A knowledge tracking system based on heterogeneous graph contrastive learning

By constructing a hierarchical semantic heterogeneous graph network and structural contrastive learning, the shortcomings of existing knowledge tracing methods in terms of structural statics and singular expressions are addressed. This enables multidimensional prediction and evaluation of students' learning behaviors and enhances the support capabilities for personalized teaching.

CN120996156APending Publication Date: 2025-11-21ZHAOQING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511024281.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-24
Publication Date
2025-11-21

AI Technical Summary

Technical Problem

Existing knowledge tracing methods are unable to fully reflect the multi-level and multi-type structural relationships of students' learning behaviors, ignore the complex semantics and heterogeneous associations in educational scenarios, and lack in-depth modeling of structural evolution information during the learning process. As a result, the models are insufficient in accuracy and generalization ability, making it difficult to meet the needs of personalized teaching.

Method used

A hierarchical semantic heterogeneous graph network is constructed using a heterogeneous graph contrastive learning approach. Multi-scale structural modeling is performed through graph neural networks. By combining structural contrastive learning and perturbation consistency modeling, a comprehensive prediction of students' answer performance, knowledge mastery level, learning transfer potential, and learning path stability can be achieved.

Benefits of technology

It improves the model's accuracy and flexibility in characterizing student learning behavior, quantifies higher-order learning indicators, enhances the adaptability of personalized learning processes, and strengthens the robustness to structural changes and the interpretability of prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120996156A_ABST
    Figure CN120996156A_ABST
Patent Text Reader

Abstract

The application discloses a kind of knowledge tracking systems based on heterogeneous graph contrast learning, comprising: data preprocessing module, for collecting educational data and pre-processing;Heterogeneous graph construction module, for constructing educational semantic heterogeneous graph;Graph structure evolution module, for the student behavior stage division of educational semantic heterogeneous graph, constructs time evolution sequence graph;Student behavior subgraph extraction module, for extracting associated structure neighborhood from time evolution sequence graph, generates student behavior subgraph;Hierarchical structure modeling module, for constructing hierarchical semantic heterogeneous graph network, and student behavior subgraph node, edge and subgraph structure hierarchy are modeled by structure representation;Structure contrast optimization module, for designing structure hierarchy contrast learning mechanism and structure disturbance confrontation mechanism, jointly optimize hierarchical semantic heterogeneous graph network;Knowledge tracking module, for realizing multidimensional prediction and evaluation to student knowledge state.The application improves the evaluation capability and generalization robustness of knowledge tracking system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to a knowledge tracking system based on heterogeneous graph contrastive learning. Background Technology

[0002] With the development of artificial intelligence and intelligent education technology, knowledge tracking technology has gradually become an important research direction in the field of intelligent education. Knowledge tracking models students' historical learning behavior data to dynamically assess their knowledge mastery and predict their future learning performance, thereby supporting personalized teaching recommendations and interventions. Existing knowledge tracking methods are mainly divided into two categories: sequence-based modeling and graph-based modeling. Sequence-based methods typically use recurrent neural networks or Transformer models, utilizing the time information and feedback results in the order in which students answer questions for prediction. However, these methods simplify student learning behavior into a linear sequence, ignoring the multi-level and multi-type structural relationships between students, questions, and knowledge points, making it difficult to fully reflect the complex semantics and heterogeneous relationships existing in educational scenarios.

[0003] In recent years, graph structure methods have used knowledge graphs or graph neural networks to model students, questions, and knowledge points as nodes in a graph, with the relationships between student behaviors or questions as edges, in order to capture the structural information inherent in the learning scenario. By introducing graph convolutional networks or graph attention networks, the structural relationships between nodes can be effectively modeled, thereby improving the modeling accuracy and prediction effect of student knowledge states. However, existing graph structure methods are generally based on static graph construction, assuming that the structure of the educational scenario remains unchanged during the learning process, making it difficult to reflect the dynamic structural changes caused by the evolution of student learning behaviors over time. Furthermore, most methods use a single type of node and edge for modeling, ignoring the multiple types of entities and their diverse semantic relationships in educational data. This results in significant limitations in expressing the heterogeneity of educational semantics and modeling complex behavioral interactions, thus restricting the accuracy and generalization ability of knowledge tracking.

