Examination proposition person behavior model construction method and system based on deep learning

Abstract

The invention provides an examination proposition person behavior model construction method and system based on deep learning, and relates to the field of artificial intelligence and educational evaluation, and the method comprises the steps: obtaining proposition records, constructing a knowledge application graph, analyzing proposition preferences based on a time window, training a proposition behavior recognition network, and optimizing a migration path and difficulty distribution. And anomaly detection of proposition behaviors is realized. According to the method, the knowledge preference and the behavior mode of the proposition person can be accurately described, the proposition quality monitoring efficiency is improved, the proposition risk is reduced, and the examination fairness is guaranteed.

Description

Technical Field

[0001] The present invention relates to artificial intelligence and educational assessment technology, and in particular to a method and system for constructing a behavior model of test setters based on deep learning. Background Art

[0002] With the development of educational informatization, intelligent question-setting systems have been widely used in the generation of examination questions. In the intelligent question-setting process, the question-setting behavior characteristics of the question setter have a significant impact on the quality of the test questions. Existing technologies mainly analyze question-setting behavior from two dimensions: question difficulty distribution and knowledge point coverage. However, these methods have the following shortcomings: existing methods only focus on the frequency of use of individual knowledge points, ignoring the migration relationship between knowledge points, and are unable to effectively characterize the question setter's overall grasp of the knowledge system; existing technologies treat question-setting behavior as a static feature, failing to consider the dynamic changes in question setter preferences over time, resulting in insufficient adaptability of the question-setting behavior model; and the lack of an abnormal monitoring mechanism for question-setting behavior makes it impossible to timely detect and intervene in inappropriate question-setting behavior. In addition, existing technologies often analyze knowledge point selection and difficulty setting separately, failing to reveal the inherent relationship between the two, and reducing the accuracy of question-setting behavior modeling. Therefore, a question-setting behavior modeling method that can comprehensively consider multiple dimensions such as knowledge point migration relationship, dynamic preference changes, and abnormal behavior monitoring is needed to improve the quality control level of intelligent question-setting systems. Summary of the Invention

[0003] The embodiments of the present invention provide a method and system for constructing an examination question setter behavior model based on deep learning, which can solve the problems in the prior art.

[0004] A first aspect of an embodiment of the present invention provides a method for constructing an examination question setter behavior model based on deep learning, comprising: Obtaining the test setter's question setting records, constructing a knowledge association analysis module, calculating the usage frequency based on the knowledge point tags in the question setting records, constructing a knowledge point mapping network based on a preset knowledge system, obtaining migration weights by calculating the conditional probabilities between knowledge points, combining the knowledge point mapping network and the migration weights to generate a knowledge application graph, and outputting a knowledge application vector based on the knowledge application graph; A dynamic preference analysis module is constructed based on the knowledge application vector. The dynamic preference analysis module groups proposition behaviors into time windows according to the time information in the proposition record, extracts features of the knowledge point selection sequence within each time window, and generates a proposition preference vector based on the difficulty information. A proposition behavior recognition network is trained based on proposition preference vectors. The proposition behavior recognition network models the knowledge point selection sequence and difficulty information respectively, and obtains proposition behavior features by optimizing the migration path and difficulty distribution in the knowledge application graph. The proposition behavior characteristics are input into the anomaly detection module, and the anomaly detection module calculates the warning threshold according to the migration path and difficulty distribution, and outputs warning information and the cause of the anomaly when the proposition behavior characteristics deviate from the warning threshold.

[0005] Constructing a knowledge association analysis module, wherein the knowledge association analysis module calculates the usage frequency according to the knowledge point tags in the proposition record and constructs a knowledge point mapping network based on a preset knowledge system, including: Convert the proposition text of the proposition record into a word vector sequence, perform bidirectional attention calculation on the query vector and the word vector sequence, and identify candidate knowledge point segments based on the attention distribution; Performing semantic vectorization on the candidate knowledge point segments, calculating the semantic similarity between the semantic vectorization and the standard knowledge points in the preset knowledge system, and extracting the corresponding knowledge point labels based on the similarity matching results; Calculate the direct usage frequency of knowledge point tags in the proposition record. At the same time, based on the hierarchical structure of the preset knowledge system, count the usage frequencies of the parent node and child node corresponding to each knowledge point tag in the proposition record and perform weighted combination to obtain the hierarchical usage frequency of knowledge points containing hierarchical relationships. Using knowledge point labels as nodes and hierarchical usage frequencies as initial node weights, node connections are established based on the dependencies between knowledge points in the preset knowledge system to construct an initial knowledge point mapping network. The order of knowledge point selection in the proposition process is extracted from the proposition records. The migration rules between knowledge points are analyzed based on the knowledge point selection order, and the correlation between nodes is calculated based on the migration rules. The connection strengths between nodes in the initial knowledge point mapping network are updated according to the degree of correlation, and feature extraction is performed on the updated node weights to finally construct a knowledge point mapping network that represents the association relationship between knowledge points.

[0006] The conditional probability between knowledge points is calculated to obtain the transfer weight, and the knowledge point mapping network and the transfer weight are combined to generate a knowledge application graph. The knowledge application vector output based on the knowledge application graph includes: Obtain a knowledge point selection sequence in the proposition record, set multiple sliding time windows of different scales for the knowledge point selection sequence, count the number of occurrences of knowledge points and the number of consecutive occurrences of knowledge point pairs in each sliding time window, calculate the conditional probability of the knowledge points in the current window, and perform a weighted combination of the conditional probabilities of the knowledge points in different sliding time windows to obtain the migration weights between the knowledge points; Combining the migration weights between knowledge points with the knowledge point mapping network to generate a knowledge application graph, wherein the knowledge application graph includes nodes, connection relationships between nodes, and migration weights corresponding to the connection relationships; Performing a transformation matrix transformation on the node feature vectors in the knowledge application graph to obtain a transformed feature vector of the node, performing a splicing operation on the transformed feature vectors of adjacent nodes and processing them with an activation function, obtaining an attention coefficient between the nodes through a normalization operation, and performing weighted aggregation on the transformed feature vectors of adjacent nodes based on the attention coefficient to obtain an updated feature vector that integrates feature information of adjacent nodes; An average pooling operation is performed on the updated feature vectors of all nodes in the knowledge application graph to generate a knowledge application vector, which is used to characterize the knowledge point selection pattern and migration law.

[0007] According to the time information in the proposition record, the proposition behavior is grouped into time windows, and the feature extraction of the knowledge point selection sequence in each time window is performed. The proposition preference vector is generated by combining the difficulty information, including: Constructing a time series for the time information in the proposition record, performing kernel density estimation on the time series to obtain a proposition activity density curve, dynamically calculating the time window size based on the proposition activity density curve, and obtaining the time window division of the proposition behavior; Based on the time window division, extract the knowledge point selection sequence within each time window from the proposition record, construct a knowledge point word vector mapping table, map the knowledge point selection sequence within each time window into a word vector sequence, and incorporate relative position encoding information to obtain a position-aware vector sequence; Inputting the position-aware vector sequence into a bidirectional recurrent neural network, extracting context-dependent features of knowledge point selection to obtain a hidden layer state sequence; The pre-acquired knowledge application vector is transformed nonlinearly to generate a query vector, the similarity between the query vector and the hidden layer state sequence is calculated to obtain an attention score, and the attention score is normalized to obtain an attention weight; The context vector is obtained by weighted summing of the hidden layer state sequence based on the attention weight, and the context vector is fused with the query vector and then subjected to nonlinear transformation to generate a proposition preference vector that represents the knowledge point selection preference of the propositioner within the time window.

[0008] The proposition behavior recognition network is trained based on the proposition preference vector. The proposition behavior recognition network models the knowledge point selection sequence and difficulty information respectively. By optimizing the migration path and difficulty distribution in the knowledge application graph, the proposition behavior features obtained include: A migration path identification module is constructed based on the proposition preference vector, the proposition preference vector is input into the feature extraction network to obtain initial features, a feature propagation matrix is ​​generated based on the migration weights between nodes in the knowledge application graph, and the feature propagation matrix is ​​used to aggregate the initial features to obtain the migration path features of the knowledge point selection sequence; Encoding difficulty information into a difficulty vector, calculating a correlation score between the difficulty vector and a migration path feature, and identifying a migration pattern of a knowledge point selection sequence at each difficulty level based on the correlation score; Constructing a difficulty distribution weight based on the migration pattern of the knowledge point selection sequence, performing weighted calculation on the difficulty distribution weight and the migration path feature to obtain a difficulty-adaptive migration path feature, and simultaneously correcting the difficulty vector based on the difficulty-adaptive migration path feature to obtain a corrected difficulty vector; The difficulty-adaptive migration path feature and the modified difficulty vector are subjected to feature fusion to generate a proposition behavior feature that reflects the collaborative relationship between the knowledge point selection sequence and the difficulty information.

