A bidirectional dynamic graph knowledge tracing method with information bottleneck enhancement

Abstract

The present invention relates to the field of graph neural networks and knowledge tracking, and provides a bidirectional dynamic graph knowledge tracking method enhanced by information bottlenecks, including: (1) bidirectional dynamic graph construction based on student-question interaction; (2) bidirectional dynamic graph neural network construction; (3) graph multi-view learning enhanced by information bottlenecks; (4) multi-view information selection and fusion; (5) student-question prediction. The method of the present invention comprehensively utilizes graph neural networks and information bottlenecks to conduct in-depth mining of students' answer data, can scientifically predict students' learning performance, and accurately assist students in personalized learning.

Description

Technical Field

[0001] The present invention relates to the field of graph neural networks and knowledge tracking, and in particular to a bidirectional dynamic graph knowledge tracking method with enhanced information bottleneck. Background Art

[0002] In recent years, online education has attracted more and more attention because it breaks the time and space limitations of teaching and has flexible and diverse characteristics. However, the urgent need for intelligent tutoring systems has also put forward a new requirement: to track students' knowledge status online based on their learning records and help them achieve personalized learning. Knowledge tracking aims to track students' mastery of knowledge concepts by analyzing their response records in the intelligent tutoring system, thereby predicting their future performance.

[0003] Traditional Bayesian knowledge tracing (BKT) tracks students' knowledge status by inputting their learning records into a hidden Markov model. The model assumes that students' knowledge status is a set of binary variables, namely mastered and not mastered. With the development of deep learning technology, more researchers have applied deep learning technology to the field of knowledge tracing, using the deep learning knowledge tracing model (DLKT) to model the dynamic changes of learners' knowledge status over time. DKT is the first model to apply recurrent neural networks to knowledge tracing tasks, SAKT introduces attention mechanisms into knowledge tracing, SAINT uses transformers to analyze student performance, and GKT uses graph neural networks to transform knowledge concepts and relationships between concepts into graphs.

[0004] In the knowledge tracking scenario, the factors that affect students' performance on test questions can be attributed to two aspects: test questions and students. On the one hand, the higher the students' knowledge mastery, the higher the probability of answering the test questions correctly. In addition, the students' learning status will also affect their performance, that is, mistakes and guesses. On the other hand, if the test questions are difficult, students may not be able to answer correctly even if they have reached the required mastery level; if the test questions are not difficult, it is impossible to distinguish students of different levels. In the existing knowledge tracking research, some methods improve the accuracy of the model by enhancing the representation of test questions, such as EKT adding test question text information, DIMKT considering the difficulty of test questions, and GKT combining test questions and concepts into heterogeneous graphs; others improve the model performance by extracting more student characteristics, such as DKT-DSC and ABKT considering student ability, CKT considering students' prior knowledge, and CoKT adding similar student retrieval. It is worth noting that the model's accurate grasp of both test questions and students is the key to improving performance. Students' mastery of knowledge needs to be reflected through answering questions, and attributes such as the difficulty of test questions also need to be comprehensively determined by the performance of a large number of students on test questions. How to build a model based on students' answer data to track knowledge status and test questions in a two-way manner, and then achieve in-depth analysis of the interaction between test questions and knowledge status to improve prediction performance is a thorny issue. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and to provide a bidirectional dynamic graph knowledge tracking method enhanced by information bottlenecks. It comprehensively utilizes technical methods such as graph neural networks and information bottlenecks to conduct in-depth mining of students' answer data, which can scientifically predict students' learning performance and accurately assist students in personalized learning.

[0006] The purpose of the present invention is achieved through the following technical measures.

[0007] A bidirectional dynamic graph knowledge tracking method with information bottleneck enhancement includes the following steps:

[0008] (1) Construction of a bidirectional dynamic graph based on student-question interaction: Generate a dynamic heterogeneous graph based on the interaction data of students in the intelligent tutoring system, where nodes include students and questions, and edge features include responses and timestamps. Then, generate the original subgraph through subgraph selection in the heterogeneous graph.

