A knowledge tracking method based on multi-view graph contrast learning

Abstract

This invention provides a knowledge tracking method based on multi-view graph contrastive learning, comprising: constructing a multi-view heterogeneous graph; using a time-aware graph convolutional network to temporally encode the multi-view heterogeneous graph, and outputting the node representation matrix of each viewpoint by incorporating a time decay factor into the message passing process of graph convolution; performing intra-view comparison, inter-view comparison, and student-level comparison on the node representation matrix of each viewpoint sequentially through a cross-view contrastive alignment method to eliminate the semantic gap caused by the multi-view graph construction, and obtaining aligned multi-view node representations; using an attention mechanism to dynamically fuse the aligned multi-view node representations to obtain a multi-view fused representation of the question; inputting the multi-view fused representation corresponding to the student's historical answer sequence into a gated recurrent unit for sequence state encoding to obtain the student's knowledge state, and then predicting the probability of correct answer. This invention significantly improves the prediction accuracy and robustness of knowledge tracking.

Description

Technical Field

[0001] This invention relates to the field of knowledge tracking technology, specifically providing a knowledge tracking method based on multi-view graph comparison learning. Background Technology

[0002] With the rapid development of artificial intelligence technology, online education has become one of the important forms of education. In intelligent tutoring systems, Knowledge Tracking (KT) technology aims to analyze students' historical answer records, model their mastery of various knowledge points, and predict their future learning performance, thereby providing data support for personalized teaching for different students. Specifically, Knowledge Tracking is essentially a supervised sequence prediction problem: given a student's answers to a sequence of historical questions, predict their accuracy on the next question. This technology lays the core foundation for personalized education applications such as learning resource recommendation, adaptive testing, and educational games. In recent years, the accumulation of massive amounts of educational data has endowed deep learning models with extremely rich statistical prior features, while also posing unprecedented engineering challenges to the feature fusion dimensions and temporal modeling of the underlying Knowledge Tracking architecture.

[0003] Currently, existing technologies in the field of knowledge tracing have primarily evolved from traditional probabilistic models to deep learning data-driven models. Models like Deep Knowledge Tracing (DKT) were the first to use Long Short-Term Memory (LSTM) networks instead of Hidden Markov Chains, achieving end-to-end feature extraction and continuous modeling and updating of students' implicit knowledge states. Subsequently, models based on self-attention mechanisms and Transformer architectures (such as SAKT, SAINT, and AKT) have been widely applied. These models utilize self-attention mechanisms to capture the correlation between practice and response from students' past interactions with exercises, thereby enabling predictions based on these knowledge states.

[0004] Considering the inherent structure between knowledge concepts, some existing technologies (such as GKT and GIKT) propose to represent knowledge concepts using graphs, establishing a correspondence graph between test questions and knowledge concepts. These models utilize graph convolutional neural networks (GCNs) to delve into the deep relationships between questions and between questions and concepts, perform feature aggregation and updates on the graph structure, and ultimately predict the probability of a student answering a question correctly in the next time step.

[0005] The current method has the following main drawbacks: (1) Over-reliance on individual historical answer sequences for independent modeling ignores the implicit group structured relationship between questions, resulting in poor generalization ability of the model in data sparsity and cold start scenarios.

[0006] (2) Existing graph neural network models are generally limited by the static graph assumption and fail to effectively introduce the time decay mechanism, making it difficult to accurately depict the memory forgetting pattern in the learning process of students.

[0007] (3) When fusing multi-source heterogeneous features, there is a lack of effective representation alignment constraints, making it difficult to eliminate the semantic gap between features. Furthermore, the model is easily affected by noise from students' random guesses or mistakes, resulting in insufficient robustness. Summary of the Invention

[0008] To overcome the above-mentioned shortcomings, this invention is proposed to provide solutions or at least partially solve the problems of insufficient modeling of structured associations between questions in data sparse scenarios, lack of time decay mechanism for graph feature propagation, and semantic gap of multi-source features.

[0009] This invention provides a knowledge tracking method based on multi-view graph contrast learning, comprising the following steps: A multi-perspective heterogeneous graph is constructed, including: a question-knowledge point bipartite graph, a question co-answer graph, and a time evolution graph; the question-knowledge point bipartite graph is constructed based on the explicit association between questions and knowledge points, and is used to represent the explicit association perspective; the question co-answer graph is constructed based on the objective difficulty similarity relationship of questions mined from group answer data, and is used to represent the implicit difficulty perspective; the time evolution graph is constructed based on the interaction relationship between questions and time windows, and is used to represent the macro-evolutionary perspective; The multi-view heterogeneous graph is temporally encoded using a time-aware graph convolutional network. By incorporating a time decay factor into the message passing process of graph convolution, the representation matrix of each view node is output. By using a cross-view comparison alignment method, the node representation matrices of each view are sequentially compared within the view, between the view, and at the student level to eliminate the semantic gap caused by multi-view composition and obtain the aligned multi-view node representations. An attention mechanism is used to dynamically fuse the aligned multi-view node representations to obtain a multi-view fused representation of the problem. The multi-perspective fusion representation of each question in the student's historical answer sequence is input into the gated recurrent unit for sequence state encoding to obtain the student's current knowledge state. Combined with the embedding of the question to be predicted, the probability of the student answering the next question correctly is predicted.

[0010] Preferably, the construction of the question co-answer graph includes: For each question, the percentage of students who answered the question correctly was calculated as the objective difficulty of the question for that group. Calculate the difficulty difference between any two questions in the question bank, select a preset number of other questions with the smallest difficulty difference as difficulty neighbors for each question, and establish undirected connecting edges between questions that are difficulty neighbors. The question co-answer graph is composed of all question nodes and the undirected connecting edges.

[0011] Preferably, the time evolution graph is represented by a quartic adjacency matrix; in the quartic adjacency matrix: The top-left corner is a zero matrix, indicating that there are no direct connections between the problem nodes from a macroscopic evolution perspective; The top right corner block records the normalized interaction frequency of each question within each time window; The bottom left corner piece is the transpose of the top right corner piece; The bottom right block is a strip matrix with a value of 1 assigned only at the corresponding positions of adjacent time windows, used to represent the sequential evolution relationship of time windows.

[0012] Preferably, in the time-aware graph convolutional network, the time decay factor is incorporated into the message passing process of graph convolution in the following way: Calculate the time difference matrix between nodes, and map the time difference matrix into a decay factor using a preset time decay function; The attenuation factor is multiplied element-wise with the original adjacency matrix of the current viewpoint to obtain the time-aware adjacency matrix. Multi-layer graph convolution propagation is performed based on the time-aware adjacency matrix, so that the aggregation weight of node features is adaptively decayed according to the time distance between nodes.

[0013] Preferably, a time-aware adjacency matrix : In the formula, This is the original adjacency matrix. This is a preset time sensitivity hyperparameter used to control the decay rate; Time difference matrix It is the identity matrix; Represents the Hadamard product; Multi-layer graph convolutional propagation is performed according to the following formula: In the formula, This is the time perception matrix. For the first The trainable weight matrix of the layer; This represents the maximum number of convolutional layers. It is a ReLU nonlinear activation function. For the first The node feature matrix of the layer, when hour, The initial characteristic matrix .