[0009] The proposition behavior characteristics are input into the anomaly detection module, which calculates the warning threshold based on the migration path and difficulty distribution. When the proposition behavior characteristics deviate from the warning threshold, the module outputs warning information and the cause of the anomaly, including: Inputting the proposition behavior features into an anomaly detection module, the anomaly detection module extracts migration path information and difficulty distribution information, calculates the migration probability between knowledge points based on the migration path information to obtain a forward coupling matrix, calculates the difficulty constraint strength of knowledge point pairs based on the difficulty distribution information to obtain a reverse coupling matrix, and combines the forward coupling matrix and the reverse coupling matrix to construct a bidirectional coupling representation; Using the bidirectional coupling representation, forward propagation is performed on the migration path information to obtain path features, and reverse propagation is performed on the difficulty distribution information to obtain difficulty features. The path features and the difficulty features are interactively calculated to obtain a coupling anomaly score. Calculating a stability index of knowledge point migration based on the bidirectional coupling representation, dynamically correcting the distribution of coupling anomaly scores using the stability index to obtain a warning threshold, and triggering anomaly detection when the coupling anomaly score exceeds the warning threshold; An anomaly propagation network is constructed based on the bidirectional coupling representation. In the anomaly propagation network, the knowledge point migration sequence that causes the anomaly is located along the forward path, and the knowledge point combination that violates the difficulty constraint is identified along the reverse path. The located abnormal knowledge point migration sequence and the knowledge point combination that violates the difficulty constraint are used as the causes of the anomaly to output warning information.

[0010] Calculating the migration probability between knowledge points based on the migration path information to obtain a forward coupling matrix, calculating the difficulty constraint strength of knowledge point pairs based on the difficulty distribution information to obtain a reverse coupling matrix, and combining the forward coupling matrix and the reverse coupling matrix to construct a bidirectional coupling representation includes: Based on the migration sequence and temporal features in the migration path information, the number of migration occurrences between knowledge points is counted according to the time window. The migration statistics in different time windows are weighted and combined to calculate the migration probability between knowledge points to obtain a forward coupling matrix. The forward coupling matrix represents the migration probability from the source knowledge point to the target knowledge point. Based on the difficulty distribution information, a hierarchical difficulty constraint graph is constructed. In the hierarchical difficulty constraint graph, knowledge points are organized into layers according to difficulty values. The difficulty gradients of knowledge point pairs within a layer and the cross-layer constraints of knowledge point pairs between layers are calculated to obtain a reverse coupling matrix. The reverse coupling matrix represents the strength of the difficulty constraints between knowledge point pairs. The forward coupling matrix and the reverse coupling matrix are normalized, and the weighted combination coefficient of the migration probability and the difficulty constraint strength is calculated. The migration probability and the difficulty constraint strength are nonlinearly combined based on the weighted combination coefficient to construct a bidirectional coupling representation, which includes a migration probability vector, a constraint strength vector and a coupling weight coefficient.

[0011] A second aspect of an embodiment of the present invention provides a system for constructing a behavior model of an examination question setter based on deep learning, including: The first unit is used to obtain the question setting records of the test setter and construct a knowledge association analysis module. The knowledge association analysis module calculates the usage frequency based on the knowledge point tags in the question setting records, constructs a knowledge point mapping network based on a preset knowledge system, obtains migration weights by calculating the conditional probabilities between knowledge points, combines the knowledge point mapping network and the migration weights to generate a knowledge application graph, and outputs a knowledge application vector based on the knowledge application graph; The second unit is used to build a dynamic preference analysis module based on the knowledge application vector. The dynamic preference analysis module groups the proposition behavior into time windows according to the time information in the proposition record, extracts features of the knowledge point selection sequence within each time window, and generates a proposition preference vector based on the difficulty information; The third unit is used to train a proposition behavior recognition network based on the proposition preference vector. The proposition behavior recognition network models the knowledge point selection sequence and difficulty information respectively, and obtains proposition behavior characteristics by optimizing the migration path and difficulty distribution in the knowledge application graph; The fourth unit is used to input the proposition behavior characteristics into the anomaly detection module. The anomaly detection module calculates the warning threshold according to the migration path and difficulty distribution, and outputs warning information and the cause of the anomaly when the proposition behavior characteristics deviate from the warning threshold.

[0012] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0013] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0014] In this embodiment, by constructing a deep learning-based behavior model for test setters, accurate characterization of test setting behavior and abnormal monitoring are achieved, effectively improving the quality and fairness of test setting. The introduction of knowledge application graphs and dynamic preference analysis can capture the behavioral patterns of test setters in knowledge point selection and difficulty setting, making the extraction of test setting behavior features more comprehensive and accurate, and helping to timely discover deviations and irregular behaviors in the test setting process. The use of deep learning methods to optimize knowledge transfer paths and difficulty distribution can not only automatically identify test setting behavior features, but also actively issue warning information and analyze abnormal causes based on warning thresholds, providing a powerful quality monitoring tool for test management departments and improving the scientificity and standardization of test setting. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of a method for constructing an exam question setter behavior model based on deep learning according to an embodiment of the present invention; Figure 2 This is a flowchart of abnormal detection and analysis of proposition behavior in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, 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. 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 ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0017] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0018] Figure 1 This is a flow chart of a method for constructing a behavior model of an examination question setter based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method includes: Obtaining the test setter's question setting records, constructing a knowledge association analysis module, calculating the usage frequency based on the knowledge point tags in the question setting records, constructing a knowledge point mapping network based on a preset knowledge system, obtaining migration weights by calculating the conditional probabilities between knowledge points, combining the knowledge point mapping network and the migration weights to generate a knowledge application graph, and outputting a knowledge application vector based on the knowledge application graph; A dynamic preference analysis module is constructed based on the knowledge application vector. The dynamic preference analysis module groups proposition behaviors into time windows according to the time information in the proposition record, extracts features of the knowledge point selection sequence within each time window, and generates a proposition preference vector based on the difficulty information. A proposition behavior recognition network is trained based on proposition preference vectors. The proposition behavior recognition network models the knowledge point selection sequence and difficulty information respectively, and obtains proposition behavior features by optimizing the migration path and difficulty distribution in the knowledge application graph. The proposition behavior characteristics are input into the anomaly detection module, and the anomaly detection module calculates the warning threshold according to the migration path and difficulty distribution, and outputs warning information and the cause of the anomaly when the proposition behavior characteristics deviate from the warning threshold.

[0019] In an optional embodiment, a knowledge association analysis module is constructed, wherein the knowledge association analysis module calculates the usage frequency according to the knowledge point tags in the proposition record, and constructs a knowledge point mapping network based on a preset knowledge system, including: Convert the proposition text of the proposition record into a word vector sequence, perform bidirectional attention calculation on the query vector and the word vector sequence, and identify candidate knowledge point segments based on the attention distribution; Performing semantic vectorization on the candidate knowledge point segments, calculating the semantic similarity between the semantic vectorization and the standard knowledge points in the preset knowledge system, and extracting the corresponding knowledge point labels based on the similarity matching results; Calculate the direct usage frequency of knowledge point tags in the proposition record. At the same time, based on the hierarchical structure of the preset knowledge system, count the usage frequencies of the parent node and child node corresponding to each knowledge point tag in the proposition record and perform weighted combination to obtain the hierarchical usage frequency of knowledge points containing hierarchical relationships. Using knowledge point labels as nodes and hierarchical usage frequencies as initial node weights, node connections are established based on the dependencies between knowledge points in the preset knowledge system to construct an initial knowledge point mapping network. The order of knowledge point selection in the proposition process is extracted from the proposition records. The migration rules between knowledge points are analyzed based on the knowledge point selection order, and the correlation between nodes is calculated based on the migration rules. The connection strengths between nodes in the initial knowledge point mapping network are updated according to the degree of correlation, and feature extraction is performed on the updated node weights to finally construct a knowledge point mapping network that represents the association relationship between knowledge points.

[0020] In this embodiment, a data set containing multiple proposition records is first received, each proposition record containing information such as proposition text, proposition time, and proposition person. For each proposition record, its proposition text is converted into a word vector sequence. Specifically, a pre-trained word embedding model (such as Word2Vec) is used to convert each word in the text into a 300-dimensional word vector to form a word vector sequence. At the same time, key knowledge points are extracted from the preset knowledge system as query vectors.

[0021] Candidate knowledge point segments are identified by performing bidirectional attention calculations between the query vector and the word vector sequence. In this bidirectional attention mechanism, a similarity score is calculated between the query vector and each word vector in the word vector sequence. Similarity scores are calculated using the inner product of the vectors. For example, the similarity calculation between the query vector "quadratic function" and the word vectors in the proposition text "Analyze the properties of a quadratic function" yields a similarity distribution of [0.1, 0.2, 0.8, 0.7, 0.3]. Based on this similarity distribution, consecutive text segments with scores above a threshold of 0.5 are identified as candidate knowledge point segments. In this example, "quadratic function" is identified as a candidate segment.