[0009] (2) Construction of a bidirectional dynamic graph neural network; mining test information through message propagation from students to test questions and node updates, and tracking student status through message propagation from test questions to students and node updates;

[0010] (3) Information bottleneck enhanced graph multi-perspective learning: First, the node and edge deletion matrices are calculated based on the node representation and edge representation obtained after the original subgraph passes through the bidirectional dynamic graph neural network. The retained perspective after node deletion and the retained perspective after edge deletion are created through the deletion matrix. Then, multi-perspective learning is performed based on the information bottleneck theory to obtain the student node representation and test question node representation of each perspective.

[0011] (4) Multi-perspective information selection and fusion: Use a gating network to select and fuse the student node representation and question node representation obtained from the retained perspective after edge deletion and the retained perspective after node deletion;

[0012] (5) Student-question prediction: predicting the probability of a student correctly answering the next question and the difficulty of the question currently being answered by the student.

[0013] In the above technical solution, the construction of the two-way dynamic graph based on student-question interaction in step (1) is specifically as follows:

[0014] (1-1) Global graph construction: construct a dynamic heterogeneous graph based on students’ interaction data in the intelligent tutoring system Where V represents the test question and student nodes, and E represents the edge connecting the test question node and the student node. Represents edge features, including answer responses and timestamps;

[0015] (1-2) Subgraph selection: For each student s, a subgraph is created at each answering moment. At each time point t, first, with student s as the center, select the l closest to s in time from the global graph. s test questions as first-order neighbors, and then take each test question as the center and select l q The students connected to the test question q are taken as the second-order neighbors, and so on, the multi-hop neighbors of the student node s are obtained, forming About the k-order subgraph of student s

[0016] In the above technical solution, the bidirectional dynamic graph neural network construction in step (2) is specifically as follows:

[0017] (2-1) Mining test information based on the one-way graph from students to test questions: Use the attention mechanism to propagate messages from students to test questions, and then aggregate the test question embeddings obtained in the message propagation stage and the test question embeddings of the previous layer to achieve information mining. Update of test question nodes;

[0018] (2-2) Student status tracking based on the test-to-student unidirectional graph: Use the GRU network to propagate messages from the test questions to the students, and then implement tracking by aggregating the student embedding obtained in the message propagation stage and the student embedding of the previous layer. Update of the middle school student node.

[0019] In the above technical solution, the information bottleneck enhanced graph multi-view learning in step (3) is specifically as follows:

[0020] (3-1) Calculation of node and edge deletion matrices: The node representation and edge representation obtained after the original subgraph passes through the bidirectional dynamic graph neural network are respectively input into the multi-layer perceptron (MLP), Sigmoid activation function and Gumbel-Softmax to obtain the node deletion matrix and edge deletion matrix. The deletion matrix is ​​used to determine whether the nodes and edges should be retained;

[0021] (3-2) Creating a retained perspective after node deletion and a retained perspective after edge deletion: the original subgraph deletes certain nodes according to the node deletion matrix to obtain a retained perspective after node deletion; the original subgraph deletes certain edges according to the edge deletion matrix to obtain a retained perspective after edge deletion;

[0022] (3-3) Multi-perspective learning enhanced by information bottleneck: The retained perspective after node deletion and the retained perspective after edge deletion are input into the bidirectional dynamic graph neural network to obtain the student node representation and the test question node representation; the optimization goal is designed based on the information bottleneck theory:

[0023]

[0024] in, represents the original subgraph, Indicates the retained perspective after the node is deleted, represents the retained view after edge deletion, Y represents the class label of the predicted target, and β 1 and β 2 ∈[0,1] are two trade-off factors, I(·) represents mutual information, and max represents maximization.

[0025] In the above technical solution, the multi-view information selection and fusion in step (4) is specifically as follows:

[0026] By constructing a gated network, the student node representation and question node representation from the two perspectives are selected and fused respectively. First, the student node representation and question node representation obtained after the original subgraph passes through the bidirectional dynamic graph neural network are used as selectors, and the weights of the two perspectives are obtained through the MLP network and Softmax. Then the weights are multiplied with the corresponding perspectives, and the two perspective information after multiplication are added together to obtain the fused feature vector.