[0004] Furthermore, existing knowledge tracing models primarily focus on answer accuracy, lacking in-depth modeling of the structural evolution of students' learning processes. In particular, they fail to fully explore the comparative learning potential between different structural levels, resulting in insufficient sensitivity and discernment of potential microstructural changes in students' learning trajectories. Simultaneously, current technologies are inadequate in quantitatively assessing higher-order cognitive characteristics such as students' knowledge transfer abilities and the stability of their learning paths, making it difficult to meet the practical needs of personalized support and refined intervention in intelligent teaching.

[0005] Therefore, how to provide a knowledge tracking system based on heterogeneous graph contrastive learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One object of the present application is to provide a knowledge tracking system based on heterogeneous graph contrastive learning. The present application combines graph neural networks, multi-scale structure modeling and contrastive learning optimization methods, constructs a hierarchical semantic heterogeneous graph network, and performs structure contrastive learning and structure disturbance consistency modeling at the node level, edge level and subgraph level, which can realize comprehensive prediction of student performance, knowledge mastery level, learning transfer potential and learning path stability, has the advantages of high structure modeling accuracy, sensitivity to time sequence changes and strong interpretability of prediction results, and is suitable for personalized learning analysis and intelligent teaching recommendation scenarios.

[0007] According to an embodiment of the present application, a knowledge tracking system based on heterogeneous graph contrastive learning comprises the following steps:

[0008] A data preprocessing module is used to collect and preprocess educational data to form standardized educational data.

[0009] A heterogeneous graph construction module is used to construct an educational semantic heterogeneous graph through a behavior semantic edge modeling mechanism.

[0010] A graph structure evolution module is used to divide the educational semantic heterogeneous graph into student behavior stages according to the time dimension and construct a time evolution sequence graph.

[0011] A student behavior subgraph extraction module is used to extract a correlation structure neighborhood from the time evolution sequence graph and generate a student behavior subgraph.

[0012] A hierarchical structure modeling module is used to construct a hierarchical semantic heterogeneous graph network and model the structure representation of student behavior subgraph nodes, edges and subgraph structure levels.

[0013] A structure contrastive optimization module is used to design a structure level contrastive learning mechanism and a structure disturbance adversarial mechanism to jointly optimize the hierarchical semantic heterogeneous graph network.

[0014] A knowledge tracking module is used to model and reason the optimized structure semantic features to realize multi-dimensional prediction and evaluation of student knowledge state.

[0015] Optionally, the modules are realized through the following methods:

[0016] S1, collect and preprocess educational data to form standardized educational data.

[0017] S2, based on the standardized educational data, an educational semantic heterogeneous graph is constructed through a behavior semantic edge modeling mechanism.

[0018] S3, the educational semantic heterogeneous graph is divided into student behavior stages through a time evolution modeling mechanism to construct a time evolution sequence graph.

[0019] S4, extracting a correlation structure neighborhood from the time evolution sequence diagram according to a historical behavior track of the target student, and generating a student behavior subgraph;

[0020] S5, modeling the student behavior subgraph in a node, an edge, and a subgraph structure level by a hierarchical semantic heterogeneous graph network, to generate a node structure feature, an edge level structure feature, and a subgraph level structure feature;

[0021] S6, designing a structure level contrast learning mechanism and a structure disturbance adversarial mechanism, and jointly optimizing a hierarchical semantic heterogeneous graph network parameter;

[0022] S7, based on the optimized node structure feature, the edge level structure feature, and the subgraph level structure feature, performing multi-dimensional prediction and evaluation on a student knowledge state.

[0023] Optionally, the education data includes student answer records, question texts, question attribute information, knowledge point labels, question type classification information, and answer timestamps, and the preprocessing includes: removing outliers, completing missing items, unifying formats, and normalizing identification codes of the education data, and performing time sorting on the student answer records based on the answer timestamps to form standardized education data.