[0022] For the identified candidate knowledge point segments, semantic vectorization is performed. A text encoder (such as BERT) is used to encode the candidate segments, obtaining a 768-dimensional semantic vector. The semantic similarity between this semantic vector and standard knowledge points in the predefined knowledge system is calculated. For example, the cosine similarity between the semantic vector of the candidate segment "quadratic function" and the semantic vectors of standard knowledge points in the knowledge system, such as "quadratic function," "function," and "linear function," is calculated, yielding similarities of 0.92, 0.75, and 0.68, respectively. Based on the similarity matching results, the standard knowledge point "quadratic function" with the highest similarity is extracted as the knowledge point label for the candidate segment. The frequency of use of the knowledge point label in the question record is calculated. The direct frequency of use is calculated by counting the number of times each knowledge point label appears in all question records. For example, if "quadratic function" appears 120 times in 1000 question records, its direct frequency of use is 120. Furthermore, based on the hierarchical structure of the predefined knowledge system, the frequency of use of the parent and child nodes corresponding to each knowledge point label in the question record is counted. For example, the parent node "Function" for the knowledge point "Quadratic Function" appears 350 times, while its child nodes "Properties of Quadratic Function" and "Graph of Quadratic Function" appear 50 and 40 times, respectively. The hierarchical frequency of "Quadratic Function" is calculated as follows: the direct frequency of 120 plus the parent node frequency of 350 multiplied by the upward weight of 0.3, plus the child node frequency (50 + 40) multiplied by the downward weight of 0.5. The resulting hierarchical frequency is 120 + 350 × 0.3 + (50 + 40) × 0.5 = 270.

[0023] When constructing a knowledge point mapping network, knowledge point labels are used as nodes, and hierarchical usage frequencies are used as initial node weights. Node connections are established based on the dependencies between knowledge points in the pre-set knowledge system to construct the initial knowledge point mapping network. For example, in the initial network, a connection is established between "Function" and "Quadratic Function," with an initial connection strength of 0.7, indicating a strong dependency between the two.

[0024] Extract the order in which knowledge points were selected during the question-setting process from the question-setting records. For example, during one question-setting process, the teacher first selected "function," then "quadratic function," and finally "properties of quadratic functions," forming the knowledge point selection sequence ["function," "quadratic function," "properties of quadratic functions"]. By analyzing the knowledge point selection sequence across multiple question-setting processes, the probability of transition between knowledge points was calculated. For example, the probability of transitioning from "function" to "quadratic function" was 0.4, and the probability of transitioning from "quadratic function" to "properties of quadratic functions" was 0.6.

[0025] The correlation between nodes is calculated based on the migration pattern. For a node pair (A, B), the correlation is calculated by combining the direct migration probability and the indirect migration path probability. For example, the correlation between "function" and "quadratic function property" is the direct migration probability of 0.1 plus the probability of the indirect path "function → quadratic function → quadratic function property" of 0.4 × 0.6 = 0.24, for a total correlation of 0.34. The connection strengths between nodes in the initial knowledge point mapping network are updated based on the calculated correlations. For example, the connection strength between "function" and "quadratic function" is updated from the initial value of 0.7 to 0.7 × 0.8 + 0.4 × 0.2 = 0.64, where 0.8 and 0.2 are the weights of the initial connection strength and migration probability, respectively.

[0026] Finally, feature extraction is performed on the updated network, including calculating network features such as the centrality and clustering coefficient of each node. For example, the degree centrality of the "quadratic function" node is 0.85, indicating that this knowledge point has high importance in the entire network. Using these features, a knowledge point mapping network is ultimately constructed that represents the relationships between knowledge points.

[0027] In this embodiment, by introducing deep semantic understanding and knowledge graph modeling, accurate identification and mapping of implicit knowledge points in the proposition text are achieved, and the accuracy of knowledge point label extraction is improved. At the same time, by integrating the hierarchical structure of the knowledge system and the proposition order, a graph structure reflecting the frequency of use and migration relationship of knowledge points is constructed, which helps to depict the real pattern of knowledge application of the propositioner. The updated knowledge point mapping network can not only reflect the semantic and logical dependencies between knowledge points, but also reflect their evolution path in the actual proposition process, providing higher quality structured input for subsequent behavior modeling and preference analysis.

[0028] In an optional embodiment, the conditional probabilities between knowledge points are calculated to obtain transfer weights, and the knowledge point mapping network and the transfer weights are combined to generate a knowledge application graph. Outputting the knowledge application vector based on the knowledge application graph includes: Obtain a knowledge point selection sequence in the proposition record, set multiple sliding time windows of different scales for the knowledge point selection sequence, count the number of occurrences of knowledge points and the number of consecutive occurrences of knowledge point pairs in each sliding time window, calculate the conditional probability of the knowledge points in the current window, and perform a weighted combination of the conditional probabilities of the knowledge points in different sliding time windows to obtain the migration weights between the knowledge points; Combining the migration weights between knowledge points with the knowledge point mapping network to generate a knowledge application graph, wherein the knowledge application graph includes nodes, connection relationships between nodes, and migration weights corresponding to the connection relationships; Performing a transformation matrix transformation on the node feature vectors in the knowledge application graph to obtain a transformed feature vector of the node, performing a splicing operation on the transformed feature vectors of adjacent nodes and processing them with an activation function, obtaining an attention coefficient between the nodes through a normalization operation, and performing weighted aggregation on the transformed feature vectors of adjacent nodes based on the attention coefficient to obtain an updated feature vector that integrates feature information of adjacent nodes; An average pooling operation is performed on the updated feature vectors of all nodes in the knowledge application graph to generate a knowledge application vector, which is used to characterize the knowledge point selection pattern and migration law.

[0029] In one embodiment, the user's question-setting records during the learning process are first obtained, and a knowledge point selection sequence is extracted from them. For example, for a student learning mathematics, their knowledge point selection sequence may be: [trigonometric functions, vectors, derivatives, integrals, differential equations, probability theory]. Multiple sliding time windows of different scales are set for this sequence, such as four windows of 3 days, 7 days, 14 days, and 30 days. Within each window, the number of occurrences of each knowledge point and the number of consecutive occurrences of knowledge point pairs are counted. Taking a 7-day window as an example, if "trigonometric functions" appears 5 times, "vectors" appears 3 times, and "trigonometric functions" is followed by "vectors" 2 times, then the conditional probability of "trigonometric functions" to "vectors" is 2 / 5 = 0.4.

[0030] For each sliding time window, the conditional probability between knowledge points is calculated. For example, the conditional probability from "trigonometric function" to "vector" is 0.3 for a 3-day window, 0.4 for a 7-day window, 0.45 for a 14-day window, and 0.5 for a 30-day window. These conditional probabilities are weighted, with weights set to [0.1, 0.2, 0.3, 0.4], corresponding to the importance of different time windows. Through weighted summation, the transfer weight from "trigonometric function" to "vector" is 0.3 × 0.1 + 0.4 × 0.2 + 0.45 × 0.3 + 0.5 × 0.4 = 0.44. Similar calculations are performed for all pairs of knowledge points, ultimately resulting in a complete transfer weight matrix.

[0031] The transfer weights calculated above are combined with the pre-built knowledge point mapping network to generate a knowledge application graph. A knowledge point mapping network is a graph structure built based on the intrinsic connections between knowledge points, where nodes represent knowledge points and edges represent the relationships between them. For example, in a mathematics knowledge point mapping network, "derivative" has direct connections with knowledge points such as "differential" and "integral." Transfer weights are assigned to these connections to form a weighted knowledge application graph. In this graph, the connection weight from "trigonometric functions" to "vectors" is 0.44, indicating the probability strength of a learner's transfer from "trigonometric functions" to "vectors."

[0032] To fully utilize the information in the knowledge application graph, the node features in the graph are processed. Assume that each knowledge point node initially has a 64-dimensional feature vector representing the attribute information of the knowledge point. A 64×32 transformation matrix is ​​used to transform the node feature vector to obtain a 32-dimensional transformed feature vector. For example, the initial feature vector of the "derivative" node is transformed to obtain a new 32-dimensional transformed feature vector.

[0033] For each node in the graph, the transformed feature vectors of its neighboring nodes are concatenated. For example, the transformed feature vectors of the "derivative" node and its neighboring "differential" node are concatenated into a 64-dimensional vector. This concatenated vector is processed using the LeakyReLU activation function and then normalized to obtain the attention coefficients between the nodes. For example, the attention coefficient from "derivative" to "differential" is 0.6, and the attention coefficient from "derivative" to "integral" is 0.4. Based on the calculated attention coefficients, the transformed feature vectors of neighboring nodes are weighted and aggregated. For the "derivative" node, its updated feature vector is the weighted sum of the transformed feature vectors of the neighboring "differential" and "integral" nodes, with weights of 0.6 and 0.4 respectively. This process ensures that each node's features not only contain its own information but also incorporate information from neighboring nodes, reflecting the relevance between knowledge points.