[0027] In the above technical solution, the student-question prediction in step (5) is specifically as follows:

[0028] (5-1) Knowledge tracking: Apply the multilayer perceptron and sigmoid activation function to the fused student node representation to obtain the weights of the student node representation and the question node representation for the knowledge tracking task. Then, multiply the weights by the corresponding node representation to obtain the student hidden state. Finally, the hidden state is activated by sigmoid to predict the answer and predict the probability that the student will correctly answer the next question.

[0029] (5-2) Difficulty tracking: The difficulty of the test questions in the knowledge tracking dataset is defined as: Where q represents the test question, count q Indicates the number of people who answered the question, correct q It represents the number of people who answered the question correctly. The multilayer perceptron and Sigmoid activation function are applied to the fused question node representation to obtain the question node and student node weights for the difficulty tracking task. The weights are then multiplied by the corresponding node representation to obtain the question hidden state. Finally, the hidden state is activated by Softmax to predict the difficulty of the question currently being answered by the student.

[0030] The present invention proposes a bidirectional dynamic graph knowledge tracking method enhanced by information bottleneck, which uses dynamic graph neural network technology in the field of deep learning to characterize test questions and student characteristics; further, it combines multi-view learning with information bottleneck to ensure that the model reduces noisy interactions while extracting information from multiple perspectives; finally, it predicts the difficulty of test questions and student answer performance based on student and test question characteristics to achieve the recognition of learner knowledge status. The present invention can track student knowledge status and predict student answer performance based on obtaining test question characteristics, so as to achieve the purpose of assisting students in personalized learning and teachers in precise teaching. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A framework diagram of a method according to an embodiment of the present invention.

[0032] Figure 2 An example diagram for constructing a student-question interactive bidirectional dynamic diagram according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention discloses a bidirectional dynamic graph knowledge tracking method enhanced by information bottleneck. Specifically, a dynamic heterogeneous graph is first generated based on the interaction data of students in an intelligent tutoring system. Then, the graph neural network technology is used to model the students' knowledge status and mine the characteristics of test questions. In the modeling process, the multi-perspective information of students and test questions is mined from the retained perspective after the edge deletion and the retained perspective after the node deletion of the heterogeneous graph, and the information bottleneck theory is used to help the model remove noise and redundant information. Afterwards, a gating network is used to select and fuse multi-perspective information, and the interaction between students and test questions is analyzed based on the tracked student knowledge status and test question characteristics. Finally, the learner's future performance and the difficulty of the test questions are predicted.

[0034] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0035] like Figure 1 As shown, an embodiment of the present invention provides a bidirectional dynamic graph knowledge tracking method with information bottleneck enhancement, comprising the following steps:

[0036] (1) Construction of a two-way dynamic graph based on student-question interaction.

[0037] (1-1) Global Graph Construction

[0038] The first step of the model is to generate a dynamic heterogeneous graph based on the students’ interaction data in the intelligent tutoring system. Figure 2 As shown in the global graph construction steps, when learner s answers question q at time t, an edge e is established between s and q, and the time t when the answer occurs and the answer result r are the edge features. In the intelligent tutoring system, different students are connected by answering the same questions, forming a dynamic graph:

[0039]

[0040] The node set V = S∪Q, S is the set of all learners, and Q is the set of all test questions. are edge features, including responses and timestamps. In particular, in order to mine the characteristics of test questions and learners respectively, there are two types of edges in the dynamic graph, one from students to test questions and the other from test questions to students.