[0014] Preferably, the intra-view comparison includes: For each question node, perform two independent random perturbations to generate two enhanced views that form a positive sample pair; Randomly select the representations of other question nodes within the current training batch to form negative sample pairs; The in-view contrast loss is calculated based on the positive and negative sample pairs, and then summed after calculating the losses for the three perspectives of explicit association, implicit difficulty, and macro-evolution to obtain the total in-view loss.

[0015] Preferably, the comparison between viewpoints includes: Pair the representation vectors of the same question from the perspectives of explicit association, implicit difficulty, and macro-evolution to form positive sample pairs; The representation of the current question from one perspective is compared with the representations of other questions in the batch from another perspective to form a negative sample pair; The comparative loss is calculated and summed for the explicit correlation perspective and the implicit difficulty perspective, the explicit correlation perspective and the macro-evolutionary perspective, and the implicit difficulty perspective and the macro-evolutionary perspective, respectively, to obtain the total loss between perspectives.

[0016] Preferably, the student-level comparison includes: Two independent random subsequence sampling operations are performed on the current student's historical answer sequence. The feature vectors of the corresponding questions are extracted from the node representation matrix of each viewpoint and then pooled and concatenated to obtain two subsequence augmented views as positive sample pairs. The complete sequences of other students in the current training batch are combined as negative samples. The student-level contrastive loss is calculated based on the positive and negative sample pairs.

[0017] Preferably, the total loss within the viewpoint, the total loss between viewpoints, and the student-level contrast loss are weighted and summed to construct a joint contrast loss; By minimizing the joint contrastive loss through backpropagation, the weight parameters of the time-aware graph convolutional network and the initial embedding matrix of all nodes are updated synchronously to obtain the aligned multi-view node representation.

[0018] Preferably, the timestamp of the node is determined in the following manner: The timestamp of the question node represents the average time it took for the student group to answer the question. The timestamp of a knowledge point node represents the average answering time for all questions associated with it. The timestamp of the supernode of the time window is the center time point of that time window.

[0019] The beneficial effects of this invention are as follows: (1) Effectively overcome the bottlenecks of data sparsity and cold start. By constructing a question co-answer graph and a time evolution graph, the similarity of difficulty and the macro-time evolution law in the group's answering behavior are transformed into graph structure prior information, so that the model can transfer knowledge with the help of the group structure when individual answer records are sparse, and significantly improve the generalization ability.

[0020] (2) Achieve deep coupling between temporal features and graph topology. The time-aware graph convolutional network directly applies the time decay factor to the message propagation process of graph convolution, which is different from the existing technology that applies a forget gate in the sequence coding layer or a temporal bias in the attention layer. This allows the aggregation weights of graph convolution to decay adaptively according to the time distance between nodes, naturally following the memory forgetting law in cognitive psychology, and more accurately depicting the dynamic evolution of students' knowledge status.

[0021] (3) Eliminate the semantic gap between multiple perspectives and enhance the robustness of the model. Through a three-layer comparative learning mechanism of intra-perspective perturbation comparison, inter-perspective semantic alignment comparison and student-level sequence consistency comparison, the representations of different graph perspectives are forced to align in high-dimensional space without increasing the complexity of the network structure, which effectively suppresses the interference of noise from students' random guesses, mistakes and other behaviors.

[0022] (4) High computational efficiency and easy engineering deployment. Cross-view comparison alignment only imposes constraints during the training phase and does not add extra computational overhead during inference; the prediction module adopts lightweight gated recurrent units and linear mappings, which are suitable for the real-time response requirements of large-scale online education scenarios. Attached Figure Description

[0023] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is a flowchart illustrating a knowledge tracking method based on multi-view graph contrast learning according to an embodiment of the present invention.

[0024] Figure 2 This is a structural diagram of the MGCLKT model of an example of the present invention. Detailed Implementation

[0025] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0026] like Figure 1-2As shown, this invention provides a knowledge tracking method based on multi-view graph contrast learning, comprising the following steps: Step S1: Construct a multi-perspective heterogeneous graph, including: a question-knowledge point bipartite graph, a question co-answer graph, and a time evolution graph; the question-knowledge point bipartite graph is constructed based on the explicit association between questions and knowledge points and is used to represent the explicit association perspective; the question co-answer graph is constructed based on the objective difficulty similarity relationship of questions mined from group answer data and is used to represent the implicit difficulty perspective; the time evolution graph is constructed based on the interaction relationship between questions and time windows and is used to represent the macro-evolution perspective.

[0027] In this embodiment, the question co-answer graph is an isomorphic graph structure that contains only question nodes as a set of vertices. The undirected edges in the graph are used to represent the implicit similarity between different questions in terms of objective difficulty. By constructing this graph, the originally isolated question entities can be clustered and connected based on the extensive answering performance of historical student groups, thereby transforming the answering behavior characteristics of a massive number of individuals into structured graph prior information of the question bank difficulty dimension.

[0028] Constructing a bipartite graph of questions and knowledge points The process is as follows: First, obtain the corresponding annotation information of questions and knowledge points in the existing dataset, and the vertex set. ,in To correspond to the set of nodes in the question, A set of knowledge point nodes; in any question node With any knowledge point node it contains Establish undirected edges between them, and all undirected edges constitute edge set ; Subsequently, the representation is constructed. Adjacency matrix of connections between nodes ,in The row dimension corresponds to the set of question nodes. , The column dimensions correspond to the set of knowledge point nodes. If the question Includes knowledge points ,but The Middle Line number Column elements Assigned value Otherwise, assign a value . Used to guide the two-way transfer of information between questions and knowledge points in graph convolution propagation.

[0029] In this context, a dataset refers to a collection of structured educational interaction data gathered from online education platforms or intelligent tutoring systems to record student learning behaviors and question attribute characteristics. Such datasets contain two main parts: first, structural mapping information, meaning the dataset explicitly includes each question node from the question bank. And the specific knowledge points and nodes being examined The correspondence between them; and the dynamic interaction record, which includes a sequence of behaviors such as student ID, the question answered, whether the answer is correct or incorrect, the time taken to answer the question and the timestamp.

[0030] The specific datasets applicable to this invention include, but are not limited to, standard public datasets in the field of knowledge tracing, such as the ASSISTments series datasets (e.g., ASSIST2009, ASSIST2012), the EdNet dataset, the Junyi Academy dataset, and NIPS-EduCup, etc.

[0031] For example: Suppose there are 3 question nodes in the current educational scenario. and 2 knowledge point nodes Among them, the question This is a "fraction addition problem," which includes the following knowledge points. (Addition); Problem This is a "decimal subtraction word problem," which includes the following knowledge points. (Subtraction); This is a "mixed addition and subtraction problem," which includes multiple knowledge points. and Then a bipartite graph of question and knowledge point can be constructed. The set of points Undirected edge set .

