative cheating identification model to be optimized is obtained. The changes in the examination room environment include the examination time period and the distribution density of candidates in the examination room; The cheating behavior data accumulated in historical exams are used as a priori knowledge base, and transfer learning technology is applied to train the collaborative cheating identification model to be optimized to obtain an optimized collaborative cheating identification model.

4. The method according to claim 3, characterized in that: The cheating behavior data accumulated in the historical exams is used as a priori knowledge base, and the transfer learning technology is applied to train the collaborative cheating identification model to be optimized to obtain the optimized collaborative cheating identification model, including: Collect cheating behavior data accumulated in historical exams to obtain a priori knowledge base, wherein the cheating behavior data includes confirmed cheating cases, suspected cheating behavior records, and related environmental information; The prior knowledge base is used as source domain data for transfer learning, and the collaborative cheating identification model to be optimized is preliminarily trained by using domain adversarial training technology to optimize the key feature extraction performance of the collaborative cheating identification model to be optimized, thereby obtaining the collaborative cheating identification model being optimized; The fine-tuning technology and the prior knowledge base are used to perform secondary training on the optimized collaborative cheating identification model to accelerate the convergence process of the optimized collaborative cheating identification model. Regularization technology is introduced in the secondary training process to prevent the optimized collaborative cheating identification model from overfitting. The optimized collaborative cheating identification model is obtained after the training.

5. The method according to claim 1, characterized in that The extreme gradient boosting learning algorithm is used to weightedly fuse the abnormal behavior information, the collaborative cheating identification result and the target detection strategy. In the weighted fusion process, different weight coefficients are set to allocate the contribution of each information, and a comprehensive score is obtained after calculation. Based on the comprehensive score, an abnormal behavior determination report is generated, including: An extreme gradient boosting learning algorithm is used to perform weighted fusion on the abnormal behavior information, the collaborative cheating identification result, and the target detection strategy. In the weighted fusion process, different weight coefficients are set to allocate the contribution of each information to obtain an initial comprehensive evaluation result; Based on the initial comprehensive evaluation result, the feature importance evaluation function of the extreme gradient boosting learning algorithm is used to adjust the weight coefficients of the abnormal behavior information, the collaborative cheating identification result, and the target detection strategy, and the Shapley value method is introduced to quantify the contribution of each information to obtain an optimized comprehensive evaluation result; Constructing an initial extreme gradient boosting learning model, inputting the optimized comprehensive evaluation result into the initial extreme gradient boosting learning model, and introducing a cross-validation method and a Bayesian optimization technique for training to obtain an optimized extreme gradient boosting learning model; The optimized extreme gradient boosting learning model and multiple learning models are used to predict the cheating behavior data respectively to obtain multiple original prediction results, and the multiple original prediction results are integrated by an ensemble learning method to obtain a fused prediction result. The scoring model is initialized based on the fused prediction result to obtain a preliminary comprehensive scoring model, and an adaptive weighted fusion method is introduced into the preliminary comprehensive scoring model to dynamically adjust the weights of each model according to the performance of the preliminary comprehensive scoring model on different subsets to obtain an optimized comprehensive scoring model. The multiple learning models include a random forest and a support vector machine; A comprehensive score is calculated based on the optimized comprehensive scoring model, and an abnormal behavior determination report is generated according to the comprehensive score. The abnormal behavior determination report includes a description of the abnormal behavior, an assessment of the possibility of cheating, and handling suggestions.

6. The method according to claim 5, characterized in that The optimized extreme gradient boosting learning model and multiple learning models are used to predict the cheating behavior data respectively to obtain multiple original prediction results, and the multiple original prediction results are integrated by an integrated learning method to obtain a fusion prediction result. The initial scoring model is trained based on the fusion prediction result to obtain a preliminary comprehensive scoring model, and an adaptive weighted fusion method is introduced into the preliminary comprehensive scoring model to dynamically adjust the weights of each model according to the performance of the preliminary comprehensive scoring model on different subsets to obtain an optimized comprehensive scoring model. The multiple learning models include random forests and support vector machines, including: The optimized extreme gradient boosting learning model and multiple learning models are used as initial base learners; the initial base learners are trained on the cheating behavior data to obtain multiple target base learners, the cheating behavior data is predicted based on the multiple target base learners to obtain multiple prediction results, the multiple prediction results are fused by weighted average to obtain a fused prediction result, and the fused prediction result is used as a meta-feature; Constructing an initial scoring model, using the initial scoring model as a meta-learner, training the meta-learner based on the meta-features and true labels to obtain a preliminary comprehensive scoring model, wherein the true labels include the cheating behavior data and data labels, and the data labels include normal behavior, abnormal behavior, and cheating behavior; The comprehensive behavioral feature set is divided into multiple training sets and validation sets, the performance of the preliminary comprehensive scoring model is evaluated on each validation set to obtain an evaluation performance result, and the prediction errors of the multiple target base learners are recorded, based on the evaluation performance result and the prediction error, the weights of the multiple target base learners are dynamically adjusted using an online gradient descent method to obtain adjusted weights, and the preliminary comprehensive scoring model is trained using the adjusted weights to obtain an optimized comprehensive scoring model.

7. The method according to claim 1, characterized in that The application is based on the time series prediction algorithm of the long short-term memory network, and combined with the attention mechanism optimization model, analyzes the behavior pattern on the time dimension of the comprehensive behavior feature set to obtain abnormal behavior information, including: Arranging the data in the comprehensive behavior feature set in chronological order and performing feature extraction to obtain time series data, wherein the time series data includes statistical features within a time window; A time series prediction model based on a long short-term memory network is constructed, and an attention mechanism is introduced into the time series prediction model to optimize the model, thereby obtaining an optimized time series prediction model; An optimized time series prediction model is used to analyze the time series data to obtain analysis results, and the analysis results are processed using an anomaly detection algorithm to obtain behavior anomaly values. A qualified threshold is preset, and when the behavior anomaly value is greater than the qualified threshold, it is judged as abnormal behavior, and the candidate's abnormal behavior information is obtained.

8. A system for analyzing abnormal behavior of examinees in a computer-based examination, characterized in that: include: An integration module is used to extract the examinee's facial expression and body posture information from the monitoring image, and use a behavior recognition algorithm to analyze the keyboard tapping frequency and the mouse movement trajectory to obtain behavior information, and integrate the examinee's facial expression, the body posture information and the behavior information to obtain a comprehensive behavior feature set; An analysis module, for applying a time series prediction algorithm based on a long short-term memory network and combining it with an attention mechanism optimization model to analyze the behavior pattern in the time dimension of the comprehensive behavior feature set to obtain abnormal behavior information; The recognition module is used to create a target examination room environment map according to the candidate seat layout through a graph neural network algorithm, identify collaborative cheating behavior by analyzing the abnormal behavior information, obtain collaborative cheating identification results, introduce a Q-learning-based reinforcement learning framework, dynamically adjust the detection strategy according to the abnormal behavior information and the collaborative cheating identification results, and obtain a target detection strategy; The generation module is used to adopt an extreme gradient boosting learning algorithm to weightedly fuse the abnormal behavior information, the collaborative cheating identification result and the target detection strategy, and in the weighted fusion process, the contribution of each information is allocated by setting different weight coefficients, and a comprehensive score is obtained after calculation, and an abnormal behavior determination report is generated based on the comprehensive score.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a method for analyzing abnormal behavior of examinees in a computer-based examination as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, a method for analyzing abnormal behavior of examinees in a computer-based examination room as described in any one of claims 1 to 7 is implemented.

Images