[0024] Optionally, the step S2 specifically includes:

[0025] S21, based on the standardized education data, extracting student entities, question entities, knowledge point entities, and question type entities by a field mapping and attribute label analysis method, and constructing a node set, the node set including a student node set, a question node set, a knowledge point node set, and a question type node set;

[0026] S22, constructing a semantic edge set by a behavior semantic edge modeling mechanism, the behavior semantic edge modeling mechanism combining a behavior category and a structure semantic, generating an edge set containing multiple semantic relationships based on a student answer behavior, a question and knowledge point mapping relationship, a question and question type attribution relationship, and a question dependency relationship, the semantic edge set including an answer behavior edge, a knowledge point association edge, a question type attribution edge, and a question dependency edge;

[0027] S23, constructing an education semantic heterogeneous graph by the node set and the semantic edge set;

[0028] S24, setting an initial attribute vector for each type of node, wherein the student node is set as a student attribute vector, the question node is set as a question information vector, the knowledge point node is set as a knowledge point feature vector, and the question type node is set as a question type category vector, and the attribute vectors of various nodes are collectively constructed into an initial node feature set;

[0029] S25, recording a semantic type and a connection relationship of each edge.

[0030] Optionally, the time evolution modeling mechanism is to divide the educational semantic heterogeneous graph into multiple stage graphs according to the answering time, and to construct a sequence of graph structures arranged in time sequence, and the step S3 specifically comprises:

[0031] S31, according to the answering time stamp in the standardized education data, the answering behavior of each student is divided into several continuous time stages in time sequence, forming a time stage set;

[0032] S32, for each time stage t i , the active nodes and associated semantic edges in the current stage are extracted from the educational semantic heterogeneous graph to construct the corresponding time slice graph;

[0033] S33, the time slice graphs are arranged in time sequence to construct a sequence of time evolution graphs;

[0034] S34, a structure evolution pair is constructed between any two continuous time slice graphs, and the structure evolution pair establishes the time sequence correspondence between the graph structures by comparing the node state changes, edge semantic type adjustments and local connection structure differences between adjacent time slice graphs.

[0035] Optionally, the step S4 comprises generating a student behavior subgraph based on the behavior trajectory of the target student, specifically comprising the following steps:

[0036] S41, according to the unique identification information of the target student, the historical answering behavior sequence of the target student is extracted from the standardized education data, and combined with the time evolution graph sequence, a plurality of time slice graphs matched with the student answering time period are selected;

[0037] S42, set the neighborhood radius r, take the student node corresponding to the target student as the center, and limit the structure depth and edge type in the node selection process, select the r-hop semantic neighbor nodes connected with the target student from the time slice graph, construct a set of semantic neighbor nodes, and extract a set of corresponding semantic edges, generate a local structure subgraph of the target student;

[0038] S43, for each candidate node in the local structure subgraph, define a set of meta-paths, the meta-path is composed of a pre-defined sequence of node types and a sequence of edge types;

[0039] S44, based on the meta-path, the embedding vector of the candidate node and the student node in the meta-path semantic is generated by the graph embedding method, and the structure correlation degree of the two is obtained by the normalized dot product similarity;

[0040] S45, the structure correlation degrees on all paths are weighted and fused to obtain a weighted structure correlation degree;

[0041] S46, set an association strength threshold value, filter the candidate nodes and the semantic edges corresponding to the candidate nodes according to the weighted structure association degree value, remove the nodes and edges with a structure association degree lower than the threshold value with respect to the target student node structure, only keep the nodes and edges with a weighted structure association degree greater than or equal to the association strength threshold value, and generate a stage filtering graph;

[0042] S47, sequentially combine all the stage filtering graphs in time sequence to form a behavior structure subgraph facing the target student.