[0032] and Corresponding adjacency matrix As shown below: In the In the middle, the first line Indicates the topic Only related to knowledge points Explicit associations exist. In the subsequent graph convolutional propagation process, this adjacency matrix specifically acts as the topological connectivity path and aggregation mask for information transfer between nodes. A matrix element of 1 represents a connected path, and 0 represents a blocked path; this structural constraint makes... The characteristics can spread to ,at the same time The aggregation characteristics can also be fed back to This enables feature interaction between the semantics of the question and the attributes of the knowledge points.

[0033] In one specific implementation, the construction process of the question co-answer graph is as follows: First, the objective difficulty of the questions is calculated based on the group's answering behavior. For each question... Count the set of students who answered the question. In the middle, each student Correct and incorrect answer record If the student answers correctly, then ,otherwise ; Calculation problems objective difficulty of the group ; Next, calculate from the complete set of problems. With any other question Difference in difficulty And select the one with the smallest difference in difficulty. Other questions as The neighboring nodes, in and stated Establish undirected edges between each of the other problems to form an edge set. ; Construct a question co-answer graph ,in The point set contains only the problem node set. At the same time, an adjacency matrix is ​​generated. When the question and When elements are neighbors of different difficulty levels, corresponding elements Assigned value ,otherwise, The remaining elements in the array have a value of 0.

[0034] The full question set refers to the collection of all deduplicated question nodes extracted from the aforementioned dataset, covering all questions within the current education system. It forms the logical boundary for constructing an implicit difficulty perspective, ensuring that the model can establish connections by comparing objective difficulty differences between questions across the global question bank.

[0035] For example: Suppose the dataset in the current educational scenario contains 4 question nodes. Based on statistical calculations of historical answer records, the objective difficulty levels for each question were determined as follows: , , , The higher the value, the higher the correct answer rate, meaning the simpler and less difficult the question.

[0036] To simplify the description, the difficulty neighbor parameter is now set. .for The question with the smallest difference in difficulty is Therefore, undirected edges are constructed in the graph. Similarly, for Its nearest neighbor is ;for The nearest neighbor is Therefore, a question co-answer graph was constructed. Two distinct clusters were automatically formed: The easy problem clusters and The resulting cluster of problems.

[0037] Corresponding adjacency matrix As shown below: In the In, elements and The system records the high degree of difficulty similarity between pairs of questions. In subsequent feature propagation, this implicit graph topology guides the model to direct students' choices of questions. The cognitive state is smoothly transferred to problems of similar difficulty that have not been encountered before. This effectively addresses the issues of data sparsity and cold start.

[0038] A temporal evolution graph is a heterogeneous graph structure that includes question nodes and time window supernodes. The edges in the graph represent the activity level of questions in a specific macro-time stage and the sequential evolutionary relationship between adjacent time stages. By constructing a temporal evolution graph, static question entities can be projected into a dynamic temporal context, capturing the systematic drift and evolution of the group's knowledge state and the objective difficulty of questions as a result of the macro-teaching process (such as the beginning of the semester, mid-term review, and final exam preparation).

[0039] In one specific implementation, the construction process of the time evolution graph is as follows: First, divide the entire interaction timeline into... Define a set of supernodes for each consecutive time window with an equal number of interactions. ; Constructing a time evolution graph ,in, The point set contains the set of question nodes. and time window supernode set , The edge set contains the interaction edge set between the question and the window. and the evolutionary edge set between adjacent windows If the question In the Time window If there are answer records in memory, then... With the corresponding time window super node Establish an undirected edge between them. Simultaneously, within any adjacent time window, the supernode... and Establish undirected edges between them.

[0040] Using a four-block adjacency matrix Representing the time evolution diagram , The top left corner block is Top right corner block Records the normalized activity frequency of questions in each window; the bottom left corner is a matrix. transpose matrix Used to characterize the symmetric interaction relationships in an undirected graph; bottom right corner block Set the elements on both sides of the main diagonal as A strip matrix is ​​used to define the cascaded evolution path of the time window.

[0041] The interaction timeline refers to the timeline formed by arranging all student-question interaction events (i.e., answer records) covered in the dataset in chronological order of their occurrence. It serves as a benchmark for dividing macroscopic time windows and aims to provide a dynamic time reference coordinate system for static question entities, thereby enabling the model to quantify and capture the systematic evolution of group cognitive states and question attributes as the teaching process (such as semester cycle and course progress) progresses.

[0042] Time window supernode: In constructing a time evolution graph, a global virtual entity node is artificially introduced to represent a specific macro-time span (such as the beginning of a semester, mid-term, or a specific review phase). Unlike "question nodes" representing specific test content, supernodes act as topological hubs that aggregate all interactive information within that specific time period. By connecting all question nodes with recorded responses within a time window to their corresponding time window supernodes, interactive events that were originally discretely distributed along the timeline can be aggregated in the graph space. This allows the model to transcend individual differences and directly capture the collective activity of questions, as well as the time-varying and dynamic evolution of knowledge difficulty as the macro-teaching process progresses.

[0043] For example: Divide the complete teaching cycle of a course into 3 consecutive time windows (e.g., the beginning of the semester). Midterm End of term ), to obtain the set of supernodes for the time window Meanwhile, it is assumed that the current question bank contains 3 question nodes. .

[0044] After statistical extraction of global historical interaction records: (Question) exist and All stages were answered extensively by students. The normalized activity frequencies for these two windows were 0.6 and 0.4, respectively; (Question) Only during the period The events occurred in concentrated phases, with a normalized frequency of 1.0; (Question) As a final exam question, only for the end of the semester. The phase was answered, with a normalized frequency of 1.0.

[0045] Based on this, construct cross-boundary interaction edge sets for heterogeneous graphs. The corresponding activity block matrix in the upper right corner and the symmetric block matrix in the lower left corner At the same time, to ensure the continuity of temporal evolution, in and Establish evolutionary edge sets between them The corresponding diagonal strip matrix in the lower right corner .

[0046] The final composition Dimensional time evolution block adjacency matrix As shown below: In the block adjacency matrix In the middle, the top left corner A zero matrix indicates that there are no direct edges between pure problem nodes in this specific view; the upper right part of the matrix (such as 0.6 and 0.4 in the fourth and fifth columns of the first row) accurately records the problem. The lifecycle spans the first two time phases; the symmetric submatrix in the lower right corner clarifies the sequential connection of the teaching process on the time axis by establishing bidirectional edges between adjacent time phases (such as the beginning and middle of the semester, and the middle and end of the semester), thus representing the continuous evolution of the time state. This block matrix directly transforms the macroscopic time pattern into the topological constraints of the graph, pre-setting a clear information transmission path for the subsequent time-aware graph convolutional network, enabling it to automatically follow the time logic during feature aggregation and more accurately capture the dynamic changes in students' knowledge mastery level over time.

[0047] Of course, multi-perspective heterogeneous graphs are not limited to the three graph structures mentioned above. Other perspective graph structures (such as student collaboration perspective, textbook chapter perspective, etc.) can be added according to specific application scenarios. Those skilled in the art can flexibly adjust and set them.