[0034] An average pooling operation is performed on the updated feature vectors of all nodes in the knowledge application graph. This operation adds the updated feature vectors of all nodes and divides them by the number of nodes to generate a unified knowledge application vector. For example, for a knowledge application graph containing 100 knowledge point nodes, these 100 32-dimensional updated feature vectors are averaged to obtain a 32-dimensional knowledge application vector. This vector comprehensively reflects the learner's knowledge point selection patterns and knowledge transfer patterns, and can be used for subsequent applications such as learning path recommendation and learning effect prediction.

[0035] In practice, the knowledge application graph and knowledge application vector can be updated regularly to adapt to changes in the learner's knowledge mastery. For example, they can be updated weekly based on the latest learning records to ensure they accurately reflect the learner's current knowledge status and learning patterns. This approach can provide learners with more personalized and effective learning support.

[0036] In this embodiment, the co-occurrence and migration patterns of knowledge points at different time granularities are captured through a multi-scale sliding window, thereby improving the accuracy of modeling the knowledge migration rules in proposition behavior. The knowledge application graph generated by combining conditional probability calculation with the knowledge point mapping network can fully reflect the dynamic associations and semantic relationships between knowledge points. Under the graph neural network mechanism, the attention mechanism is introduced to perform weighted aggregation of node features, which effectively enhances the model's ability to identify key knowledge migration paths. The knowledge application vector finally generated can represent the propositioner's behavioral preferences and logical patterns in the process of knowledge selection and migration in a high-dimensional manner, providing key feature support for subsequent behavior recognition and anomaly detection.

[0037] In an optional embodiment, the proposition behavior is grouped into time windows according to the time information in the proposition record, and feature extraction is performed on the knowledge point selection sequence within each time window. The proposition preference vector is generated in combination with the difficulty information, including: Constructing a time series for the time information in the proposition record, performing kernel density estimation on the time series to obtain a proposition activity density curve, dynamically calculating the time window size based on the proposition activity density curve, and obtaining the time window division of the proposition behavior; Based on the time window division, extract the knowledge point selection sequence within each time window from the proposition record, construct a knowledge point word vector mapping table, map the knowledge point selection sequence within each time window into a word vector sequence, and incorporate relative position encoding information to obtain a position-aware vector sequence; Inputting the position-aware vector sequence into a bidirectional recurrent neural network, extracting context-dependent features of knowledge point selection to obtain a hidden layer state sequence; The pre-acquired knowledge application vector is transformed nonlinearly to generate a query vector, the similarity between the query vector and the hidden layer state sequence is calculated to obtain an attention score, and the attention score is normalized to obtain an attention weight; The context vector is obtained by weighted summing of the hidden layer state sequence based on the attention weight, and the context vector is fused with the query vector and then subjected to nonlinear transformation to generate a proposition preference vector that represents the knowledge point selection preference of the propositioner within the time window.

[0038] In this embodiment, a time series is constructed based on the time information in the proposition record. The time series includes a set of time points when the propositioner performs proposition activities. For example, for a proposition record of a propositioner in the past three months, the timestamp information of each proposition activity can be extracted to form a time series {t1, t2, ..., t n}.

[0039] Kernel density estimation is performed on the time series to obtain the proposition activity density curve. The kernel density estimation process uses a Gaussian kernel function, and the bandwidth parameter is determined by the cross-validation method. For example, for the above example, a bandwidth of 0.5 days can be selected. The generated density curve represents the distribution of proposition activities on the time axis, and the peak of the curve corresponds to the time period with frequent proposition activities. The time window size is dynamically calculated based on the proposition activity density curve. The specific method is to identify the local minimum point in the density curve as the window dividing point to ensure that the proposition activity in each window is relatively concentrated and that there are differences in activity patterns between windows. For the example data, five time windows may be identified, with window sizes of 12 days, 18 days, 25 days, 16 days, and 20 days respectively.

[0040] Based on time window division, the sequence of knowledge point selections within each time window is extracted from the question record. For example, in the first window, the sequence of knowledge points selected by the question setter might be {"linear algebra," "calculus," "probability theory," "linear algebra," "numerical analysis"}. To process these textual knowledge points, a knowledge point word vector mapping table is constructed. This mapping table is generated using a pretrained word embedding model, mapping knowledge point concepts into a 64-dimensional vector space. For common subject knowledge points, a word embedding model fine-tuned for the education domain is used to ensure that the vectors accurately reflect the semantic relationships between knowledge points. The sequence of knowledge point selections within each time window is mapped into a sequence of word vectors. For the above example, a sequence of five 64-dimensional vectors is obtained. To capture the sequential information of knowledge point selections, relative position encoding information is incorporated to produce a position-aware vector sequence. This position encoding uses the sine-cosine positional encoding method, encoding the position information into a vector of the same dimension as the knowledge point vector. The position encoding is then fused with the knowledge point vector through addition to obtain a position-aware knowledge point representation.

[0041] The position-aware vector sequence is input into a bidirectional recurrent neural network to extract context-dependent features of knowledge point selection, generating a hidden state sequence. The bidirectional recurrent neural network uses a gated recurrent unit structure with a hidden dimension of 128 and consists of two hidden layers. For each knowledge point, forward propagation captures information about the previously selected knowledge point, while backpropagation captures information about the subsequent selected knowledge point. The two are combined to form a complete contextual representation. For example, for the "Probability Theory" in the example sequence, its hidden state includes the influence of "Linear Algebra" and "Calculus," as well as the influence on the subsequent choices of "Linear Algebra" and "Numerical Analysis." This generates a 128-dimensional vector representing the semantic and positional importance of this knowledge point in the overall selection sequence.

[0042] The previously acquired knowledge application vector is transformed nonlinearly to generate a query vector. The knowledge application vector is extracted from the question record and reflects the characteristics of the question setter's application of knowledge points. It has a dimension of 32 and is expanded to 128 dimensions through a fully connected layer transformation. The activation function uses the hyperbolic tangent function to obtain a query vector that matches the dimension of the hidden layer state. The similarity between the query vector and each state vector in the hidden layer state sequence is calculated using a dot product operation to obtain an attention score. Softmax normalization is performed on the attention score to obtain the attention weight, which reflects the importance of each knowledge point in forming the question preference.

[0043] The context vector is obtained by weighted summing of the hidden state sequence based on the attention weights. For example, if the attention weights of the five knowledge points in a window are [0.15, 0.25, 0.3, 0.2, 0.1], the corresponding hidden state vectors are weighted and summed according to these weights to generate a context vector that can reflect the knowledge point selection preference within the entire window. The context vector and the query vector are fused through a splicing operation and then transformed nonlinearly by a fully connected layer. The activation function uses ReLU, and the output dimension is 96. This generates a proposition preference vector that represents the question setter's knowledge point selection preference within the time window. This vector not only contains the question setter's knowledge point selection pattern information but also integrates the characteristics of knowledge application. It can comprehensively reflect the question setter's proposition preference within a specific time window.

[0044] In practical applications, such as analyzing a mathematics question setter's three-month set records, the above method can be used to obtain question preference vectors for each time window. These vectors can be further used for question setting behavior analysis, anomaly detection, or personalized question recommendation. For example, cluster analysis revealed that the setter's preferences exhibited significant regularity across different time windows, such as a preference for basic concept questions at the beginning of the month, computational application questions in the middle of the month, and comprehensive application questions at the end of the month. This provides data support for understanding the setter's behavioral patterns.

[0045] In this embodiment, by performing kernel density estimation on the proposition time information, dynamic time window division of proposition behavior is achieved, which can more accurately capture the phased change characteristics of the propositioner's behavior. The position-aware vector sequence is constructed in combination with the knowledge point selection order, and a bidirectional recurrent neural network is introduced to effectively model the contextual dependencies of knowledge points, thereby improving the ability to depict the propositioner's thinking path. At the same time, by utilizing the attention mechanism associated with the knowledge application vector, weighted extraction of the propositioner's focus of interest is achieved, thereby generating a proposition preference vector that can better reflect proposition preferences and knowledge transfer tendencies, providing more discriminative feature representations for subsequent proposition behavior analysis and personalized recognition.