[0043] Optionally, the step S5 comprises:

[0044] S51, construct a hierarchical semantic heterogeneous graph network, and model the student behavior subgraph at a node level, an edge level and a subgraph level, wherein the hierarchical semantic heterogeneous graph network comprises a node-level semantic fusion module, an edge-level semantic aggregation module and a subgraph-level structure fusion module;

[0045] S52, input the student behavior subgraph into the node-level semantic fusion module, and perform feature extraction and multi-semantic information fusion on each type of node in the student behavior subgraph;

[0046] By inheriting the initial attribute vector of each node in the student behavior subgraph set in the educational semantic heterogeneous graph as an input feature, and setting a type mapping function to encode the node features in a type-specific manner, a typed feature is obtained;

[0047] According to the context of the multiple meta-paths involved by the node in the student behavior subgraph, the structural context features are calculated according to the path attention mechanism, and the global behavior statistical features of the target student are introduced to supplement the role information;

[0048] The typed feature, the structural context feature and the global behavior statistical feature are fused to generate a node structure feature;

[0049] S53, input the student behavior subgraph into the edge-level semantic aggregation module, and model the semantic edge;

[0050] An edge type perception mechanism is introduced, which sets an edge semantic gating function for different types of semantic edges in the student behavior subgraph, calculates an edge-level information transmission weight, and dynamically adjusts the information transmission strength of each edge according to the structural feature difference between the starting node and the target node connected by the edge, and combines the edge semantic type;

[0051] The multi-type neighbor node information of the target node is weighted and aggregated by the edge-level information transmission weight to generate an edge-level structure feature;

[0052] S54, the node structure feature vector and the edge-level structure feature vector are fused at a structure level by a subgraph-level structure fusion module;

[0053] The node structure feature vector and the edge level structure feature vector are fused by a residual connection manner to generate a multi-layer substructure feature;

[0054] A path semantic learning function is designed to model semantics of the multi-layer substructure feature, adaptive weights are assigned based on path types, and the multi-layer substructure features are combined by weighting to generate a subgraph level structure feature.

[0055] Optionally, the step S6 specifically includes:

[0056] S61, a structure level contrast learning mechanism is constructed to respectively establish a contrast learning process for a student behavior subgraph at a node level, an edge level and a subgraph level;

[0057] The node level constructs a structure triple (v + ,v - ) for each target node v based on a time evolution graph sequence, wherein a positive sample node v + represents a corresponding node having similar types, connection relationships and local semantic structures in adjacent time slice graphs, and a negative sample node v - represents a heterogeneous node having obvious changes in structure topology or semantic features;

[0058] The edge level constructs an edge level structure contrast triple (e + ,e - ) based on a semantic edge as a basic unit, wherein a positive sample edge e + represents a semantic edge connecting the same node pair and having an unchanged edge type, and a negative sample edge e - represents an edge having a changed edge type, direction or topological position while the node pair is consistent;

[0059] The subgraph level constructs a subgraph contrast sample triple (G based on a local structure subgraph as a modeling unit, wherein a positive sample subgraph G represents adjacent time subgraphs having stable structure topology and semantic edges, and a negative sample subgraph G represents a phased subgraph having a significantly changed structure activity or a reconstructed connection mode;

[0060] A structure level contrast loss function is formed by maximizing the structure representation similarity between positive samples and minimizing the structure representation similarity between negative samples;

[0061] S62, based on the student behavior subgraph, a structure disturbance adversarial mechanism is introduced to obtain a structure disturbance subgraph by performing slight disturbance on node connection structure, semantic edge type and subgraph topological structure;

[0062] S63, an adversarial consistency loss term is constructed by modeling the adversarial consistency of the student behavior subgraph and the structure disturbance subgraph by using a mean square error.

[0063] S64, the structure level contrast loss is jointly modeled with the adversarial perturbation consistency loss to form an optimization objective function, and the node level semantic fusion module parameters, the edge level semantic aggregation module parameters and the subgraph level structure fusion module parameters in the hierarchical semantic heterogeneous graph network are jointly optimized to obtain the optimized.

[0064] Optionally, the step S7 specifically comprises:

[0065] S71, the optimized node level structure features, edge level structure features and subgraph level structure features are fused and modeled to extract a structure semantic vector of the target student;

[0066] S72, based on the structure semantic vector, the behavior performance of the target student in future learning activities is comprehensively modeled to obtain an answer correctness probability, a knowledge point mastery degree vector, a learning transfer potential score and a learning path stability index;

[0067] S73, the structure semantic vector is matched with the node features of the to-be-predicted question to predict the answer correctness probability of the student to the question title;

[0068] S74, based on the relationship between the structure semantic vector and the knowledge point node, a multi-label classifier is used to predict the mastery degree of the student to multiple knowledge points involved in the current stage, and a knowledge point mastery score set is output;

[0069] S75, the transfer deviation between the historical stage structure semantic vector and the current fusion representation is modeled, the ability of the student to transfer existing knowledge to new question types or new knowledge points in the learning process is quantified, and a learning transfer potential score is output;

[0070] S76, according to the structure perturbation subgraph, the consistency score of the knowledge tracking output before and after the perturbation is calculated to generate a learning path stability index.