[0048] Step S2: Use a time-aware graph convolutional network to perform time encoding on the multi-view heterogeneous graph. By incorporating the time decay factor into the message passing process of graph convolution, output the representation matrix of each view node.

[0049] In this embodiment, the Time-Aware Graph Convolutional Network (TAGCN) is used to incorporate a time decay factor into the message passing process of graph convolution, so as to dynamically adjust the weight of feature aggregation according to the time span of node interaction, thereby simulating the memory forgetting pattern and time effect in the student's cognitive learning process.

[0050] The TAGCN network first extracts the timestamps of graph nodes to calculate the time difference matrix and generate a time-aware adjacency matrix. Then, based on the time-aware adjacency matrix, it performs multi-layer graph convolutional feature propagation on the constructed multi-view heterogeneous graph, achieving deep coupling encoding of graph topology and temporal features. This ensures that the encoded node features retain the objective correlation between questions and knowledge points while accurately mapping the dynamic evolution of students' knowledge mastery decaying over time, highly aligning with the cognitive mechanisms in real teaching scenarios. It outputs three node representation matrices that integrate dynamic temporal characteristics from explicit association perspectives, implicit difficulty perspectives, and macro-evolutionary perspectives, respectively represented as follows: , and .

[0051] In one specific implementation, step S2 includes: Step S21: Construct the node timestamp vector and calculate the time difference matrix. In this step, timestamps are assigned to various nodes in the multi-view heterogeneous graph. Setting timestamps quantifies the absolute time position of node features, providing a numerical benchmark for subsequent calculations of the time span between nodes, i.e., the time difference. Its core function is to drive the graph structure to break the static assumption, enabling the model to dynamically adjust the decay weights of information aggregation based on the temporal distance between nodes when performing graph convolution feature propagation, thereby more accurately simulating the temporal effects and forgetting patterns in cognitive learning.

[0052] For a question node, calculate the average answering time for that question among the student population and use it as the timestamp for that question node; for a knowledge point node, calculate the average answering time for all questions associated with it and use it as the timestamp for that knowledge point node; for a time window super node in the time evolution graph, directly use the center time point of that time window as the timestamp.

[0053] First, from the perspective of the current process Next, The corresponding graph structure contains Aggregate the timestamps of each node to construct a column vector of node timestamps. ,in Indicates the first i The timestamp of each node express A dimensional real-valued column vector space. The current processing perspective. This refers to one of the aforementioned explicit association perspectives, implicit difficulty perspectives, or macro-evolutionary perspectives; Then, using the node timestamp vector Calculate the time difference matrix between nodes : In the formula, For dimension is A column vector of all 1s. elements in Represents a node With nodes The time interval between; This represents absolute value operations.

[0054] Timestamp: A numerical variable used to mark the exact absolute time point of a physical or business event. In this invention, it is assigned as a temporal prior attribute to various nodes in the graph structure to record the time reference of teaching interaction behaviors related to that node.

[0055] Of course, the methods for determining timestamps are not limited to those mentioned above. The timestamp for a question node can also be the time when the question was first answered or the most recent time it was answered; the timestamp for a knowledge point node can be the earliest or latest time of answering its associated question; and the timestamp for a time window node can be the start or end time of that window. Regardless of the method used to determine the timestamp, as long as it reasonably reflects the node's position on the timeline, it can be used to calculate the time difference between nodes.

[0056] Step S22: Map the time difference matrix to a decay factor using a time kernel function, and generate a time-aware adjacency matrix by weighting the original adjacency matrix. From the current processing perspective Let the original graph topology be the original adjacency matrix. ,Right now Represents the construction in S1 above , , One of them. The time difference matrix is ​​processed using a time kernel function. Mapped to a decay factor, and applied to the original adjacency matrix. After weighting, a time-aware adjacency matrix is ​​obtained. : In the formula, This is a preset time sensitivity hyperparameter used to control the decay rate; It is the identity matrix; Represents the Hadamard product; time-aware adjacency matrix This represents the graph topology after the time decay mechanism has been applied, reflecting the objective fact that the connection weights between nodes with large time intervals are significantly weakened.

[0057] The time kernel function is a mathematical mapping mechanism used to quantify the nonlinear decay of human memory over time. In this invention, it is specifically expressed as the exponential decay term in the formula. The design of this function is based on the forgetting theory in educational psychology. When the time span between two nodes (such as a student's answering behavior) is smaller, the decay factor of the function output is closer to 1, which indicates that the knowledge is closely related and the memory is well preserved. However, as the time span increases, the decay factor of the function output decreases rapidly in an exponential manner, which indicates that the association is weakened and the knowledge is forgotten.

[0058] Step S23: Perform symmetric normalized multi-layer graph convolutional propagation using the time-aware adjacency matrix to iteratively update the node feature representation. Using time-aware adjacency matrix Perform feature aggregation and updating between each node and its neighboring nodes: First, construct the time-awareness matrix. , It is a diagonal matrix, and the elements on its main diagonal are... Time-aware adjacency matrix No. The sum of all elements in the row is obtained, i.e. , ,in For matrix The Middle Line number Column elements; All off-diagonal elements are .

[0059] Then, a symmetric normalized graph convolution propagation operation is performed, the first... The update formula for the node representation of the layer is as follows: In the formula, For the first The trainable weight matrix of the layer; This represents the maximum number of convolutional layers. It is a ReLU nonlinear activation function; For the first The node feature matrix of the layer, when hour, The initial characteristic matrix in S1 .

[0060] Step S24: Repeat the above time encoding operation for each composition viewpoint to output the final node representation matrix under multiple views. go through After convolutional propagation of the layered graph, the final node representation matrix is ​​obtained. For the explicit association perspective, implicit difficulty perspective, and macro-evolutionary perspective, repeat the operations from S21 to S23 above respectively to obtain the corresponding explicit association perspective node representation matrix. Implicit difficulty perspective node representation matrix Node representation matrix from a macro-evolutionary perspective Each row vector in the above matrix This represents the structured feature vector of the question from the corresponding perspective, which serves as the input for the subsequent comparison and alignment module.

[0061] For example: setting time sensitivity hyperparameters Initial feature matrix of nodes The dimension is The system calculates the pairwise time difference matrix between 1000 nodes using the node timestamp vectors. When performing graph convolutional propagation, for simplicity of description, the number of convolutional layers is set without loss of generality. The first layer weight matrix The dimension is The second layer weight matrix The dimension is After two layers of convolutional aggregation, the output is a matrix representing the question nodes from three perspectives: explicit association, implicit difficulty, and macro-evolution. , and Their dimensions are all maintained as .

[0062] Step S3: Using a cross-view comparison alignment method, perform intra-view comparison, inter-view comparison, and student-level comparison on the representation matrices of each view node in sequence to eliminate the semantic gap caused by multi-view composition and obtain the aligned multi-view node representation.