[0041] For each time point t, divide the edge set E into several subsets E t , represents all the edges established between learners and test questions before time t. According to the definition of dynamic graph, dynamic graph It can be viewed as a continuous time graph, with each time point corresponding to a subgraph:

[0042]

[0043] (1-2) Sub-graph selection

[0044] In real learning scenarios, as time goes by, the interaction sequences between learners and test questions continue to increase, and new learners continue to join. Therefore, the scale of the dynamic graph based on the interaction between learners and test questions is also gradually expanding. On the one hand, the huge data scale will lead to an increase in computing costs, and on the other hand, a large number of neighbor nodes will introduce too much noise and cause information redundancy. In order to help the model more effectively mine test questions and learner information, it is necessary to sample the global graph.

[0045] The task of knowledge tracking is to predict the probability that the learner will answer the test correctly at the next moment, so a subgraph is established for each learner s at each answering moment. Figure 2 As shown in the subgraph selection step, at each time point t, first, learner s is taken as the center, and Select l that is closest to s in time. s The test questions are taken as first-order neighbors, denoted as Then, for each test question As the center, select l q The learners connected to question q are regarded as second-order neighbors, denoted as By analogy, we can get the multi-hop neighbors of the learner node s, forming About the k-order subgraph of learner s After sampling, each subgraph The nodes in are connected to each other through learner-to-question and question-to-learner relationships.

[0046] (2) Construction of bidirectional dynamic graph neural network

[0047] In order to model the dynamic evolution of students' knowledge status and explore the inherent static characteristics of test questions, it is necessary to design appropriate message propagation and node update mechanisms respectively. The constructed bidirectional dynamic graph neural network is called BDGN (Bidirectional Dynamic Graph Neural Network).

[0048] (2-1) Mining test information based on the one-way graph from students to test questions

[0049] The characteristics of the test questions can be reflected by the students' performance on the test questions. For example, the difficulty of the test questions. Most students get the questions wrong, which are difficult, while the correct rate of simple test questions is high. Generally speaking, the more people answer the questions, the more accurate the information analyzed is. Based on this, the attention mechanism is used for message propagation, with the aim of making the model pay attention to global information. The core formula of message propagation is as follows:

[0050]

[0051] Where Q L-1 is the L-1th layer of test embedding, S L-1 is the learner embedding of layer L-1, R is the response embedding, and d is the dimension of the embedding.

[0052] This method achieves this by aggregating the question embedding obtained in the message propagation stage and the question embedding of the previous layer. The update mechanism from the L-1th layer to the Lth layer is:

[0053]

[0054] in, Is the weight matrix, used to select and merge and Q L-1 Information in.

[0055] (2-2) Student status tracking based on the test question to student unidirectional graph

[0056] The degree of students' mastery of knowledge is reflected in their answering performance. For example, if students have a high degree of mastery of specific knowledge, they are more likely to correctly answer the corresponding test questions of this knowledge point. At the same time, students' knowledge status changes over time and is affected by factors such as knowledge consolidation, memory and forgetting. Based on this, the GRU network is used for message propagation, with the aim of exploring the law of students' knowledge status changing over time. The core formula is as follows:

[0057]

[0058] in, is a learnable parameter, σ is the Sigmoid activation function, ⊙ represents the element-wise multiplication of the vector, is the update gate used to update information, It is a reset gate used to filter information.

[0059] Similar to the question node update, the learner embedding obtained in the message propagation module and the learner embedding of the previous layer are aggregated to achieve The update mechanism from layer L-1 to layer L is:

[0060]

[0061] in, is a learnable parameter used to select and fuse and S L-1 Information in.

[0062] (3) Graph Multi-View Learning with Information Bottleneck Enhancement

[0063] In the subgraph selection in section (1-2), this method takes student s as the center and selects the questions that s has answered within a given time range, as well as other students who have answered these questions within this time range. This is just a screening of the student-question bipartite graph from the time range, and there are still many noisy interactions in the graph. Therefore, this method introduces the information bottleneck theory and optimizes the graph structure by adaptively learning whether to delete edges or nodes to avoid capturing irrelevant information between different perspectives.