[0071] The beneficial effects of the present application are:

[0072] Firstly, the present application proposes a hierarchical semantic heterogeneous graph network for student behavior modeling, which is used to depict the multi-level structure features of nodes, edges and subgraphs. Compared with the traditional graph model which only relies on single expression of nodes or edges, the representation ability of multi-source heterogeneous data in the education scene is significantly enhanced. By introducing type-aware encoding, semantic path attention fusion and structure level residual connection mechanism, the limitations of existing methods in structure heterogeneity expression are effectively solved, and the description accuracy of complex learning behavior of students is improved.

[0073] Secondly, the application designs a structure level contrast learning mechanism and a structure disturbance confrontation mechanism, constructs contrast samples for different structure levels, excavates micro semantic differences in structure changes, and introduces an adversarial disturbance consistency constraint to improve the robustness of the model to structure changes. Compared with the traditional optimization method which only takes the answer accuracy as the target, the application significantly enhances the generalization performance and representation consistency of the model in the dynamic evolution scene, and can effectively capture the continuity and mutability of the student knowledge state evolution over time.

[0074] In addition, the application can not only predict the correct rate of the student's answer, but also further quantify the high-order learning indicators such as the knowledge point mastery degree, learning transfer potential and path stability, realize the leap from surface behavior prediction to deep cognitive state evaluation, and improve the adaptive ability of the system to the individualized learning process.

[0075] To sum up, by introducing the hierarchical heterogeneous structure modeling, the structure contrast optimization mechanism and the multi-index knowledge tracking method, the application significantly improves the accuracy, flexibility and interpretability of the student knowledge state modeling, overcomes the defects of the prior art in structure staticity, expression singularity and insufficient prediction dimension, and has important application value and popularization significance. BRIEF DESCRIPTION OF DRAWINGS

[0076] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, which together with the embodiments of the application, is used to explain the application, and does not constitute a limitation on the application. In the drawings:

[0077] Fig. 1 is a schematic diagram of a knowledge tracking system based on heterogeneous graph contrast learning proposed by the application;

[0078] Fig. 2 is a student behavior subgraph construction flowchart in the application;

[0079] Fig. 3 is a hierarchical semantic heterogeneous graph network modeling schematic diagram in the application. DETAILED DESCRIPTION

[0080] The application will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, so they only show the components related to the application.

[0081] Reference Figs. 1-3 A knowledge tracking system based on heterogeneous graph contrast learning, comprising:

[0082] A data preprocessing module for collecting and preprocessing educational data to form standardized educational data;

[0083] a heterogeneous graph construction module configured to construct an educational semantic heterogeneous graph by a behavior semantic edge modeling mechanism;

[0084] a graph structure evolution module configured to divide the educational semantic heterogeneous graph into student behavior stages according to a time dimension, and construct a time evolution sequence graph;

[0085] a student behavior subgraph extraction module configured to extract a correlation structure neighborhood from the time evolution sequence graph, and generate a student behavior subgraph;

[0086] a hierarchical structure modeling module configured to construct a hierarchical semantic heterogeneous graph network, and model a structure representation of nodes, edges and subgraph structure levels of the student behavior subgraph;

[0087] a structure comparison optimization module configured to design a structure level comparison learning mechanism and a structure disturbance adversarial mechanism, and jointly optimize the hierarchical semantic heterogeneous graph network;

[0088] a knowledge tracking module configured to model and reason the optimized structure semantic features, and realize multi-dimensional prediction and evaluation of a student knowledge state.

