Dynamic graph knowledge tracking method and system fusing behavior characteristics of learner
By using a dual-branch, dual-layer GRU structure and a multi-set indicator approach, the problems of complexity and poor scalability in the fusion of student behavioral features in existing technologies are solved, enabling efficient utilization and accurate prediction of educational data and reducing training costs.
Patent Information
- Application Number
- CN202510927307.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-07
- Publication Date
- 2025-10-28
AI Technical Summary
Existing knowledge tracking methods suffer from problems such as cumbersome feature engineering, complex internal structure, and poor scalability when integrating student behavioral characteristics, resulting in insufficient data utilization and high training costs.
We employ a two-branch, two-layer GRU structure, combined with multi-set indicators and multi-head attention mechanisms. By acquiring the historical records of students and questions, we generate embeddings of concepts, answers, timestamps, and behavioral features. We then utilize the connections and fully connected layers of the two-layer GRU to generate behavioral and knowledge level features. Finally, we predict the probability of correct answers using the sigmoid function.
It improves adaptability to different educational data, reduces training costs, and significantly enhances prediction accuracy and feature representation capabilities, while also allowing for easy modification of the number of fused features.
Smart Images
Figure CN120851160A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of educational data mining, specifically involving a dynamic graph knowledge tracking method and system that integrates learner behavioral characteristics. Background Technology
[0002] Knowledge tracing tasks predict students' knowledge status and future performance based on their historical learning records, assess student learning outcomes by mining educational data, and assign personalized learning plans. Existing knowledge tracing methods, such as Dynamic Graph Learning for Knowledge Tracing (DyGKT), use dynamic graph methods to process datasets, automatically extracting learning sequences from educational data, simplifying data processing, reducing training costs, and demonstrating high practicality. However, they do not consider the integration of student behavioral features, resulting in insufficient data utilization. Most existing knowledge tracing methods that integrate student behavioral features are extensions of Learning Behavior-oriented Knowledge Tracing (LBKT). While these methods can fully utilize student behavioral features, their complex internal structure, cumbersome feature engineering, and the need to modify the internal structure and conduct extensive validation experiments to add integrated features result in poor scalability. Summary of the Invention
[0003] In view of this, the purpose of this invention is to provide a dynamic graph knowledge tracking method and system that integrates learner behavioral characteristics. This method can easily modify the number of integrated features while ensuring practicality, and enhance the adaptability to different educational data.
[0004] The objective of this invention is achieved through the following technical solution:
[0005] A dynamic graph knowledge tracking method that integrates learner behavioral characteristics includes:
[0006] Given a student and a question, retrieve the corresponding specific student ID and specific question ID;
[0007] Based on a specific student ID and a specific question ID, retrieve the student's historical answer records and the question's historical answer records respectively. The student's historical answer records and the question's historical answer records form the student sequence and the question sequence respectively. Each student's historical answer record and each question's historical answer record contains the student ID, question ID, concept, answer, timestamp, and learner's behavioral characteristics.
[0008] For each student's historical answer record and each question's historical answer record, an encoder is used to generate embeddings for the concept, answer, timestamp, and feature in each record, denoted as concept embedding, answer embedding, time embedding, and feature embedding;
[0009] A multiset indicator is used to generate multiset embeddings in each record. For each student's historical answer record, two binary tokens are used to generate the multiset embedding. The first binary token is used to determine whether a given student has answered a given question, and the second binary token is used to determine whether a given student has answered the concept associated with the given question. For each question's historical answer record, a single binary token is used to generate the multiset embedding, where the binary token is used to determine whether a given question has been answered by a given student.
[0010] For student sequences and question sequences, a two-branch structure is used for knowledge tracing. Each branch adopts a two-layer GRU structure. The input of the first layer GRU in each branch is the concatenation of concept embeddings, answer embeddings, feature embeddings, and multiset embeddings in the sequence corresponding to that branch. The input of the second layer GRU in each branch is the concatenation of concept embeddings, answer embeddings, temporal embeddings, feature embeddings, and multiset embeddings in the sequence corresponding to that branch, as well as the result of multi-head attention on the hidden states from time 1 to time t of the first layer GRU of that branch.