[0046] In an optional embodiment, a proposition behavior recognition network is trained based on proposition preference vectors. The proposition behavior recognition network models the knowledge point selection sequence and difficulty information respectively. By optimizing the migration path and difficulty distribution in the knowledge application graph, proposition behavior features are obtained, including: A migration path identification module is constructed based on the proposition preference vector, the proposition preference vector is input into the feature extraction network to obtain initial features, a feature propagation matrix is ​​generated based on the migration weights between nodes in the knowledge application graph, and the feature propagation matrix is ​​used to aggregate the initial features to obtain the migration path features of the knowledge point selection sequence; Encoding difficulty information into a difficulty vector, calculating a correlation score between the difficulty vector and a migration path feature, and identifying a migration pattern of a knowledge point selection sequence at each difficulty level based on the correlation score; Constructing a difficulty distribution weight based on the migration pattern of the knowledge point selection sequence, performing weighted calculation on the difficulty distribution weight and the migration path feature to obtain a difficulty-adaptive migration path feature, and simultaneously correcting the difficulty vector based on the difficulty-adaptive migration path feature to obtain a corrected difficulty vector; The difficulty-adaptive migration path feature and the modified difficulty vector are subjected to feature fusion to generate a proposition behavior feature that reflects the collaborative relationship between the knowledge point selection sequence and the difficulty information.

[0047] In this implementation, a migration path identification module is first constructed based on the proposition preference vector. The core function of this module is to capture the migration patterns of the question setter during the knowledge point selection process. In specific implementation, the 96-dimensional proposition preference vector generated above is input into a feature extraction network to obtain initial features. The feature extraction network uses a multi-layer perceptron architecture, comprising three fully connected layers with 128, 64, and 32 neurons in each layer. Reluctant linear unit (ReLU) is used as the activation function, and a batch normalization layer is used to improve training stability. For example, for a mathematics question setter, their proposition preference vector is processed by the feature extraction network to obtain a 32-dimensional initial feature vector. A feature propagation matrix is ​​generated based on the migration weights between nodes in the knowledge application graph. The knowledge application graph is a pre-constructed directed graph that reflects the relationships between knowledge points, where nodes represent knowledge points and edges represent migration relationships. For example, in mathematics, there is a clear migration relationship from "differentiation of univariate functions" to "differentiation of multivariate functions," with a migration weight of 0.75, while the migration relationship from "probability theory" to "linear algebra" is weaker, with a weight of 0.15. For a knowledge application graph consisting of n core knowledge points, an n×n-dimensional feature propagation matrix is ​​generated. The matrix element values ​​are determined by the migration weights between the corresponding knowledge points. The feature propagation matrix is ​​used to aggregate the initial features to obtain the migration path features of the knowledge point selection sequence. This aggregation process utilizes a graph convolutional network mechanism. For each knowledge point node, information is aggregated based on its connection relationships with adjacent nodes. After two iterations, a 32-dimensional migration path feature vector containing information about the knowledge point migration pattern is generated.

[0048] Difficulty information is encoded as a difficulty vector. This information is derived from the difficulty level annotations of the questions in the question-setting records. Assuming a five-level difficulty scale: basic, easy, moderate, difficult, and hard, one-hot encoding is used to convert the difficulty levels into a five-dimensional difficulty vector. For question-setting records containing multiple questions, the distribution ratio of each difficulty level is calculated. For example, if the difficulty distribution of a question setter is [0.1, 0.2, 0.4, 0.2, 0.1], it means that basic questions account for 10%, easy questions account for 20%, and so on. The correlation score between the difficulty vector and the migration path features is calculated. Specifically, the difficulty vector is mapped to the same 32-dimensional space as the migration path features through a fully connected layer. The cosine similarity between the two vectors is then calculated to obtain the correlation score between each of the five difficulty levels and the migration path features. The correlation score is used to identify the migration pattern of the knowledge point selection sequence at each difficulty level. A high correlation score indicates a high correlation between the difficulty level and the specific migration pattern. In actual application, it may be found that the question setter tends to choose the migration path from "differential equations" to "complex functions" when setting "relatively difficult" and "difficult" level questions, with correlation scores of 0.78 and 0.82 respectively, while when setting "basic" and "simple" level questions, he prefers the migration path from "linear algebra" to "analytic geometry", with correlation scores of 0.85 and 0.76.

[0049] The difficulty distribution weight is constructed based on the migration pattern of the knowledge point selection sequence. The specific method is to normalize the correlation score of each difficulty level through the Softmax function to obtain a weight vector reflecting the importance of each difficulty level. The difficulty distribution weight and the migration path feature are weighted to obtain the difficulty-adaptive migration path feature. The weighted calculation uses vector element-level multiplication, that is, each dimensional component of the migration path feature is multiplied by the corresponding difficulty distribution weight to generate a 32-dimensional feature vector that can adaptively reflect the characteristics of the migration pattern under different difficulty levels. At the same time, the difficulty vector is corrected based on the difficulty-adaptive migration path feature to obtain a corrected difficulty vector. The correction process uses the attention mechanism, taking the difficulty-adaptive migration path feature as the query vector and the difficulty vector as the key-value vector, and generating a 5-dimensional corrected difficulty vector that is more in line with the actual difficulty preference of the question setter through attention calculation. For example, although a question setter claims to prefer setting questions of medium difficulty, analysis of their actual setting behavior shows that the difficulty of the questions they set on the "Probability and Statistics" knowledge point is generally higher than their self-assessed level. The revised difficulty vector may be adjusted from [0.1, 0.2, 0.4, 0.2, 0.1] to [0.05, 0.15, 0.35, 0.3, 0.15].

[0050] The difficulty-adaptive migration path features and the corrected difficulty vector are fused to generate proposition behavior features that reflect the synergistic relationship between the knowledge point selection sequence and difficulty information. Feature fusion adopts a strategy of splicing first and then transforming, splicing the 32-dimensional difficulty-adaptive migration path features and the 5-dimensional corrected difficulty vector into a 37-dimensional vector, which is then transformed into the final 48-dimensional proposition behavior feature vector through a two-layer fully connected network. The first fully connected layer outputs 64 dimensions and uses the Leaky ReLU activation function; the second layer outputs 48 dimensions and uses the Tanh activation function to ensure that the eigenvalue distribution is within the range of [-1, 1]. The final generated proposition behavior feature vector comprehensively captures the behavioral characteristics of the propositioner in the two dimensions of knowledge point selection and difficulty control, providing an effective feature representation for subsequent proposition behavior analysis, personalized proposition recommendation, and proposition quality assessment.

[0051] In a case study of the propositional behavior analysis of a mathematics teacher at a certain university, the propositional behavior characteristics extracted by the above method successfully identified that the teacher preferred a propositional pattern of gradually increasing difficulty from the shallow to the deep when propositioning in the probability theory unit, while in the numerical analysis unit, he tended to adopt a propositional strategy of mixed difficulty and highlighting key knowledge points. This personalized behavioral characteristic provides an important reference basis for the intelligent proposition assistance system.

[0052] In this embodiment, by constructing a migration path identification module and combining it with the proposition preference vector, the migration features in the knowledge point selection sequence are effectively extracted to achieve accurate modeling of the knowledge application path of the propositioner. At the same time, the difficulty information and the migration path are correlated and modeled to identify the knowledge point migration patterns at different difficulty levels, so that the model has a fine-grained understanding of the proposition strategy. By fusing the difficulty distribution and migration features, generating a difficulty adaptive representation, and correcting the difficulty vector, the model's recognition accuracy of the relationship between the question structure and cognitive load is improved. The proposition behavior characteristics finally output can accurately reflect the propositioner's collaborative behavior in knowledge transfer and difficulty control, and provide a high-quality foundation for proposition behavior prediction and anomaly detection.

[0053] like Figure 2 As shown, the proposition behavior anomaly detection and analysis process of this embodiment is demonstrated.

[0054] In an optional embodiment, the proposition behavior characteristics are input into an anomaly detection module, and the anomaly detection module calculates a warning threshold based on the migration path and difficulty distribution. When the proposition behavior characteristics deviate from the warning threshold, the module outputs warning information and the cause of the anomaly, including: Inputting the proposition behavior features into an anomaly detection module, the anomaly detection module extracts migration path information and difficulty distribution information, calculates the migration probability between knowledge points based on the migration path information to obtain a forward coupling matrix, calculates the difficulty constraint strength of knowledge point pairs based on the difficulty distribution information to obtain a reverse coupling matrix, and combines the forward coupling matrix and the reverse coupling matrix to construct a bidirectional coupling representation; Using the bidirectional coupling representation, forward propagation is performed on the migration path information to obtain path features, and reverse propagation is performed on the difficulty distribution information to obtain difficulty features. The path features and the difficulty features are interactively calculated to obtain a coupling anomaly score. Calculating a stability index of knowledge point migration based on the bidirectional coupling representation, dynamically correcting the distribution of coupling anomaly scores using the stability index to obtain a warning threshold, and triggering anomaly detection when the coupling anomaly score exceeds the warning threshold; An anomaly propagation network is constructed based on the bidirectional coupling representation. In the anomaly propagation network, the knowledge point migration sequence that causes the anomaly is located along the forward path, and the knowledge point combination that violates the difficulty constraint is identified along the reverse path. The located abnormal knowledge point migration sequence and the knowledge point combination that violates the difficulty constraint are used as the causes of the anomaly to output warning information.