[0064] (3-1) Node and edge deletion matrix calculation

[0065] This method uses Gumbel-Softmax to obtain the deletion matrix p L First, the original subgraph obtained after the subgraph selection in section (1-2) Input the bidirectional dynamic graph neural network (BDGN) designed in Section (2) to obtain the node embedding of each layer, that is, and edge embedding

[0066]

[0067] Note that here and in the following content, this method is based on replace That is, the original subgraph. L represents the Lth layer, S L represents student embedding, Q L Represents test question embedding.

[0068] Then use Multi-layer Perceptron (MLP) and Sigmoid activation function to learn the probability matrix of nodes and edges in each layer

[0069]

[0070] in N represents the number of nodes or edges, The value range is (0,1).

[0071] Finally, Gumbel-Softmax is used to Sampling is performed to decide whether to keep or remove the node edges.

[0072]

[0073] is a two-dimensional one-hot vector, Indicates that the node is retained. Indicates deleting the node. So, the final node and edge deletion matrix pL for:

[0074]

[0075] (3-2) Create the retained perspective after node deletion and the retained perspective after edge deletion

[0076] In order to minimize the impact of noisy student and test question nodes in the graph on test question information mining and student status tracking, this method creates the retained perspective after node deletion and the retained perspective after edge deletion based on the deletion matrix obtained in Section (3-1):

[0077]

[0078] in Represents the graph after deleting nodes at layer L, Represents the graph after deleting edges at layer L. p i L and The node deletion matrix and edge deletion matrix are used, where 0 means deleting the node or edge, and 1 means keeping the node or edge.

[0079] Preserve perspective after creating node deletion and the preserved perspective after edge deletion Finally, the two perspectives are sent to the bidirectional dynamic graph neural network (BDGN) to obtain the test node representation and student node representation of the retained perspective after node deletion and the retained perspective after edge deletion, respectively.

[0080] (3-3) Multi-view learning with information bottleneck enhancement

[0081] The goal of knowledge tracking alone cannot guide the deletion process well and create the best multi-view. Therefore, in order to ensure that each view retains as much task-related information as possible while reducing noisy interactions and redundant information, this method introduces the information bottleneck theory, called GIB (Graph Information Bottleneck). Specifically, the goal of GIB is defined as:

[0082]

[0083] GIB consists of three items, β 1 and β 2 There are two trade-off factors. The first The purpose of is to maximize the mutual information between the two views and the class label. The purpose of is to limit the information flow and minimize the shared information between the two views and the original image. The first two items encourage the model to optimize the intermediate representation and extract important interactive features from the views. The model is explicitly constrained to extract view-specific information for different views, ensuring and promoting view separation.

[0084] maximize The purpose is to hope that the model can make accurate predictions. Therefore, in order to reduce the complexity of the task, this method uses the method of minimizing the loss of the prediction task instead of optimizing Therefore, the optimization objective function of this method is converted into:

[0085]

[0086] in represents the sum of the answer prediction cross loss in knowledge tracking and the difficulty prediction cross loss in difficulty tracking. As for the mutual information of the last two items, this method estimates it by using variational self-distillation (VSD). The mutual information of the last two items consists of the mutual information from the student side and the test side, namely:

[0087]

[0088] (4) Selection and fusion of multi-perspective information

[0089] After obtaining the student and test question representations from the retained perspective after node deletion, and the student and test question representations from the retained perspective after edge deletion, this method needs to select and fuse the information from the two perspectives.

[0090] By constructing a gating network, we select and fuse multi-view information. First, we select the student node representation S that needs to be predicted from all node representations of each viewpoint according to the task. t and the question node representation Q t , and then concatenate the student node representation and question node representation from the two perspectives:

[0091]

[0092] in and d is the embedding dimension, NV represents the preserved perspective after node deletion, and EV represents the preserved perspective after edge deletion.

[0093] Then, the student node representation and question node representation obtained after the original subgraph passes through the bidirectional dynamic graph neural network (BDGN) are used as selectors, and the weight of each perspective is obtained through the MLP network and Softmax:

[0094]

[0095] in is the weight of the student node representation, It is about the weight of the test question node representation.