[0063] In this embodiment, to eliminate the representation space differences caused by different compositional perspectives, the present invention designs a cross-view comparison alignment mechanism, aiming to enhance the robustness of node feature representations against interference and promote semantic consistency of multi-view features within a unified feature space. Specifically, this mechanism receives three viewpoint node representation matrices output by S2. , and The process involves sequentially performing intra-view comparison, inter-view comparison, and student-level comparison to construct positive and negative sample pairs and calculate the corresponding InfoNCE loss. This operation aims to enhance the robustness of node features under constrained perturbations, force semantic space alignment between multiple views, and capture the inherent consistency of individual macro-knowledge states.

[0064] In one specific implementation, step S3 includes: Step S31: Perform self-comparison of node representations within the viewpoint to enhance the robustness of node features against disturbances under a single viewpoint. For any perspective Node representation matrix after graph convolution The model applies local perturbations to its internal features to construct contrastive learning positive and negative samples. Two enhanced representations of the same question node generated through different perturbations are paired together, denoted as a positive sample pair. Any enhanced representation of the question node is combined with the representations of other question nodes randomly selected from the current batch to form a negative sample pair. Through the construction of these positive and negative samples, the model establishes a constraint mechanism for intra-class consistency and inter-class discriminability in the feature space. Positive samples force the model to learn essential features that are insensitive to structural perturbations, thus maintaining robustness in the face of random noise in student answer data. Negative samples prevent the model from mapping all questions to similar degenerate spaces, thus ensuring high discriminability between different questions.

[0065] First, for each node Two enhanced views are generated by independently performing two random masking or Gaussian noise perturbations. and Forming positive sample pairs; Then, randomly select other nodes within the current training batch. The representation of This constitutes a negative sample; Finally, the InfoNCE loss function is used to calculate the in-view contrast loss. : In the formula, This is the cosine similarity calculation function. For temperature coefficient, This represents the set of nodes in the current batch.

[0066] The contrastive loss values ​​within each of the three perspectives—explicit association, implicit difficulty, and macro-evolution—were calculated separately. , and .

[0067] Finally, the contrast loss values ​​from the three perspectives are summed to obtain the total loss within the perspective. : Total loss from the above perspective This will serve as the basic loss term in the subsequent joint optimization stage to drive the gradient update of the model parameters, ensuring that the node representations of each perspective are robust against structural noise.

[0068] The constraint mechanism of intra-class consistency and inter-class discriminability refers to an optimization rule that guides the distribution of node feature vectors in a high-dimensional representation space through a contrastive loss function during the model training phase. Intra-class consistency mandates that after experiencing different structural perturbations (i.e., positive sample pairs), the multiple view vectors generated by the same question node remain highly clustered and close to each other in the feature space, prompting the model to ignore random noise and extract the essential attributes of the question. Inter-class discriminability, on the other hand, mandates that the feature vectors of different question nodes (i.e., negative sample pairs) are mutually exclusive and spaced out in the space, thus preventing feature smoothing during repeated graph convolution aggregation and avoiding blurring of semantic boundaries between different questions. This mechanism synergistically ensures that the question features output by the model possess both robust stability against interference and high independent discriminability.

[0069] The weighting coefficients for each loss are adjustable hyperparameters, which can be flexibly configured by those skilled in the art according to the importance of each comparison task. It should be noted that the weighting coefficients are not limited to fixed values, and dynamic adjustment strategies can also be adopted, such as gradually increasing or decreasing the weight of a certain loss according to the training process, as long as the optimization objectives of each comparison task can be effectively balanced.

[0070] Step S32: Perform a comparison of node representations across different viewpoints to achieve semantic space alignment of the same node across multiple viewpoints. For the same question node This maximizes the cosine similarity between the node's representation vectors from different perspectives, making the node's vector representations in the high-dimensional feature space closer to each other. Even though it's the same question node... The graph structures, representing explicit associations, implicit difficulty, and macroscopic evolution, capture features with different emphases, but the essence of the question as an objective physical entity remains unique. Forcing these features to align closely in the feature space encourages cross-domain interaction and information complementarity between features from different perspectives, preventing the model from falling into the bias of a single perspective. Through this consistency constraint, the model can ultimately extract a globally comprehensive representation for each question that integrates multi-dimensional contextual attributes and possesses inherently unified semantics.

[0071] First, for any problem Node representation matrix from three perspectives , and Extract the corresponding representation vector , and Using cosine similarity to maximize the distance between pairs of nodes, taking explicit association perspectives and implicit question perspectives as examples, this approach is applied to the node set within the current batch. Its comparative loss Defined as: Similarly, the contrast loss between the relational perspective and the macro-evolutionary perspective can be calculated. A comparison of the losses from the implicit problem perspective and the macro-evolutionary perspective : Then, the three contrast losses mentioned above are summed to obtain the total interview loss. : Finally, the total loss between viewpoints is calculated using the backpropagation algorithm. Minimize the loss function. Driven by gradient descent, compare the loss functions from the above perspectives. , , The molecular parts, namely respectively , , This will be continuously amplified, while its denominator (representing the cross-perspective similarity between different problem nodes) will be continuously reduced. Thus, at the mathematical and parameter update level, this forces the same problem... The multi-view representation vectors are brought closer together in the feature space, achieving semantic alignment.

[0072] Step S33: Perform aggregation and comparison of student-level historical sequences to achieve internal consistency alignment of individual macro-knowledge states. In this step, a student-level comprehensive representation is constructed based on the student's historical interaction sequence to capture the inherent consistency of the individual's macro-knowledge state.

[0073] (1) Construct a comprehensive representation of the entire sequence as a negative sample pool: for each student and its historical interactive question sequence The three viewpoint nodes output from S2 represent matrices respectively. , and In the process, the feature row vectors corresponding to each question node in the interaction sequence are extracted and average pooled to obtain the aggregated representation from each perspective. , and Then, they are pieced together to form a comprehensive student representation. ; in, The plus sign indicates average pooling operation, and the plus sign indicates concatenation operation. (2) Construct subsequence augmented view and construct positive sample pairs: for the current student's historical interaction sequence Perform two independent random subsequence samplings to obtain two subsequences. and For these two subsequences, repeat the feature extraction, pooling, and concatenation operations in (1) above to calculate the first enhanced view. Second Enhanced View .

[0074] (3) Calculate student-level contrast loss: The current student Two enhanced views As positive sample pairs; at the same time, extract other students in the current training batch (denoted as...). ,and The complete synthetic representation calculated in step (1) above. This is used as a negative sample for the contrast task. The student-level contrast loss is calculated. : In the formula, The total number of students in the batch. Refers to students who are not currently enrolled in the batch. A comprehensive representation of other individuals.

[0075] In this context, "batch" refers to a basic data partitioning unit during the training process of a deep learning model. Since the massive amount of real student answer records cannot be loaded into the computer's video memory / RAM all at once, the system divides the entire dataset into several fixed-size subsets (i.e., "batches" or "Mini-batches," such as the batch size set in the example) during model training. The model reads only the data in the current batch (such as the answer sequence of 32 students) each time to perform comparative loss calculation and gradient backpropagation. After traversing all batches, one epoch of global training is completed.