[0011] The hidden states at time t of the first layer GRU in the two branches are connected and then fully connected to generate behavioral-level features.
[0012] The hidden states at time t of the second-layer GRU in the two branches are connected and then fully connected to generate knowledge level features.
[0013] The knowledge level features and behavioral level features are connected, and then the sigmoid function is used after two fully connected layers to obtain the probability that a given student will answer a given question correctly.
[0014] Furthermore, learner behavioral characteristics include one or more combinations of factors such as the speed at which students answer questions, the number of prompts required for a student to answer a question correctly, and the number of attempts a student makes to answer a question.
[0015] Furthermore, a fully connected layer is used to encode both concepts and responses.
[0016] Furthermore, for features, two fully connected layers are used for encoding, and each fully connected layer is followed by a ReLU function for activation.
[0017] Furthermore, the timestamp is encoded using a two-layer fully connected structure, and the time interval between two adjacent answers is compared with a time threshold. Different two-layer fully connected structures are used for cases where the time interval between two adjacent answers is greater than or less than the time threshold.
[0018] Furthermore, for each student's historical answer record, two binary tokens are used to generate multi-set embeddings. The first binary token is used to determine whether a given student has answered a given question, and the second binary token is used to determine whether a given student has answered the associated concept of a given question.
[0019] Furthermore, for each historical answer record of a question, a binary token is used to generate a multiset embedding, where the binary token is used to determine whether a given question has been answered by a given student.
[0020] Furthermore, the input of the second-layer GRU in each branch also includes the final hidden state of the first-layer GRU in the other branch.
[0021] This invention also provides a dynamic graph knowledge tracking system that integrates learner behavioral characteristics, the system comprising:
[0022] Memory, configured to store computer programs;
[0023] The processor is configured to execute the computer program to implement the dynamic graph knowledge tracing method that integrates learner behavioral characteristics as described above.
[0024] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the dynamic graph knowledge tracking method for integrating learner behavioral features as described above.
[0025] The beneficial effects of this invention are:
[0026] This invention acquires and fuses learners' behavioral characteristics, making full use of these characteristics to increase the accuracy of the final prediction results. Moreover, the method can adjust the number of incorporated behavioral features for different data.
[0027] The time encoder used in this invention can analyze the impact of students answering questions multiple times in a short period of time and answering questions for a long period of time on students' learning by setting a time threshold;
[0028] In the knowledge tracing of the two branches of this invention, the first layer GRU uses connections of other embeddings besides temporal embeddings, and the second layer GRU uses connections of all embeddings including temporal embeddings. It also uses the multi-head attention results of the output of the first layer GRU, which can capture long-term dependencies and key interactions, improve the expressive power of features, and improve the prediction performance of the method by cross-passing the final hidden state of the first layer GRU into the second layer GRU.
[0029] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:
[0031] Figure 1 This is a schematic flowchart illustrating a dynamic graph knowledge tracing method that integrates learner behavioral characteristics;
[0032] Figure 2 This is a schematic diagram of subgraph construction;
[0033] Figure 3 It is a diagram illustrating the encoding of various types of information;
[0034] Figure 4 This is a diagram illustrating knowledge tracing based on a two-layer GRU.
[0035] Figure 5 This is a schematic diagram of the prediction layer;
[0036] Figure 6 This is the processing structure of the ASSISTment17 dataset;
[0037] Figure 7 This is a comparison chart of the scores of models with different fusion features added. Detailed Implementation
[0038] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.
[0039] This invention provides a dynamic graph knowledge tracing method that integrates learner behavioral characteristics, including subgraph construction, information encoding, knowledge tracing, and prediction. Figure 1 This is a schematic flowchart of a dynamic graph knowledge tracing method that integrates learner behavioral characteristics.