[0055] Exemplarily, the anomaly detection module first extracts migration path information and difficulty distribution information from the proposition behavior characteristics. The migration path information records the conversion relationship between knowledge points in the proposition process, such as the migration frequency from "quadratic function" to "function extreme value". The difficulty distribution information contains the difficulty value of each knowledge point and its combination. Based on the migration path information, the migration probability between knowledge points is calculated, and a forward coupling matrix is ​​constructed. In specific implementation, for the migration from knowledge point A to knowledge point B, the frequency of A followed by B in the historical proposition data is divided by the total frequency of A appearing to obtain the migration probability. For example, if the frequency of "linear equation" followed by "system of two linear equations" is 80 times, and "linear equation" appears a total of 120 times, then the migration probability is 0.67.

[0056] Based on the difficulty distribution information, the difficulty constraint strength of knowledge point pairs is calculated and a reverse coupling matrix is ​​constructed. The difficulty constraint strength reflects the degree of correlation between the difficulty of two knowledge points. During implementation, the distribution of difficulty differences between knowledge point pairs in historical data is analyzed, and the inverse of the standard deviation of the difficulty difference is calculated as the constraint strength. For example, if the standard deviation of the difficulty difference between the knowledge points "derivative" and "differential" in historical data is 0.2, then the difficulty constraint strength between them is 5.

[0057] The forward coupling matrix and the reverse coupling matrix are combined to construct a bidirectional coupling representation. This representation contains two pieces of information: the transition relationship between knowledge points and the difficulty constraint relationship. In practical implementation, the bidirectional coupling representation can be stored in a three-dimensional data structure, where each pair of knowledge points has both a transition probability value and a difficulty constraint strength value. Using the bidirectional coupling representation, the transition path information is forward propagated to obtain path characteristics. During the forward propagation process, the rationality score of the path is accumulated along the sequence of knowledge point transitions in the proposition. For example, for the transition sequence "function → derivative → integral", the overall rationality of the path is calculated based on the corresponding transition probabilities in the forward coupling matrix. If the transition probability of "function → derivative" is 0.8 and the transition probability of "derivative → integral" is 0.7, the path characteristic value may be 0.56.

[0058] At the same time, backpropagation of the difficulty distribution information is performed to obtain a difficulty signature. During backpropagation, the system checks whether the difficulty differences between adjacent knowledge points in the problem are consistent with historical statistical patterns. For example, if the historical data shows a typical difficulty difference of around 0.2 from "vector" to "matrix," but the current problem shows a difference of 0.5, the system will calculate a higher difficulty outlier.

[0059] The path and difficulty features are interactively calculated to obtain the coupling anomaly score. This interactive calculation takes into account the combined impact of path rationality and difficulty rationality. For example, through weighted averaging, the path feature is given a weight of 0.6 and the difficulty feature a weight of 0.4 to obtain the final coupling anomaly score.

[0060] Based on the bidirectional coupling representation, calculate the stability index of knowledge point migration. The stability index reflects the consistency of the knowledge point migration pattern in the historical data. In implementation, the entropy of the migration probability in the forward coupling matrix can be calculated. Lower entropy values ​​indicate more stable migration patterns. For example, if the outbound migration probability distribution of the knowledge point "Equation" is concentrated and the entropy value is 0.3, then its stability index is 0.7.

[0061] The stability index is used to dynamically adjust the distribution of coupling anomaly scores to obtain a warning threshold. The strictness of the warning threshold is adjusted based on the stability index, with stricter thresholds used for knowledge point migrations with high stability. For example, for a knowledge point migration with a stability index of 0.8, the warning threshold might be set to 0.7; for a knowledge point migration with a stability index of 0.4, the warning threshold might be relaxed to 0.85. Anomaly detection is triggered when the coupling anomaly score exceeds the warning threshold.

[0062] Based on the bidirectional coupling representation, an anomaly propagation network is constructed. This network simulates the propagation of anomalies within the knowledge point graph, helping to locate the source of the anomaly. In the anomaly propagation network, the knowledge point migration sequence that leads to the anomaly is located along the forward path. For example, if an anomaly is detected containing the migration sequence "geometric proof → algebraic operation → calculus," and the forward coupling matrix shows an extremely low probability of the migration from "geometric proof → algebraic operation" (e.g., 0.05), this migration is identified as one of the causes of the anomaly. Simultaneously, along the reverse path, combinations of knowledge points that violate difficulty constraints are identified. For example, if the difficulty of "simple functions" in a proposition is set to 0.8, while the difficulty of "complex integrals" is set to 0.6, violating the constraint that the difficulty should increase step by step, this combination is identified as the cause of the anomaly.

[0063] Finally, the identified abnormal knowledge point migration sequence and the combination of knowledge points that violate the difficulty constraint are identified as the cause of the anomaly, and a warning message is generated. The warning message includes the anomaly type, location, and severity. For example, it may say, "An abnormal migration from the knowledge point 'Linear Algebra' to 'Probability Theory' was detected, with a migration probability of only 0.03. The proposition logic should be re-examined." or "The difficulty settings of the knowledge points 'Elementary Functions' (difficulty 0.9) and 'Function Graphs' (difficulty 0.4) are found to be unreasonable, violating the principle of progressive difficulty." These warning messages help teachers improve their question setting and enhance test quality.

[0064] In this embodiment, by constructing a bidirectional coupling representation that integrates the probability of knowledge transfer and the strength of difficulty constraints, the intrinsic relationship between the knowledge point selection path and difficulty control in the proposition behavior can be fully characterized. In the anomaly detection process, the interactive calculation of path characteristics and difficulty characteristics is used to generate a coupled anomaly score, and dynamic threshold correction is performed in combination with the knowledge point migration stability index, effectively improving the sensitivity and robustness of anomaly detection. Further introducing an anomaly propagation network can accurately locate the migration sequence and difficulty conflict point that triggers the anomaly, and provide a clear explanation of the cause of the anomaly, thereby achieving timely and explainable early warning of anomalies in proposition behavior and improving proposition quality monitoring and risk management capabilities.

[0065] In an optional embodiment, calculating the migration probability between knowledge points based on the migration path information to obtain a forward coupling matrix, calculating the difficulty constraint strength of knowledge point pairs based on the difficulty distribution information to obtain a reverse coupling matrix, and combining the forward coupling matrix and the reverse coupling matrix to construct a bidirectional coupling representation includes: Based on the migration sequence and temporal features in the migration path information, the number of migration occurrences between knowledge points is counted according to the time window. The migration statistics in different time windows are weighted and combined to calculate the migration probability between knowledge points to obtain a forward coupling matrix. The forward coupling matrix represents the migration probability from the source knowledge point to the target knowledge point. Based on the difficulty distribution information, a hierarchical difficulty constraint graph is constructed. In the hierarchical difficulty constraint graph, knowledge points are organized into layers according to difficulty values. The difficulty gradients of knowledge point pairs within a layer and the cross-layer constraints of knowledge point pairs between layers are calculated to obtain a reverse coupling matrix. The reverse coupling matrix represents the strength of the difficulty constraints between knowledge point pairs. The forward coupling matrix and the reverse coupling matrix are normalized, and the weighted combination coefficient of the migration probability and the difficulty constraint strength is calculated. The migration probability and the difficulty constraint strength are nonlinearly combined based on the weighted combination coefficient to construct a bidirectional coupling representation, which includes a migration probability vector, a constraint strength vector and a coupling weight coefficient.

[0066] In this embodiment, based on the migration sequence and temporal characteristics in the migration path information, the number of migrations between knowledge points is counted according to the time window. In specific implementation, the question setter's question records are grouped according to the time windows divided by the above method. Within each time window, the sequence of knowledge points selected by the question setter is extracted. For example, the sequence of knowledge points selected by a mathematics teacher in five time windows within a month is as follows: Window 1: ["Limit", "Derivative", "Integral", "Differential Equation"]; Window 2: ["Linear Algebra", "Matrix", "Eigenvalue"]; Window 3: ["Probability Theory", "Random Variable", "Mathematical Statistics"]; Window 4: ["Differential Equation", "Numerical Analysis", "Interpolation"]; Window 5: ["Complex Function", "Series Expansion", "Residue Theorem"]. For the sequence of knowledge points within each window, the number of migrations between adjacent knowledge points is counted. For example, the migration from "Limit" to "Derivative" is counted as 1, and the migration from "Derivative" to "Integral" is counted as 1. A weighted combination of migration statistics from different time windows is performed, with migration data from more recent time windows given higher weights. Assuming the weights for the five windows are 0.1, 0.15, 0.2, 0.25, and 0.3, the number of migrations between each pair of knowledge points is multiplied by the corresponding window weight and summed. The migration probabilities between knowledge points are calculated to produce a forward coupling matrix, which represents the migration probability from the source knowledge point to the target knowledge point. For n knowledge points, an n×n dimensional matrix is ​​constructed, where the matrix elements represent the probability of migration from the row-indexed knowledge point to the column-indexed knowledge point. For example, the migration probability from "Differential Equations" to "Numerical Analysis" might be 0.75, indicating that after selecting "Differential Equations," the question setter has a 75% chance of selecting "Numerical Analysis" as the next knowledge point.