[0076] Step S34: Integrate multiple losses to construct a global joint optimization objective, collaboratively drive the update of various parameters, and output an explicit association viewpoint node representation matrix. Implicit problem perspective node representation matrix Node representation matrix from a macro-evolutionary perspective ; The previously independently calculated contrastive losses are integrated into a global joint optimization objective. : In the formula, This is a weight hyperparameter used to dynamically adjust the importance of each comparison constraint in the global optimization process.

[0077] It is the final backpropagation optimization metric of the entire cross-view contrast alignment mechanism, which balances the gradient contributions of each level of the contrast task through weighted summation. Utilizing... Guided gradient descent can collaboratively drive the synchronous update of various parameters, which refer to the weight parameters of the Time-Aware Graph Convolutional Network (TAGCN) and the initial embedding matrix of the underlying nodes.

[0078] During the training phase of the TAGCN network, continuous backpropagation is used. The system minimizes this value until the model converges. The final output is a jointly optimized, fully aligned, and highly robust explicit relational viewpoint node representation matrix. Implicit problem perspective node representation matrix Node representation matrix from a macro-evolutionary perspective This provides high-quality global feature support for downstream knowledge tracing and probability prediction tasks.

[0079] The initial embedding matrix of the bottom-level nodes refers to the basic parameter matrix used to map various discrete nodes (such as question nodes, knowledge point nodes, and time window super nodes) into continuous, low-dimensional dense real vectors before performing any graph feature propagation and aggregation operations. In this invention, it is the initial feature matrix constructed in step S1. .

[0080] The significance of "bottom layer" is that it emphasizes that the matrix is ​​located at the very beginning of the entire deep learning network architecture, and is the starting point for all subsequent multi-view graph convolutions and feature transformations.

[0081] The initial purpose is to obtain the optimized node representation (such as the initial embedding matrix of the underlying node and the final output after multi-layer network propagation and comparison alignment mechanisms) from the initial embedding matrix of the underlying node. , and Make clear distinctions about the life cycle.

[0082] During the joint training phase, the initial embedding matrix of the bottom node not only provides prior input, but also serves as the core learnable parameter of the model. Driven by the backpropagation mechanism, it will be dynamically updated and optimized synchronously according to the loss gradient.

[0083] For example: setting the training batch size Comparison of temperature coefficient of loss In in-view comparison, the node feature vectors are processed. A proportional random dimension masking operation (Dropout) is used to construct positive sample pairs. In student-level comparisons, if a student's historical interaction sequence is 100 units long, two non-overlapping subsequences of length 50 are randomly sampled and aggregated to construct positive samples of the student's features. Simultaneously, the combined representations of other students in the current training batch are directly extracted as negative samples for this comparison task. When constructing the global joint optimization objective, the loss weight coefficients are set as follows: (Within the field of view) (Between perspectives) and (Student level), and then all trainable parameters in TAGCN can be updated synchronously through the backpropagation algorithm.

[0084] Step S4: Dynamically fuse the aligned multi-view node representations using an attention mechanism to obtain a multi-view fused representation of the question. Input the multi-view fused representations corresponding to each question in the student's historical answer sequence into a gated recurrent unit for sequence state encoding to obtain the student's current knowledge state. Combined with the embedding of the question to be predicted, predict the probability of the student answering the next question correctly.

[0085] In this embodiment, in order to break the independence between different compositional perspectives, the present invention introduces an adaptive attention mechanism to fuse the representation of multi-view nodes, thereby obtaining comprehensive features of the question containing all contextual attributes, and then performing an inner product with the student's current implicit cognitive state to finally complete the core task of knowledge tracking that predicts the student's future answer performance (i.e., outputs the probability of answering correctly).

[0086] In one specific implementation, step S4 includes: Step S41: Perform adaptive attention fusion of multi-view features to generate a multi-view fusion representation of the question. First, regarding any question From the node representation matrix after S3 joint optimization , and Get the corresponding first from each row vector , and Let be the vector representing the problem from the corresponding perspective; for ease of description, let be the perspective. , express , or one; Then, a globally learnable query vector is randomly initialized. ,use Corresponding projection matrix Map the representation vectors to the query space and compute the view attention weights. : Finally, the weighted sum of the representation vectors from each perspective is taken to obtain the solution to the problem. Multi-view fusion representation : Among them, perspective Corresponding projection matrix This refers to the learnable parameter matrix that is independently initialized for each compositional viewpoint during the model building phase. This matrix is ​​not derived from viewpoint variables, but rather represents the trainable weights at the model's underlying level. Because different viewpoints (such as explicit associations and implicit difficulty) extract features with different focuses, their latent variable feature spaces exhibit semantic differences. The projection matrix is ​​introduced to address this. Its core function is to perform a linear transformation (i.e., spatial projection) on the representation vector of the current viewpoint, unifying heterogeneous features from different viewpoints into a common global query space. This ensures that the model can compute the features of each viewpoint and the global query vector under a unified dimensional benchmark. The degree of matching is used to calculate attention weights.

[0087] Step S42: Input the multi-view fusion representation sequence into the gated recurrent unit for encoding to obtain the student's current knowledge state vector. First, extract the students' information in chronological order. Based on the previous historical interactive question sequence, the corresponding multi-view fusion representation is obtained using S41. Construct a fusion representation sequence ; Then, the fused representation sequence is gradually input into the gated recurrent unit network (GRU) to set the initial hidden state. Given a zero vector, compute the hidden state at the current time step. As students, at all times Knowledge mastery level state vector: Step S43: Concatenate the student's knowledge state vector with the embedding vector of the question to be predicted, and output the predicted probability of the correct answer through linear mapping. In order to predict the student at time Answer the questions The performance, firstly, starts from the initial embedding of the question. Extracting questions The corresponding trainable embedding vector Students at all times Knowledge state vector embedding vector with question Concatenate the features to construct a feature combination vector. The feature combination vector is then input into a dataset containing trainable weights. and bias A linear mapping layer is applied, and after processing with the Sigmoid activation function, the output shows the student's correct answer. Predicted probability : In the formula, This represents a vector concatenation operation. Represents the Sigmoid function; Finally, based on the predicted probability The system judges the correctness of answers against preset thresholds and completes knowledge tracking prediction. The knowledge tracking prediction refers to dynamically assessing the student's knowledge mastery level at the current moment based on the student's historical answer interaction sequence and time evolution pattern captured by the model, and then predicting their answer performance when facing unknown specific questions in the future. This provides an objective prior basis for downstream personalized test question recommendation, learning situation diagnosis and analysis, and adaptive learning path planning.

[0088] Example: Setting a globally learnable query vector The dimension is 128, and the projection matrix is ​​obtained. The dimension is By calculating attention weights, the 128-dimensional feature vectors from the three perspectives are weighted and fused into a 128-dimensional fused representation. The hidden state dimension of the gated loop unit (GRU) is set to... .