[0096] After obtaining the weights, select the information from the two perspectives. First, multiply the concatenated student and question node representations by the weights:

[0097]

[0098] in and The specific matrix is ​​expressed as:

[0099]

[0100] Then add the results of the two perspectives to get the final fused vector:

[0101]

[0102] (5) Student-Test Question Prediction

[0103] (5-1) Knowledge Tracking

[0104] After the above steps, the representation of the learner node is obtained Whether learners can answer questions correctly is affected by two factors: one is the state of students' knowledge, and the other is the characteristics of the test questions. For example, learners have mastered specific knowledge, but they may still make mistakes because the test questions are difficult; or learners have low knowledge, but the test questions are too simple to distinguish between students with different knowledge levels. Based on this, this method considers both the test questions and the students to make predictions.

[0105]

[0106] in, is the weight matrix, Represents the student's hidden knowledge state.

[0107] After obtaining the student's hidden knowledge state, the learner's future performance is predicted. The calculation method is as follows:

[0108]

[0109] is a learnable weight matrix, |Q| represents the number of all test questions in the data set, It indicates that the prediction is the learner's performance on all test questions. Then, according to the task, the performance of the specific test question corresponding to the label is selected.

[0110] (5-2) Difficulty Tracking

[0111] After the above steps, the representation of the test question is obtained. In order to help the model better learn the characteristics of test questions, a task-driven strategy is adopted, that is, adding difficulty tracking tasks based on knowledge tracking.

[0112] First, the difficulty of the test questions in the knowledge tracking dataset is defined as:

[0113]

[0114] Where q represents the test question, count q Indicates the number of people who answered the question, correct q Indicates the number of people who answered the question correctly. 0≤d q ≤10, d q There are 11 values ​​in total, namely N d =11. According to the formula, d q The larger the value of , the easier the test question is.

[0115] This method extracts the difficulty of the test questions by mining the students' performance on the test questions. Therefore, the grasp of the difficulty of the test questions is inseparable from the model's observation of the students' status. For example, if most students have a high level of knowledge but still get the question wrong, the difficulty of the test question is relatively high. Based on this, this method considers both the test questions and the students to predict the difficulty of the test questions.

[0116]

[0117] After getting the hidden state of the test question After that, the difficulty of the test questions is predicted. The calculation method is as follows:

[0118]

[0119] is the learnable weight matrix.

[0120] The contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A bidirectional dynamic graph knowledge tracking method with information bottleneck enhancement, Features The method comprises the following steps: (1) Construction of a bidirectional dynamic graph based on student-question interaction: Generate a dynamic heterogeneous graph based on the interaction data of students in the intelligent tutoring system, where nodes include students and questions, and edge features include responses and timestamps. Generate the original subgraph by subgraph selection; (2) Construction of a bidirectional dynamic graph neural network; mining test information through message propagation from students to test questions and node updates, and tracking student status through message propagation from test questions to students and node updates; (3) Information bottleneck enhanced graph multi-perspective learning: First, the node and edge deletion matrices are calculated based on the node representation and edge representation obtained after the original subgraph passes through the bidirectional dynamic graph neural network. The retained perspective after edge deletion and the retained perspective after node deletion are created through the deletion matrix. Then, multi-perspective learning is performed based on the information bottleneck theory to obtain the student node representation and test question node representation of each perspective. Specifically: (3-1) Calculation of node and edge deletion matrices: The node representation and edge representation obtained after the original subgraph passes through the bidirectional dynamic graph neural network are respectively input into the multilayer perceptron, Sigmoid activation function and Gumbel-Softmax to obtain the node deletion matrix and edge deletion matrix. The deletion matrix is ​​used to determine whether the edges and nodes are retained; (3-2) Creating a retained perspective after node deletion and a retained perspective after edge deletion: the original subgraph deletes certain nodes according to the node deletion matrix to obtain a retained perspective after node deletion; the original subgraph deletes certain edges according to the edge deletion matrix to obtain a retained perspective after edge deletion; (3-3) Multi-perspective learning enhanced by information bottleneck: The retained perspective after node deletion and the retained perspective after edge deletion are input into the bidirectional dynamic graph neural network to obtain the student node representation and the test question node representation; the optimization goal is designed based on the information bottleneck theory: in, represents the original subgraph, Indicates the retained perspective after the node is deleted, represents the retained view after edge deletion, Y represents the class label of the predicted target, and β 1 and β 2 ∈[0,1 are two trade-off factors, I(·) represents mutual information, and max represents maximization; the first term The purpose of is to maximize the mutual information between the two views and the class label. The purpose of is to limit the information flow and minimize the shared information between the two views and the original image. The first two items encourage the model to optimize the intermediate representation and extract important interactive features from the views. Explicitly constraining the model to extract view-specific information for different viewpoints, ensuring and promoting viewpoint separation; (4) Multi-perspective information selection and fusion: Use a gating network to select and fuse the student node representation and question node representation obtained from the retained perspective after edge deletion and the retained perspective after node deletion; (5) Student-question prediction: predicting the probability of a student correctly answering the next question and the difficulty of the question currently being answered by the student.