[0089] In the embodiment, the modules are realized by the following methods:

[0090] S1, collecting educational data and pre-processing to form standardized educational data;

[0091] S2, constructing an educational semantic heterogeneous graph by a behavior semantic edge modeling mechanism based on the standardized educational data;

[0092] S3, dividing the educational semantic heterogeneous graph into student behavior stages by a time evolution modeling mechanism, and constructing a time evolution sequence graph;

[0093] S4, extracting a correlation structure neighborhood from the time evolution sequence graph according to a historical behavior trajectory of a target student, and generating a student behavior subgraph;

[0094] S5, modeling a structure representation of nodes, edges and subgraph structure levels of the student behavior subgraph by a hierarchical semantic heterogeneous graph network, and generating node structure features, edge level structure features and subgraph level structure features;

[0095] S6, designing a structure level comparison learning mechanism and a structure disturbance adversarial mechanism, and jointly optimizing parameters of the hierarchical semantic heterogeneous graph network;

[0096] S7, based on the optimized node structure features, edge level structure features and subgraph level structure features, performing multi-dimensional prediction and evaluation of a student knowledge state.

[0097] In the embodiment, the educational data includes student answer records, question texts, question attribute information, knowledge point labels, question type classification information and answer timestamps, and the preprocessing includes: removing outliers, completing missing items, unifying formats and normalizing identification codes of the educational data, and time sorting the student answer records based on the answer timestamps to form standardized educational data.

[0098] In the embodiment, the step S2 specifically includes:

[0099] S21, based on the standardized educational data, extracting student entities, question entities, knowledge point entities and question type entities through field mapping and attribute label analysis methods, constructing a node set, and the node set includes a student node set, a question node set, a knowledge point node set and a question type node set;

[0100] S22, constructing a semantic edge set through a behavior semantic edge modeling mechanism, the behavior semantic edge modeling mechanism combines behavior categories and structural semantics, generates an edge set containing multiple types of semantic relationships based on student answering behaviors, question and knowledge point mapping relationships, question and question type attribution relationships and question dependency relationships, and the semantic edge set includes answering behavior edges, knowledge point association edges, question type attribution edges and question dependency edges;

[0101] S23, constructing an educational semantic heterogeneous graph from the node set and the semantic edge set;

[0102] S24, setting an initial attribute vector for each type of node, wherein the student node is set as a student attribute vector, the question node is set as a question information vector, the knowledge point node is set as a knowledge point feature vector, and the question type node is set as a question type category vector, and the attribute vectors of various types of nodes are collectively constructed into an initial node feature set;

[0103] S25, recording the semantic type and connection relationship of each edge.

[0104] For example, student A answered question Q1 at timestamp 2024-03-15 14:05:23, the stem of question Q1 is “Solve the equation 2x + 3 = 7”, the type of question Q1 is multiple choice, and it is associated with two knowledge points: solving linear equations and basic properties of equations. It is also recorded that Q1 has a progressive relationship with another question Q2, which is a transformed solving type of advanced questions. After data preprocessing, five types of structured entity nodes are identified from the answer record through field mapping and label parsing methods: student A, question Q1, knowledge point KP1, knowledge point KP2, and question type TQ1, which constitute the multi-type node set in the educational semantic heterogeneous graph. Then, student A and the answered question Q1 are connected by an answer behavior edge; Q1 and KP1, KP2 are connected by knowledge point association edges; Q1 and TQ1 are connected by question type attribution edges; and Q1 and Q2 are connected by question dependence edges. In the initial node attribute setting stage, the attribute vector of student node A is constructed by the past answer accuracy, learning time, and mastery score; the attribute vector of question Q1 is generated by encoding the stem text of the question through a language model and integrating the question difficulty and cognitive dimension information; the feature vectors of knowledge points KP1 and KP2 come from the knowledge attributes and sequential relationships recorded in the knowledge graph; and the attribute vector of question type TQ1 is generated according to the question type classification and teaching function label.