[0089] When performing knowledge tracking and prediction, extract the questions to be predicted. The 128-dimensional embedding vector, and the current 256-dimensional knowledge state vector. The vectors are concatenated to generate a 384-dimensional combined feature vector. Then, the vectors are processed using a dimension of... The output weight matrix Perform linear mapping and Sigmoid processing to output the probability of correct prediction. Without loss of generality, suppose that when processing a student's answer prediction, the above-mentioned concatenated 384-dimensional combined feature vector and weight matrix... After performing the inner product operation, the resulting scalar value is 1.85. This 1.85 is then input into the Sigmoid activation function for a non-linear transformation, i.e. Therefore, the system outputs a prediction probability of 0.864, meaning the model predicts the student will answer the next question correctly. The probability of answering correctly is 86.4%.

[0090] Among them, the gated loop unit automatically captures the evolution of students' knowledge status over time through its internal gated mechanism.

[0091] Of course, sequence coding networks are not limited to gated recurrent units. Other neural network structures with sequence modeling capabilities, such as Long Short-Term Memory (LSTM) networks and Transformer encoders, can also be used. Those skilled in the art can make flexible choices.

[0092] Step S5: Perform joint training of the knowledge tracing model MGCLKT and deploy it in a real-world educational application. The knowledge tracing model described in S1 to S4 is based on the construction of multi-view heterogeneous graphs, graph structure temporal encoding, cross-view comparison alignment, and multi-view fusion prediction. It is referred to as the MGCLKT model.

[0093] Step S51: Perform backpropagation based on the global joint optimization objective, and synchronously update the weight parameters and initial embedding matrices of the time-aware graph convolutional network. In this step, based on the global joint optimization objective The gradient is calculated through backpropagation, and the weight parameters of the time-aware graph convolutional network TAGCN and the initial embedding matrix of all nodes are updated synchronously. After multiple rounds of iterative optimization, until Upon convergence, the system ultimately outputs an explicit association viewpoint node representation matrix that has undergone semantic alignment and robustness optimization. Implicit difficulty perspective node representation matrix A node representation matrix from a macro-evolutionary perspective.

[0094] Note: The node representation matrix output above provides a plug-and-play, high-quality underlying question bank feature library with strong noise resistance for downstream practical educational applications such as "student knowledge mastery diagnosis", "personalized test question recommendation" and "adaptive learning path planning". It can alleviate the problem of knowledge tracking and prediction failure caused by sparse student answer records or random errors in real teaching systems.

[0095] The backpropagation mechanism refers to the core algorithmic process by which deep learning models optimize parameters during the joint training phase. Its basic principle is to utilize the chain rule from calculus to calculate the global joint optimization objective (i.e., the loss function). The gradients of each trainable parameter in the network are calculated, and the error signal is propagated layer by layer in the opposite direction of gradient descent. In the joint training step of this invention, the backpropagation mechanism, guided by the collaborative reduction of multi-contrast loss, accurately propagates the error signal to the initial embedding matrices of all nodes at the input end and the weight parameters of the Temporally Aware Graph Convolutional Network (TAGCN) in the feature extraction layer. Through multiple iterations, these two core parameter groups are updated and optimized synchronously, continuously minimizing the global loss, and ultimately ensuring that the multi-view node representation matrix output by the model has high robustness and semantic consistency from the underlying data flow.

[0096] Step S52: Deploy the trained MGCLKT model in a real-world application scenario, and use sequence state coding to assess students' knowledge mastery and predict their answer performance. In this step, the multi-view graph contrastive learning MGCLKT model, which has been optimized through the above joint training steps, is deployed in a practical knowledge tracking application.

[0097] For the target students, extract their historical interactive question sequences in chronological order, and utilize... , and Feature representations of the corresponding questions are extracted, and a multi-view fusion representation sequence is adaptively generated through a global query vector and an attention mechanism. This fusion representation sequence is then input into a lightweight gated recurrent unit (GRU), and the student's current knowledge mastery level hidden state is obtained through sequence state encoding. Subsequently, the embedding vector of the next target question is extracted, and it is concatenated with the hidden state of the current knowledge mastery level. The concatenation is then input into a linear mapping layer and processed by a Sigmoid activation function. Finally, the predicted probability of the student answering the next target question correctly is output, thus completing knowledge tracking and state evaluation.

[0098] Example 1: The collected student answer dataset was divided into training, validation, and test sets in a 7:1:2 ratio. The model was trained using the Adam optimizer with an initial learning rate of 0.001 and a training period of 200 epochs. Early stopping was implemented to prevent overfitting when the loss on the validation set stopped decreasing after 10 consecutive epochs. In the application test, given 50 historical answer records of a student, the model could output in real-time that the student's prediction accuracy for the next question with a difficulty coefficient of 0.8 was 0.72, and simultaneously update the student's latent vector of knowledge mastery level.

[0099] Suppose a student, user A, is studying the topic of "High School Mathematics - Trigonometric Functions". The system records their most recent 10 responses in real time.

[0100] (1) Knowledge State Diagnosis: After student A completes question 10, the system-trained model is used. By inputting student A's answer sequence into the model, the model's hidden state is determined. The student's current cognitive level was automatically aggregated. At this point, The vector has a higher value in the "sine function graph" dimension and a lower value in the "application of induction formulas" dimension. This objectively reflects that student A has a strong image recognition ability, but has a weakness in formula transformation.

[0101] (2) Prediction and Recommendation Logic: The teaching platform currently offers two alternative recommended questions: (Comprehensive proof questions with higher difficulty) and (This is a basic reinforcement question targeting the reduction formulas). The system first extracts... The fused feature vector, and Calculate the predicted probability of answering correctly. Due to the extremely low probability, it was determined that pushing this question at this time would discourage students, therefore it was not recommended. Subsequently, the system extracted... The fused feature vectors are used to calculate the predicted correct answer probability. .

[0102] (3) Final application effect: system judgment This falls within the range that student A can master with a little effort (0.7-0.8 is the golden practice range), therefore... The information is pushed to student A's learning interface. In this way, the present invention achieves precise learning guidance based on deep graph representation learning. It not only quantifies the student's macro-level knowledge status but also enables personalized question recommendations through accurate probability prediction, significantly improving learning efficiency.

[0103] Example 2: Based on the method proposed in this invention, a knowledge tracing model, abbreviated as MGCLKT, is implemented. The structure diagram of the MGCLKT model is shown below. Figure 2 As shown.

[0104] The MGCLKT model was compared with some publicly available baseline models on the same dataset in terms of prediction performance, with the AUC value used as the comparison metric. The ASSISTments2012 and ASSISTments2017 datasets were obtained from the ASSISTments online tutoring platform and are typically used as homework or classroom exercises for students. The Algebra2005 dataset originated from the algebra course of the Cognitive Tutor system and records real-world interactions between junior and senior high school students. The results are shown in the table below.

[0105] Table 1 Performance comparison between the proposed model and the baseline model.