[0040] like Figure 1 As shown, the method includes
[0041] Step 1: For a given student and a given question, obtain the corresponding specific student ID and specific question ID;
[0042] Step 2: Obtain the student's historical answer record and the question's historical answer record based on the specific student ID and the specific question ID (i.e., subgraph construction). The student's historical answer record and the question's historical answer record form the student sequence and the question sequence, respectively. Each student's historical answer record and each question's historical answer record contains the student ID, question ID, concept (i.e., the knowledge points involved in the question), answer (i.e., a description of whether the question was answered correctly), timestamp, and learner's behavioral characteristics.
[0043] Step 3: For each record, generate the embedding of the information included in each record. Specifically, for each student's historical answer record and each question's historical answer record, use the encoder to generate the embedding of concept, answer, timestamp and feature in each record, denoted as concept embedding, answer embedding, time embedding and feature embedding.
[0044] A multiset indicator is used to generate multiset embeddings in each record. For each student's historical answer record, two binary tokens are used to generate the multiset embedding. The first binary token is used to determine whether a given student has answered a given question, and the second binary token is used to determine whether a given student has answered the concept associated with the given question. For each question's historical answer record, a single binary token is used to generate the multiset embedding, where the binary token is used to determine whether a given question has been answered by a given student.
[0045] Step 4: Use a two-layer GRU structure for knowledge tracking. Specifically, a two-branch structure is used for knowledge tracking of student sequences and question sequences. Each branch uses a two-layer GRU structure. The input of the first layer GRU in each branch is the concatenation of concept embeddings, answer embeddings, feature embeddings, and multiset embeddings in the sequence corresponding to that branch. The input of the second layer GRU in each branch is the concatenation of concept embeddings, answer embeddings, temporal embeddings, feature embeddings, and multiset embeddings in the sequence corresponding to that branch, as well as the result of multi-head attention on the hidden states of the first layer GRU from time 1 to time t in that branch.
[0046] The hidden states at time t of the first layer GRU in the two branches are connected and then fully connected to generate behavioral-level features.
[0047] The hidden states at time t of the second-layer GRU in the two branches are connected and then fully connected to generate knowledge level features.
[0048] Step 5: Predict the accuracy of students' answers. Specifically, connect knowledge level features and behavioral level features, and then use the sigmoid function after two fully connected layers to obtain the probability that a given student will answer a given question correctly.
[0049] Figure 2 This is an illustrative example of subgraph construction. For example... Figure 2 As shown, the acquired dataset includes multiple student nodes and multiple question nodes, connected by edges. These nodes and edges form a dynamic graph, where nodes and edges can have their attributes added, deleted, or modified over time. Each edge describes a student's answer to a question, with attributes including a timestamp (the time the student answered the question), the correctness of the answer (e.g., 1 for correct, 0 for incorrect), and the associated knowledge concept (a question may contain one or more concepts).
[0050] In some embodiments, the attributes of an edge may also include learner (i.e., student) behavioral characteristics. These learner behavioral characteristics describe aspects of a student's answering of questions, including one or more combinations of factors such as the student's answering speed, the number of prompts required for a correct answer, and the number of attempts made. The specific content and number of learner behavioral characteristics can be selected based on the specific dataset.
[0051] Given a student (corresponding to a specific student ID) and a given question (corresponding to a specific question ID), the student's historical answer records can be retrieved based on that specific student ID. For example, the most recent N historical records of a given student can be retrieved. Figure 2 In the context of time step t, given a student S... t Get some of the questions it answered in the past (q1, q2, ..., q...). t-1 This refers to the records of a question. For a given question (i.e., its specific question ID), you can retrieve the historical answers to that question. For example, you can retrieve the most recent N historical records for that given question. Figure 2 In the context of the problem q given at the current time step (i.e., time step t),... t Get some students (S1, S2, ..., S...) who have answered this question in the past. t-1 The records include the student ID, question ID, concept, answer, timestamp, and characteristics (i.e., the learner's behavioral characteristics).
[0052] After the subgraph is constructed, the encoding layer can be entered to encode the student and question sequences extracted from the subgraph. In this invention, different encoding strategies are used for different types of information. For example... Figure 3 As shown, for both answers and concepts, a fully connected layer (i.e., ...) is used. Figure 3 The FC in the code is encoded to generate answer embeddings and concept embeddings.