2. The information bottleneck enhanced bidirectional dynamic graph knowledge tracking method according to claim 1, Features The construction of the two-way dynamic graph based on student-question interaction described in step (1) is specifically as follows: (1-1) Global graph construction: First, a dynamic heterogeneous graph is constructed based on the students’ interaction data in the intelligent tutoring system. Where V represents the test question and student nodes, E represents the edge connecting the test question node and the student node, and θ represents the edge features, including the answer response and timestamp; (1-2) Subgraph selection: For each student s, a subgraph is created at each answering moment. At each time point t, first, with student s as the center, select the l closest to s in time from the global graph. s test questions as first-order neighbors, and then take each test question as the center and select l q The students connected to the test question q are taken as the second-order neighbors, and so on, the multi-hop neighbors of the student node s are obtained, forming About the k-order subgraph of student s 3. The information bottleneck enhanced bidirectional dynamic graph knowledge tracking method according to claim 1, Features The bidirectional dynamic graph neural network construction in step (2) is specifically as follows: (2-1) Mining test information based on the unidirectional graph from students to test questions: Use the attention mechanism to propagate messages from students to test questions, and then update the test question nodes in the graph by aggregating the test question embedding obtained in the message propagation stage and the test question embedding of the previous layer; (2-2) Student status tracking based on the unidirectional graph from test questions to students: Use the GRU network to propagate messages from test questions to students, and then update the student nodes in the graph by aggregating the student embeddings obtained in the message propagation stage and the student embeddings of the previous layer.

4. The information bottleneck enhanced bidirectional dynamic graph knowledge tracking method according to claim 1, Features The multi-view information selection and fusion in step (4) is specifically as follows: First, the student node representation and question node representation obtained after the original subgraph passes through the bidirectional graph neural network are used as selectors. The weights of the two perspectives are obtained through the MLP network and Softmax. Then the weights are multiplied with the corresponding perspectives, and the two perspective information after the multiplication are added together to obtain the fused feature vector.

5. The information bottleneck enhanced bidirectional dynamic graph knowledge tracking method according to claim 1, Features The student-question prediction in step (5) is specifically as follows: (5-1) Knowledge tracking: First, the multilayer perceptron and Sigmoid activation function are applied to the fused student node representation to obtain the weights of the student node representation and the question node representation for the knowledge tracking task. Then, the weights are multiplied by the corresponding node representation to obtain the student hidden state. Finally, the hidden state is activated by Sigmoid to predict the answer and predict the probability that the student will correctly answer the next question. (5-2) Difficulty tracking: First, the difficulty of the test questions in the knowledge tracking dataset is defined as: Where q represents the test question, count q Indicates the number of people who answered the question, correct q It represents the number of people who answered the question correctly. The multilayer perceptron and Sigmoid activation function are applied to the fused question node representation to obtain the question node and student node weights for the difficulty tracking task. The weights are then multiplied by the corresponding node representation to obtain the question hidden state. Finally, the hidden state is activated by Softmax to predict the difficulty of the question currently being answered by the student.

Images