[0105] In this embodiment, the time evolution modeling mechanism divides the educational semantic heterogeneous graph into multiple stage graphs according to the answer time and constructs a sequence of graph structures in time order, and the step S3 specifically includes:

[0106] S31, according to the answer timestamp in the standardized education data, the answer behavior of each student is divided into several continuous time stages in time order, forming a set of time stages;

[0107] S32, for each time stage t i extract the active nodes and associated semantic edges in the current stage from the educational semantic heterogeneous graph to construct the corresponding time slice graph;

[0108] S33, arrange the time slice graphs in time order to construct a sequence of time evolution graphs;

[0109] S34, construct a structure evolution pair between any two consecutive time slice graphs, which establishes the time sequence correspondence between the graph structures by comparing the node state changes, edge semantic type adjustments, and local connection structure differences between adjacent time slice graphs.

[0110] For example, student A answers question Q1 at 14:05:23 on March 15, 2024, and then answers question Q2 at 14:19:12 on March 15, 2024. According to the time of answering, the answering behavior is time-ordered, and a time window division strategy is set to divide the learning process into two time stages T1 and T2, and the time stage set T is equal to T1 and T2.

[0111] In stage T1, the system extracts the active nodes and edges related to question Q1 from the educational semantic heterogeneous graph, constructs a time slice graph GT1, the node set contains student A, question Q1, knowledge points one equation solving and equation basic properties, and question type multiple choice, and the edge set includes the answering behavior edge of the student to the question, the association edge between the question and the two knowledge points, and the belonging edge between the question and the question type.

[0112] In stage T2, the system identifies that student A answers question Q2. The question is a transformed solving type question, containing knowledge points equation basic properties and one-dimensional transformation, and the question type is fill-in-the-blank, and there is a teaching dependency relationship with Q1. Therefore, the system constructs a time slice graph GT2, which contains student A, question Q2, two knowledge points, question type fill-in-the-blank, and the edge set includes the answering edge of the student to the question, the association edge between the question and the knowledge point, the belonging edge between the question and the question type, and the question dependency edge from Q1 to Q2.

[0113] The two time slice graphs are arranged in time order to form a time evolution graph sequence G.

[0114] In this embodiment, the step S4 includes generating a student behavior subgraph based on the behavior trajectory of the target student, specifically including the following steps:

[0115] S41, according to the unique identification information of the target student, extracting the historical answering behavior sequence of the target student from the standardized education data, and combining the time evolution graph sequence, screening out a plurality of time slice graphs matched with the student answering time period;

[0116] S42, set the neighborhood radius r, take the student node corresponding to the target student as the center, and limit the structure depth and edge type in the node selection process, select the r-hop semantic neighbor nodes connected with the target student from the time slice graph, construct a semantic neighbor node set, and extract the corresponding semantic edge set, generate a local structure subgraph of the target student;

[0117] S43, for each candidate node in the local structure subgraph, define a meta-path set, the meta-path is composed of a pre-defined node type sequence and an edge type sequence;

[0118] S44, based on the meta-path, generate the embedding vector of the candidate node and the student node in the semantic of the meta-path through the graph embedding method, and obtain the structural correlation degree of the two through the normalized dot product similarity;

[0119] S45, weight and fuse the structural correlation degrees on all paths to obtain a weighted structural correlation degree;

[0120] S46, set an association strength threshold, filter the candidate nodes and the semantic edges corresponding to the candidate nodes according to the weighted structural correlation degree value, remove the nodes and edges with a structural correlation degree lower than the threshold, and only keep the nodes and edges with a weighted structural correlation degree greater than or equal to the association strength threshold, and generate a stage filtering graph;

[0121] S47, sequentially combine all the stage filtering graphs in time order to form a behavior structure subgraph for the target student.

[0122] For example, the historical answer record of the target student stu_001 in a certain online learning system covers questions q_001 to q_015, and each question records the answering time, the knowledge point it belongs to, the question type and the answer result. The answering time of the target student stu_001 is divided into multiple continuous stages in time order, and the corresponding answer behavior is extracted according to the time window in each stage, for example, the first stage contains the answer records of questions q_003, q_004 and q_007.

[0123] In the corresponding time slice graph, taking the target student node stu_001 as the center, setting the neighborhood radius to 2, and combining the semantic edge type and the structural depth limit, the semantic neighbor nodes with a distance of not more than 2 hops from the student node in the graph structure are screened out, and the obtained semantic neighbors include question nodes q_003, q_004, q_007, which are respectively associated with knowledge point nodes k_01, k_02, k_03 and question type nodes t_01, t_02. Since question node q_006 has an association relationship with knowledge point k_01, question node q_006 is also included in the local structure subgraph of the target student.