[0053] For the features (i.e. the learner’s behavioral features), two fully connected layers are used for encoding. Each fully connected layer (i.e., FC) is followed by a ReLU function for activation, thereby generating feature embeddings.
[0054] The timestamp is encoded using a two-layer fully connected structure, and the time interval between two adjacent responses (i.e., Figure 3 Δt in i or Δt j The time interval between two consecutive answers is compared with a time threshold (denoted as ΔT). For cases where the time interval between two consecutive answers is greater than or less than the time threshold, different two-layer fully connected structures (i.e., ...) are used. Figure 3 (Two branches in the process). Δt i This represents the time interval between two answers in the student sequence (i.e., the time interval between two answers given by a student in time t). i Time has answered a question, in t i-1 Time has also answered a question, Δt i That is, the interval between these two times); Δt j This represents the time interval between two answers in the question sequence (i.e., the time interval between two answers to a question in time t). j The time was answered once, in t j-1 Time was also answered once, Δt j This refers to the interval between these two times, and this method distinguishes the impact of short-time and long-time question-answering on changes in students' knowledge status.
[0055] If the time interval between two consecutive responses is greater than the time threshold ΔT, the first set of weights and biases is used in the first fully connected layer. If the time interval between two consecutive responses is not greater than the time threshold ΔT, the second set of weights and biases is used in the first fully connected layer. The second fully connected layer uses the same weights and biases. The first fully connected layer can also be activated using the ReLU function.
[0056] By using a two-layer fully connected structure, time encoding can be achieved to generate temporal embeddings. By using a time threshold ΔT as a criterion, time intervals greater than and less than the time threshold ΔT are passed to different two-layer fully connected structures. In this way, the impact of students answering questions multiple times in a short period of time and answering questions for a long period of time on student learning can be analyzed, that is, the impact of time on forgetting and memory can be modeled.
[0057] For each student's historical answer record, a multiset embedding is generated using two binary tokens (i.e., two 0-1 vectors). One binary token is used to determine whether a given student has answered a given question. Where s1 and q1 are the predicted students and questions. This is a problem in the student sequence, if If q1 is true, then it is marked as 1; otherwise, it is marked as 0. Another binary flag is used to determine whether a given student has answered a given question related to a concept. Where k1 is the corresponding concept for the prediction problem. It is a concept in the student sequence, if If k1 is true, then the tag is 1; otherwise, it is 0. By summing these two binary tags, a merge tag is generated. Using the merge tag, a multi-set embedding for each student's historical answer record is generated.
[0058] For each record of a question's answer history, a binary token is used to generate a multiset embedding. This binary token is used to determine whether a given question has been answered by a given student, i.e., whether the obtained record of the question's answer history includes a specific student. If a student is in the problem sequence, then If it is s1, then it is marked as 1; otherwise, it is marked as 0.
[0059] After acquiring the embedded information, we can proceed to step 4 for knowledge tracing. This invention employs a knowledge tracing method based on a two-layer GRU. Figure 4 This is a schematic diagram of knowledge tracing based on a two-branch, two-layer GRU structure.
[0060] like Figure 4 As shown, a two-branch structure is used for knowledge tracing for student sequences and problem sequences; that is, different branches are used to process student qualifications and problem sequences. Each branch adopts a two-level GRU structure.
[0061] The input to the first layer of the GRU in each branch is a concatenation of the concept embeddings, response embeddings, feature embeddings, and multiset embeddings from the sequence corresponding to that branch. Specifically, the input to the first layer of the GRU in the first branch (i.e., the student sequence branch) is a concatenation of the concept embeddings, response embeddings, feature embeddings, and multiset embeddings obtained by encoding the student's historical response records (the concatenation result can be denoted as the "student learning sequence," which does not include temporal embeddings). The input to the first layer of the GRU in the second branch (i.e., the question sequence branch) is a concatenation of the concept embeddings, response embeddings, feature embeddings, and multiset embeddings obtained by encoding the question's historical response records (the concatenation result can be denoted as the "question learning sequence," which does not include temporal embeddings). The output of the first layer of the GRU is the hidden state corresponding to all time steps (i.e., 1, 2, 3...t). The hidden state at time t of the first layer GRU in the first branch represents the student's behavioral habits (i.e., some behavioral characteristics of the student when answering the question, such as the student's answering speed, number of prompts, number of attempts, etc.); the hidden state at time t of the first layer GRU in the second branch represents the question's behavioral characteristics (i.e., the speed at which the question is answered, the number of prompts when the question is answered correctly, number of attempts, etc.).