[0106] The MGCLKT model of this invention exhibits stable predictive performance on three datasets. On the ASSISTments2017 dataset, the MGCLKT model achieves an AUC of 0.7621, which is 0.44% higher than the second-best model, sparseKT. On the Algebra2005 dataset, it achieves an AUC of 0.9274, outperforming sparseKT by 1.77%. On the ASSISTments2012 dataset, which has a relatively large dataset, the MGCLKT AUC reaches 0.7829, comparable to the best baseline.

[0107] This performance demonstrates the comprehensive characterization of complex relationships between questions using a multi-perspective heterogeneous graph structure. The explicit association perspective provides prior knowledge structure, the implicit difficulty perspective mines difficulty clustering information through group answering behavior, and the macro-evolutionary perspective integrates the time dimension into the graph representation. This complementary design of the three perspectives enables the model to understand question features from multiple dimensions, compensate for data sparsity with group information, and maintain stable predictive ability when data is abundant. The introduction of a time-aware graph convolutional network further enhances the model's sensitivity to temporal dynamics during the learning process, allowing the multi-perspective representation to reflect the evolution of knowledge over time.

[0108] The MGCLKT model uncovers implicit associations in the difficulty dimension through question co-answer graphs and uses time evolution graphs to depict the macroscopic drift of group knowledge states, transforming individual behavioral characteristics into group structural characteristics, thereby achieving effective knowledge transfer under limited data conditions. For the Algebra2005 dataset, the knowledge point structure is clear, and the inclusion relationship between questions and knowledge points is explicit. The introduction of an explicit association perspective allows the model to fully utilize the deterministic topology of the knowledge space, and, combined with a cross-perspective contrast alignment mechanism, eliminates semantic gaps, ultimately improving prediction accuracy.

[0109] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A knowledge tracking method based on multi-view graph contrast learning, characterized in that, Includes the following steps: A multi-perspective heterogeneous graph is constructed, including: a question-knowledge point bipartite graph, a question co-answer graph, and a time evolution graph; the question-knowledge point bipartite graph is constructed based on the explicit association between questions and knowledge points, and is used to represent the explicit association perspective; the question co-answer graph is constructed based on the objective difficulty similarity relationship of questions mined from group answer data, and is used to represent the implicit difficulty perspective; the time evolution graph is constructed based on the interaction relationship between questions and time windows, and is used to represent the macro-evolutionary perspective; The multi-view heterogeneous graph is temporally encoded using a time-aware graph convolutional network. By incorporating a time decay factor into the message passing process of graph convolution, the representation matrix of each view node is output. By using a cross-view comparison alignment method, the node representation matrices of each view are sequentially compared within the view, between the view, and at the student level to eliminate the semantic gap caused by multi-view composition and obtain the aligned multi-view node representations. An attention mechanism is used to dynamically fuse the aligned multi-view node representations to obtain a multi-view fused representation of the problem. The multi-perspective fusion representation of each question in the student's historical answer sequence is input into the gated recurrent unit for sequence state encoding to obtain the student's current knowledge state. Combined with the embedding of the question to be predicted, the probability of the student answering the next question correctly is predicted.

2. The method according to claim 1, characterized in that, The construction of the question co-answer graph includes: For each question, the percentage of students who answered the question correctly was calculated as the objective difficulty of the question for that group. Calculate the difficulty difference between any two questions in the question bank, select a preset number of other questions with the smallest difficulty difference as difficulty neighbors for each question, and establish undirected connecting edges between questions that are difficulty neighbors. The question co-answer graph is composed of all question nodes and the undirected connecting edges.

3. The method according to claim 1, characterized in that, The time evolution graph is represented by a four-block adjacency matrix; in the four-block adjacency matrix: The top-left corner is a zero matrix, indicating that there are no direct connections between the problem nodes from a macroscopic evolution perspective; The top right corner block records the normalized interaction frequency of each question within each time window; The bottom left corner piece is the transpose of the top right corner piece; The bottom right block is a strip matrix with a value of 1 assigned only at the corresponding positions of adjacent time windows, used to represent the sequential evolution relationship of time windows.

4. The method according to claim 1, characterized in that, In the time-aware graph convolutional network, the time decay factor is incorporated into the message passing process of graph convolution in the following way: Calculate the time difference matrix between nodes, and map the time difference matrix into a decay factor using a preset time decay function; The attenuation factor is multiplied element-wise with the original adjacency matrix of the current viewpoint to obtain the time-aware adjacency matrix. Multi-layer graph convolution propagation is performed based on the time-aware adjacency matrix, so that the aggregation weight of node features is adaptively decayed according to the time distance between nodes.

5. The method according to claim 4, characterized in that, Time-aware adjacency matrix : In the formula, This is the original adjacency matrix. This is a preset time sensitivity hyperparameter used to control the decay rate; Time difference matrix It is the identity matrix; Represents the Hadamard product; Multi-layer graph convolutional propagation is performed according to the following formula: In the formula, This is the time perception matrix. For the first The trainable weight matrix of the layer; This represents the maximum number of convolutional layers. It is a ReLU nonlinear activation function. For the first The node feature matrix of the layer, when hour, The initial characteristic matrix .

6. The method according to claim 1, characterized in that, The comparison within the specified viewpoint includes: For each question node, perform two independent random perturbations to generate two enhanced views that form a positive sample pair; Randomly select the representations of other question nodes within the current training batch to form negative sample pairs; The in-view contrast loss is calculated based on the positive and negative sample pairs, and then summed after calculating the losses for the three perspectives of explicit association, implicit difficulty, and macro-evolution to obtain the total in-view loss.

7. The method according to claim 6, characterized in that, The comparison between the viewpoints includes: Pair the representation vectors of the same question from the perspectives of explicit association, implicit difficulty, and macro-evolution to form positive sample pairs; The representation of the current question from one perspective is compared with the representations of other questions in the batch from another perspective to form a negative sample pair; The comparative loss is calculated and summed for the explicit correlation perspective and the implicit difficulty perspective, the explicit correlation perspective and the macro-evolutionary perspective, and the implicit difficulty perspective and the macro-evolutionary perspective, respectively, to obtain the total loss between perspectives.

8. The method according to claim 7, characterized in that, The student-level comparison includes: Two independent random subsequence sampling operations are performed on the current student's historical answer sequence. The feature vectors of the corresponding questions are extracted from the node representation matrix of each viewpoint and then pooled and concatenated to obtain two subsequence augmented views as positive sample pairs. The complete sequences of other students in the current training batch are combined as negative samples. The student-level contrastive loss is calculated based on the positive and negative sample pairs.

9. The method according to claim 8, characterized in that, The total loss within the viewpoint, the total loss between viewpoints, and the student-level contrast loss are weighted and summed to construct the joint contrast loss. By minimizing the joint contrastive loss through backpropagation, the weight parameters of the time-aware graph convolutional network and the initial embedding matrix of all nodes are updated synchronously to obtain the aligned multi-view node representation.

10. The method according to claim 1, characterized in that, The timestamp of the node is determined in the following manner: The timestamp of the question node represents the average time it took for the student group to answer the question. The timestamp of a knowledge point node represents the average answering time for all questions associated with it. The timestamp of the supernode of the time window is the center time point of that time window.

Images