[0067] A hierarchical difficulty constraint graph is constructed based on difficulty distribution information. This information is derived from the difficulty statistics of each knowledge point in the question record. First, the average difficulty value of each knowledge point is calculated. For example, the average difficulty value of "Limit" is 3.2, "Derivative" is 3.5, and "Integral" is 4.1 (assuming a difficulty range of 1-5). In the hierarchical difficulty constraint graph, knowledge points are organized into layers according to their difficulty value. For example, difficulty values ​​of 1-1.9 are divided into the first layer, 2-2.9 into the second layer, and so on, for a total of five layers. The difficulty gradients of knowledge point pairs within a layer and the cross-layer constraints of knowledge point pairs between layers are calculated to obtain the reverse coupling matrix. The intra-layer difficulty gradient is calculated by calculating the absolute value of the difference in difficulty values ​​for knowledge point pairs within the same layer. The smaller the difference, the stronger the constraint strength. For example, if "Matrix" and "Eigenvalue" are both located in the third layer, with difficulty values ​​of 3.3 and 3.4, respectively, the difficulty gradient between them is 0.1, and the corresponding constraint strength may be 0.9. Inter-layer and cross-layer constraints are calculated by calculating the constraint strength for pairs of knowledge points at different layers based on the difference in their level and difficulty values. The greater the difference in level and difficulty value, the weaker the constraint strength. For example, the level difference between "Limit" (layer 3) and "Residue Theorem" (layer 5) is 2, and the difficulty difference is 1.7, so the corresponding constraint strength might be 0.3. The reverse coupling matrix is ​​also an n×n matrix, with the matrix element values ​​representing the difficulty constraint strength between the row-indexed knowledge point and the column-indexed knowledge point.

[0068] The forward and reverse coupling matrices are normalized to ensure that the matrix element values ​​are distributed within the range of 0-1 and that the sum of the elements in each row is 1. Row normalization is performed, dividing each row element by the sum of all elements in that row. A weighted combination coefficient is calculated for the transfer probability and the difficulty constraint strength. This coefficient determines the relative importance of the transfer probability and the difficulty constraint strength in the bidirectional coupling representation. The calculation of the combination coefficient takes into account the characteristics of the question setter's question-setting behavior, such as their preference for question coherence and difficulty control. In specific implementation, the combination coefficient is calculated based on the components of the question-setting behavior feature vector related to coherence and difficulty control. For example, for a question setter who prioritizes knowledge point coherence, the weight for transfer probability might be 0.7, and the weight for difficulty constraint strength might be 0.3. For a question setter who prioritizes difficulty balance, the weight for transfer probability might be 0.4, and the weight for difficulty constraint strength might be 0.6.

[0069] The bidirectional coupling representation is constructed by nonlinearly combining the migration probability and difficulty constraint strength based on a weighted combination coefficient. This nonlinear combination employs a weighted harmonic mean method. For each pair of knowledge points, the bidirectional coupling strength is calculated based on the migration probability, difficulty constraint strength, and combination coefficient. The bidirectional coupling representation consists of a migration probability vector, a constraint strength vector, and a coupling weight coefficient. For n knowledge points, the migration probability vector and the constraint strength vector each have n×n elements, corresponding to the values ​​of the forward and reverse coupling matrices, respectively. The coupling weight coefficient is a scalar value that represents the relative importance of the migration probability and difficulty constraint strength.

[0070] This bidirectional coupling effectively captures the complex association patterns between knowledge points, which are influenced by both transfer logic and difficulty control constraints. This provides an important basis for building a proposition recommendation system that better reflects the actual behavior of proposition setters. For example, for teachers who are accustomed to setting questions progressively from the basics to applications, the system will prioritize knowledge point combinations with high transfer probability and moderate difficulty gradients. For teachers who prefer challenging propositions, the system will recommend knowledge point combinations with a higher difficulty span but still maintaining a certain transfer logic.

[0071] In this embodiment, by fusing migration path information with difficulty distribution information, a bidirectional coupling representation is constructed to achieve accurate modeling of the synergistic relationship between the logical migration of knowledge points and difficulty control in propositional behavior. The forward coupling matrix characterizes the migration tendency between knowledge points, and the reverse coupling matrix reflects the strength of difficulty constraints. The two are normalized and a weighted combination mechanism is introduced, so that the coupling representation can adapt to the laws of knowledge migration while taking into account difficulty gradient control, effectively improving the accuracy and discrimination ability of behavior modeling. This method can provide a richer and more interpretable semantic feature foundation for subsequent anomaly detection and behavior recognition, thereby improving the system's modeling and analysis capabilities for complex propositional behaviors.

[0072] A second aspect of an embodiment of the present invention provides a system for constructing a behavior model of an examination question setter based on deep learning, the system comprising: The first unit is used to obtain the question setting records of the test setter and construct a knowledge association analysis module. The knowledge association analysis module calculates the usage frequency based on the knowledge point tags in the question setting records, constructs a knowledge point mapping network based on a preset knowledge system, obtains migration weights by calculating the conditional probabilities between knowledge points, combines the knowledge point mapping network and the migration weights to generate a knowledge application graph, and outputs a knowledge application vector based on the knowledge application graph; The second unit is used to build a dynamic preference analysis module based on the knowledge application vector. The dynamic preference analysis module groups the proposition behavior into time windows according to the time information in the proposition record, extracts features of the knowledge point selection sequence within each time window, and generates a proposition preference vector based on the difficulty information; The third unit is used to train a proposition behavior recognition network based on the proposition preference vector. The proposition behavior recognition network models the knowledge point selection sequence and difficulty information respectively, and obtains proposition behavior characteristics by optimizing the migration path and difficulty distribution in the knowledge application graph; The fourth unit is used to input the proposition behavior characteristics into the anomaly detection module. The anomaly detection module calculates the warning threshold according to the migration path and difficulty distribution, and outputs warning information and the cause of the anomaly when the proposition behavior characteristics deviate from the warning threshold.

[0073] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0074] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0075] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0076] 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a behavior model of test setters based on deep learning, characterized by: include: Obtaining the test setter's question setting records, constructing a knowledge association analysis module, calculating the usage frequency based on the knowledge point tags in the question setting records, constructing a knowledge point mapping network based on a preset knowledge system, obtaining migration weights by calculating the conditional probabilities between knowledge points, combining the knowledge point mapping network and the migration weights to generate a knowledge application graph, and outputting a knowledge application vector based on the knowledge application graph; A dynamic preference analysis module is constructed based on the knowledge application vector. The dynamic preference analysis module groups proposition behaviors into time windows according to the time information in the proposition record, extracts features of the knowledge point selection sequence within each time window, and generates a proposition preference vector based on the difficulty information. A proposition behavior recognition network is trained based on proposition preference vectors. The proposition behavior recognition network models the knowledge point selection sequence and difficulty information respectively, and obtains proposition behavior features by optimizing the migration path and difficulty distribution in the knowledge application graph. The proposition behavior characteristics are input into the anomaly detection module, and the anomaly detection module calculates the warning threshold according to the migration path and difficulty distribution, and outputs warning information and the cause of the anomaly when the proposition behavior characteristics deviate from the warning threshold.

2. The method according to claim 1, characterized in that Constructing a knowledge association analysis module, wherein the knowledge association analysis module calculates the usage frequency according to the knowledge point tags in the proposition record and constructs a knowledge point mapping network based on a preset knowledge system, including: Convert the proposition text of the proposition record into a word vector sequence, perform bidirectional attention calculation on the query vector and the word vector sequence, and identify candidate knowledge point segments based on the attention distribution; Performing semantic vectorization on the candidate knowledge point segments, calculating the semantic similarity between the semantic vectorization and the standard knowledge points in the preset knowledge system, and extracting the corresponding knowledge point labels based on the similarity matching results; Calculate the direct usage frequency of knowledge point tags in the proposition record. At the same time, based on the hierarchical structure of the preset knowledge system, count the usage frequencies of the parent node and child node corresponding to each knowledge point tag in the proposition record and perform weighted combination to obtain the hierarchical usage frequency of knowledge points containing hierarchical relationships. Using knowledge point labels as nodes and hierarchical usage frequencies as initial node weights, node connections are established based on the dependencies between knowledge points in the preset knowledge system to construct an initial knowledge point mapping network. The order of knowledge point selection in the proposition process is extracted from the proposition records. The migration rules between knowledge points are analyzed based on the knowledge point selection order, and the correlation between nodes is calculated based on the migration rules. The connection strengths between nodes in the initial knowledge point mapping network are updated according to the degree of correlation, and feature extraction is performed on the updated node weights to finally construct a knowledge point mapping network that represents the association relationship between knowledge points.