[0062] The input to the second-layer GRU in each branch is the concatenation of concept embeddings, answer embeddings, temporal embeddings, feature embeddings, and multi-set embeddings in the sequence corresponding to that branch (i.e., all previously generated embedding information with the corresponding temporal embedding added) and the output of the first-layer GRU of that branch after multi-head attention. In some embodiments, to further enhance the fusion of question and student features, the input to the second-layer GRU in each branch also includes the final hidden state of the first-layer GRU of another branch, i.e., through cross-connection (i.e., ... Figure 4 The content represented by the black dotted lines in the text is passed to the second-level GRU.
[0063] Next, the hidden states of the first-level GRU in the two branches at time t are concatenated and then fully connected (e.g., ...). Figure 4 In this process, the results representing students' behavioral habits and problem behaviors are processed through connections, i.e. Figure 4 In Then, by passing in "Linear", it can also be activated by the ReLU function to generate behavioral level features (i.e., the behavioral level features of the question, which characterize whether the question is easy to answer correctly, and predict the behavior that students may exhibit when answering questions by using students' learning habits and the behavioral level of the question).
[0064] The final hidden states of the second-level GRU in the two branches represent the student's knowledge state and the difficulty of the problem, respectively. Figure 4The knowledge level features (i.e., the student's knowledge level) can be generated by connecting the final hidden states of the second-layer GRU in the two branches and then performing a full connection, and can also be activated by the ReLU function (i.e., the student's knowledge level, and predicting the student's application of knowledge when answering questions based on the student's knowledge state and the difficulty of the questions).
[0065] Next, in step 5, the accuracy of students' answers to the questions is predicted. Specifically, Figure 5 This is a schematic diagram of the prediction layer, such as... Figure 5 As shown, knowledge level features and behavioral level features can be connected (i.e. Figure 5 In The algorithm uses two fully connected layers followed by a sigmoid function (with ReLU activation between the two fully connected layers) to obtain the probability that a given student answers a given question correctly (i.e., the probability of the student answering the question correctly). ).
[0066] The methods described above can be used to assess any student's understanding of any problem, enabling educators to grasp each student's knowledge level or guide learners in developing personalized learning plans.
[0067] The method of this invention has low training cost, can automatically adapt to the ever-increasing educational data, and only requires modifying the parameter dimension of the feature encoder to change the number of fused features. Compared with the original method, the prediction accuracy is significantly improved.
[0068] This invention also provides a dynamic graph knowledge tracking system that integrates learner behavioral characteristics, the system comprising:
[0069] Memory, configured to store computer programs;
[0070] The processor is configured to execute the computer program to implement the dynamic graph knowledge tracing method that integrates learner behavioral characteristics as described above.
[0071] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the dynamic graph knowledge tracking method for integrating learner behavioral features as described above.
[0072] The advantages of this invention will be illustrated below through specific experiments.
[0073] The main parameters and their settings are as follows: batch size during training is 2000, number of neighbors is 50 (i.e., the length of the student sequence and question sequence extracted from the subgraph), dropout rate is 0.1, number of multi-head attention is 2, learning rate is 0.001, the optimizer is Adam, and weight decay rate is set to 0.0001.
[0074] This experiment was conducted on three real-world datasets (ASSISTment12, ASSISTment17, and Junyi), with ASSISTment17 used as an example to demonstrate the usage process.
[0075] First, the training process is demonstrated, processing the ASSISTment17 dataset as follows: Figure 6 The structure. In Figure 6 In the table, the first column is the index; the second column, user_id, represents the "student ID"; the third column, problem_id, represents the "problem ID"; the fourth column, end_time, represents the "timestamp" corresponding to the student's answer (the time the student submitted their answer); the fifth column, correct, represents the "answer" (i.e., whether the student's answer is correct or incorrect, 1 for correct and 0 for incorrect); the sixth column, skill_id, represents the "concept" (i.e., the knowledge points contained in the question, also known as skills); the seventh column, hint_count, represents the "number of hints"; the eighth column, attempt_count, represents the "number of attempts"; and the ninth column, speed, represents the "response speed". The data is sorted from low to high end_time. The number of features incorporated can be modified according to the dataset; here, only 3 features are incorporated.
[0076] Through the Figure 6The table content shown is processed to generate three files: ml_assist17.csv, ml_assist17.npy, and ml_assist17_node.npy. These three files constitute a dynamic graph storage structure. This method automatically extracts relevant data after loading the dynamic graph, without manual intervention. The dynamic graph structure is briefly introduced here. ml_assist17.csv is a list of edges, and the data is arranged in chronological order. It mainly contains the edge ID (the edge ID is the index of the record from 1 to n. In the dynamic graph, each record is regarded as an edge from the node (student) to the node (question), so it is also called edge ID), node ID and time information (such as table = (1, 20, 30, 1000), which means that the first record (edge ID) is student 20 who answered question 30 at time 1000); ml_assist17.npy is a list of edge information. In this method, it contains the student's answer status and student characteristic information. Information can be extracted by edge ID (such as array[1] = (1, 49, 2, 0, ...), where the edge ID is used to extract the information of the first record. 1 means that the student answered correctly, 49 means that the student took 49 seconds to answer, 2 means that the student tried 2 times, 0 means that the student did not use the hint, and other characteristic information can be added later). ml_assist17_node.npy contains node information. This method mainly stores the skills corresponding to the questions (e.g., array
[50] = (3, ...), which means that the skill corresponding to question 50 is 3).
[0077] The corresponding training file DyGLib has been configured; simply load the model to begin training. When starting training, adjust the parameter dimensions of the feature encoder according to the number of features incorporated (three learner behavioral features are used here). For higher performance, you can increase the `num_neighbors` parameter setting, but this will slow down the training process.
[0078] To demonstrate the effectiveness of this method, a comparative experiment with DyGKT is presented below. To facilitate comparison of the impact on different datasets, this experiment only incorporates three features common to the datasets. The results are shown in Table 1.
[0079] Table 1 Comparison of experimental results between the present invention (i.e., BEDGKT) and DyGKT
[0080]
[0081] In Table 1, AP represents the average precision, and AUC represents the area under the ROC curve. Table 1 shows that our method achieves a significant improvement over DyGKT on the three datasets.
[0082] Fusing more features can improve the prediction level. In subsequent experiments, five features were added to the ASSISment17 dataset for fusion experiments. Figure 7 This is a comparison chart of the scores of models with different fusion features added. Transduction learning refers to the model's prediction performance on the entire test set (students and questions that never appeared in the training set are also included in the prediction). The left chart shows the results of "Transduction Learning AP", and the right chart shows the results of "Transduction Learning AUC". Figure 7 The legend in the diagram represents the AP or AUC results when the model adds features to the original three features (number of attempts, response speed, and number of hints); frIsHelpRequest indicates whether help was used on the first response; AveCorrect represents the average accuracy; bottomHint indicates whether the final hint was used; AveKnow represents the learner's knowledge state level predicted by BEDGKT in the dataset; original indicates that the problem is the original problem rather than a scaffolding problem; allfeature indicates that all the above features are added to the model. Figure 7 As shown, the model that partially integrates 4 features outperforms the model that integrates 3 features. The model that integrates 8 features has an AUC of 87.39%, while the model that integrates 3 features has an AUC of 86.46%, an improvement of 0.93%. Moreover, the training costs are basically the same in the experiment. Therefore, in practical use, the model can be optimized by selecting the integration features.