3. The method according to claim 1, characterized in that The conditional probability between knowledge points is calculated to obtain the transfer weight, and the knowledge point mapping network and the transfer weight are combined to generate a knowledge application graph. The knowledge application vector output based on the knowledge application graph includes: Obtain a knowledge point selection sequence in the proposition record, set multiple sliding time windows of different scales for the knowledge point selection sequence, count the number of occurrences of knowledge points and the number of consecutive occurrences of knowledge point pairs in each sliding time window, calculate the conditional probability of the knowledge points in the current window, and perform a weighted combination of the conditional probabilities of the knowledge points in different sliding time windows to obtain the migration weights between the knowledge points; Combining the migration weights between knowledge points with the knowledge point mapping network to generate a knowledge application graph, wherein the knowledge application graph includes nodes, connection relationships between nodes, and migration weights corresponding to the connection relationships; Performing a transformation matrix transformation on the node feature vectors in the knowledge application graph to obtain a transformed feature vector of the node, performing a splicing operation on the transformed feature vectors of adjacent nodes and processing them with an activation function, obtaining an attention coefficient between the nodes through a normalization operation, and performing weighted aggregation on the transformed feature vectors of adjacent nodes based on the attention coefficient to obtain an updated feature vector that integrates feature information of adjacent nodes; An average pooling operation is performed on the updated feature vectors of all nodes in the knowledge application graph to generate a knowledge application vector, which is used to characterize the knowledge point selection pattern and migration law.

4. The method according to claim 1, wherein According to the time information in the proposition record, the proposition behavior is grouped into time windows, and the feature extraction of the knowledge point selection sequence in each time window is performed. The proposition preference vector is generated by combining the difficulty information, including: Constructing a time series for the time information in the proposition record, performing kernel density estimation on the time series to obtain a proposition activity density curve, dynamically calculating the time window size based on the proposition activity density curve, and obtaining the time window division of the proposition behavior; Based on the time window division, extract the knowledge point selection sequence within each time window from the proposition record, construct a knowledge point word vector mapping table, map the knowledge point selection sequence within each time window into a word vector sequence, and incorporate relative position encoding information to obtain a position-aware vector sequence; Inputting the position-aware vector sequence into a bidirectional recurrent neural network, extracting context-dependent features of knowledge point selection to obtain a hidden layer state sequence; The pre-acquired knowledge application vector is transformed nonlinearly to generate a query vector, the similarity between the query vector and the hidden layer state sequence is calculated to obtain an attention score, and the attention score is normalized to obtain an attention weight; The context vector is obtained by weighted summing of the hidden layer state sequence based on the attention weight, and the context vector is fused with the query vector and then subjected to nonlinear transformation to generate a proposition preference vector that represents the knowledge point selection preference of the propositioner within the time window.

5. The method according to claim 1, wherein The proposition behavior recognition network is trained based on the proposition preference vector. The proposition behavior recognition network models the knowledge point selection sequence and difficulty information respectively. By optimizing the migration path and difficulty distribution in the knowledge application graph, the proposition behavior features obtained include: A migration path identification module is constructed based on the proposition preference vector, the proposition preference vector is input into the feature extraction network to obtain initial features, a feature propagation matrix is ​​generated based on the migration weights between nodes in the knowledge application graph, and the feature propagation matrix is ​​used to aggregate the initial features to obtain the migration path features of the knowledge point selection sequence; Encoding difficulty information into a difficulty vector, calculating a correlation score between the difficulty vector and a migration path feature, and identifying a migration pattern of a knowledge point selection sequence at each difficulty level based on the correlation score; Constructing a difficulty distribution weight based on the migration pattern of the knowledge point selection sequence, performing weighted calculation on the difficulty distribution weight and the migration path feature to obtain a difficulty-adaptive migration path feature, and simultaneously correcting the difficulty vector based on the difficulty-adaptive migration path feature to obtain a corrected difficulty vector; The difficulty-adaptive migration path feature and the modified difficulty vector are subjected to feature fusion to generate a proposition behavior feature that reflects the collaborative relationship between the knowledge point selection sequence and the difficulty information.

6. The method according to claim 1, characterized in that The proposition behavior characteristics are input into the anomaly detection module, which calculates the warning threshold based on the migration path and difficulty distribution. When the proposition behavior characteristics deviate from the warning threshold, the module outputs warning information and the cause of the anomaly, including: Inputting the proposition behavior features into an anomaly detection module, the anomaly detection module extracts migration path information and difficulty distribution information, calculates the migration probability between knowledge points based on the migration path information to obtain a forward coupling matrix, calculates the difficulty constraint strength of knowledge point pairs based on the difficulty distribution information to obtain a reverse coupling matrix, and combines the forward coupling matrix and the reverse coupling matrix to construct a bidirectional coupling representation; Using the bidirectional coupling representation, forward propagation is performed on the migration path information to obtain path features, and reverse propagation is performed on the difficulty distribution information to obtain difficulty features. The path features and the difficulty features are interactively calculated to obtain a coupling anomaly score. Calculating a stability index of knowledge point migration based on the bidirectional coupling representation, dynamically correcting the distribution of coupling anomaly scores using the stability index to obtain a warning threshold, and triggering anomaly detection when the coupling anomaly score exceeds the warning threshold; An anomaly propagation network is constructed based on the bidirectional coupling representation. In the anomaly propagation network, the knowledge point migration sequence that causes the anomaly is located along the forward path, and the knowledge point combination that violates the difficulty constraint is identified along the reverse path. The located abnormal knowledge point migration sequence and the knowledge point combination that violates the difficulty constraint are used as the causes of the anomaly to output warning information.

7. The method according to claim 6, characterized in that Calculating the migration probability between knowledge points based on the migration path information to obtain a forward coupling matrix, calculating the difficulty constraint strength of knowledge point pairs based on the difficulty distribution information to obtain a reverse coupling matrix, and combining the forward coupling matrix and the reverse coupling matrix to construct a bidirectional coupling representation includes: Based on the migration sequence and temporal features in the migration path information, the number of migration occurrences between knowledge points is counted according to the time window. The migration statistics in different time windows are weighted and combined to calculate the migration probability between knowledge points to obtain a forward coupling matrix. The forward coupling matrix represents the migration probability from the source knowledge point to the target knowledge point. Based on the difficulty distribution information, a hierarchical difficulty constraint graph is constructed. In the hierarchical difficulty constraint graph, knowledge points are organized into layers according to difficulty values. The difficulty gradients of knowledge point pairs within a layer and the cross-layer constraints of knowledge point pairs between layers are calculated to obtain a reverse coupling matrix. The reverse coupling matrix represents the strength of the difficulty constraints between knowledge point pairs. The forward coupling matrix and the reverse coupling matrix are normalized, and the weighted combination coefficient of the migration probability and the difficulty constraint strength is calculated. The migration probability and the difficulty constraint strength are nonlinearly combined based on the weighted combination coefficient to construct a bidirectional coupling representation, which includes a migration probability vector, a constraint strength vector and a coupling weight coefficient.

8. A system for constructing a behavior model of test question setters based on deep learning, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the question setting records of the test setter and construct a knowledge association analysis module based on the question setting records. The knowledge association analysis module calculates the usage frequency based on the knowledge point tags in the question setting records, constructs a knowledge point mapping network based on a preset knowledge system, calculates the conditional probabilities between knowledge points to obtain migration weights, combines the knowledge point mapping network and the migration weights to generate a knowledge application graph, and outputs a knowledge application vector based on the knowledge application graph; The second unit is used to build a dynamic preference analysis module based on the knowledge application vector. The dynamic preference analysis module groups the proposition behavior into time windows according to the time information in the proposition record, extracts features of the knowledge point selection sequence within each time window, and generates a proposition preference vector based on the difficulty information; The third unit is used to train a proposition behavior recognition network based on the proposition preference vector. The proposition behavior recognition network models the knowledge point selection sequence and difficulty information respectively, and obtains proposition behavior characteristics by optimizing the migration path and difficulty distribution in the knowledge application graph; The fourth unit is used to input the proposition behavior characteristics into the anomaly detection module. The anomaly detection module calculates the warning threshold according to the migration path and difficulty distribution, and outputs warning information and the cause of the anomaly when the proposition behavior characteristics deviate from the warning threshold.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images