[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A dynamic graph knowledge tracking method that integrates learner behavioral characteristics, characterized in that, include: Given a student and a question, retrieve the corresponding specific student ID and specific question ID; Based on a specific student ID and a specific question ID, retrieve the student's historical answer records and the question's historical answer records respectively. The student's historical answer records and the question's historical answer records form the student sequence and the question sequence respectively. Each student's historical answer record and each question's historical answer record contains the student ID, question ID, concept, answer, timestamp, and learner's behavioral characteristics. For each student's historical answer record and each question's historical answer record, an encoder is used to generate embeddings for the concept, answer, timestamp, and feature in each record, denoted as concept embedding, answer embedding, time embedding, and feature embedding; A multiset indicator is used to generate multiset embeddings in each record. For each student's historical answer record, two binary tokens are used to generate the multiset embedding. The first binary token is used to determine whether a given student has answered a given question, and the second binary token is used to determine whether a given student has answered the concept associated with the given question. For each question's historical answer record, a single binary token is used to generate the multiset embedding, where the binary token is used to determine whether a given question has been answered by a given student. For student sequences and question sequences, a two-branch structure is used for knowledge tracing. Each branch adopts a two-layer GRU structure. The input of the first layer GRU in each branch is the concatenation of concept embeddings, answer embeddings, feature embeddings, and multiset embeddings in the sequence corresponding to that branch. The input of the second layer GRU in each branch is the concatenation of concept embeddings, answer embeddings, temporal embeddings, feature embeddings, and multiset embeddings in the sequence corresponding to that branch, as well as the result of multi-head attention on the hidden state at time 1-t of the first layer GRU of that branch. The hidden states at time t of the first layer GRU in the two branches are connected and then fully connected to generate behavioral-level features. The hidden states at time t of the second-layer GRU in the two branches are connected and then fully connected to generate knowledge level features. The knowledge level features and behavioral level features are connected, and then the sigmoid function is used after two fully connected layers to obtain the probability that a given student will answer a given question correctly.
2. The dynamic graph knowledge tracking method integrating learner behavioral features according to claim 1, characterized in that, Learner behavioral characteristics include one or more combinations of factors such as the speed at which students answer questions, the number of prompts required for a student to answer a question correctly, and the number of attempts a student makes to answer a question.
3. The dynamic graph knowledge tracking method integrating learner behavioral features according to claim 1, characterized in that, Both concepts and answers are encoded using a fully connected layer.
4. The dynamic graph knowledge tracking method integrating learner behavioral features according to claim 1, characterized in that, For features, two fully connected layers are used for encoding, and each fully connected layer is followed by a ReLU function for activation.
5. The dynamic graph knowledge tracking method integrating learner behavioral features according to claim 1, characterized in that, The timestamp is encoded using a double-layer fully connected structure, and the time interval between two adjacent answers is compared with a time threshold. Different double-layer fully connected structures are used for cases where the time interval between two adjacent answers is greater than or less than the time threshold.
6. The dynamic graph knowledge tracking method integrating learner behavioral features according to claim 1, characterized in that, For each student's historical answer record, two binary tokens are used to generate multiset embeddings. The first binary token is used to determine whether a given student has answered a given question, and the second binary token is used to determine whether a given student has answered the concept associated with the given question.
7. The dynamic graph knowledge tracking method integrating learner behavioral features according to claim 1, characterized in that, For each historical record of a question being answered, a binary token is used to generate a multiset embedding, where the binary token is used to determine whether a given question has been answered by a given student.
8. The dynamic graph knowledge tracking method integrating learner behavioral features according to claim 1, characterized in that, The input of the second-layer GRU in each branch also incorporates the hidden state at time t of the first-layer GRU of another branch, i.e., cross-connection.
9. A dynamic graph knowledge tracking system that integrates learner behavioral characteristics, characterized in that, include: Memory, configured to store computer programs; The processor is configured to execute the computer program to implement the dynamic graph knowledge tracing method that integrates learner behavioral features as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the dynamic graph knowledge tracking method that integrates learner behavioral features as described in any one of claims 1 to 8.