[0124] For the candidate nodes in the above local structure, multiple meta-path semantics are constructed, such as student-question-knowledge point, student-question-question type and student-question-knowledge point-question. Under each meta-path semantics, the structural correlation between the target student node and the candidate node is calculated respectively. The structural correlation threshold is set to 0.65. The structural correlation of the question node q_004 and the student node stu_001 under the student-question-knowledge point path is higher, while the structural correlation of the question node q_006 under each path is lower than the threshold. Finally, the question nodes q_003 and q_004, the knowledge point nodes k_01 and k_02, the question type node t_01 and the corresponding semantic edges with strong structural correlation are retained to form a phased filtering graph. After merging all the filtering graphs of different time stages in chronological order, the behavior structure subgraph of the student is obtained.

[0125] In the embodiment, the step S5 comprises:

[0126] S51, a hierarchical semantic heterogeneous graph network is constructed, and the student behavior subgraph is modeled at the node level, the edge level and the subgraph level. The hierarchical semantic heterogeneous graph network comprises a node-level semantic fusion module, an edge-level semantic aggregation module and a subgraph-level structure fusion module.

[0127] S52, the student behavior subgraph is input into the node-level semantic fusion module, the features of each type of node in the student behavior subgraph are extracted and the multi-semantic information is fused, and the node structure features are generated.

[0128] S521, the initial attribute vector of each node in the student behavior subgraph set in the educational semantic heterogeneous graph is inherited as the input feature;

[0129] S522, a type mapping function is set to encode the node features in a type-specific manner to obtain typed features;

[0130] S523, according to the context of the multiple meta-paths involved by the node in the student behavior subgraph, the structural context features are calculated according to the path attention mechanism;

[0131] S524, the global behavior statistical features of the target student are introduced to supplement the role information;

[0132] S525, the typed features, the structural context features and the global behavior statistical features are fused to generate the node structure features;

[0133] S53, the student behavior subgraph is input into the edge-level semantic aggregation module, the semantic edges are modeled, and the edge-level structure features are generated;

[0134] S531, a type-aware mechanism is introduced, which sets a semantic gating function for different types of semantic edges in the student behavior subgraph, calculates the edge-level information transmission weight, and dynamically regulates the information transmission strength of each edge according to the structural feature difference between the starting node and the target node of the edge connection, and combines the edge semantic type;

[0135] S532, the multi-type neighbor node information of the target node is weighted and aggregated by the edge-level information transmission weight to generate edge-level structural features;

[0136] S54, the node structure feature vector and the edge-level structure feature vector are fused by a subgraph-level structure fusion module to construct a subgraph-level structure feature:

[0137] S541, the node structure feature vector and the edge-level structure feature vector are fused by a residual connection method to generate a multi-layer substructure feature;

[0138] S542, a path semantic learning function is designed to model the multi-layer substructure feature, assign adaptive weights based on path types, and combine the multi-layer substructure feature to generate a subgraph-level structure feature.

[0139] In this embodiment, the step S6 specifically includes:

[0140] S61, a structure-level contrast learning mechanism is constructed to establish a contrast learning process for the student behavior subgraph at the node level, edge level and subgraph level;

[0141] The node level is based on the time evolution graph sequence to construct a structure triple (v + ,v - ), wherein the positive sample node v + represents a corresponding node with similar types, connection relationships and local semantic structures in adjacent time slice graphs, and v - represents a heterogeneous node with obvious changes in structural topology or semantic features;

[0142] The edge level constructs an edge-level structure contrast triple (e + ,e - ) based on semantic edges, wherein the positive sample edge e + represents a semantic edge connecting the same node pair with unchanged edge type, and the negative sample edge e - represents an edge with consistent node pairs but changed edge type, direction or topological position;

[0143] The subgraph level constructs a subgraph contrast sample triple wherein the positive sample subgraph Adjacent temporal sub-graphs representing structural topology and semantic edge stability; negative sample sub-graphs Phasic sub-graphs representing significant changes in structural activity or reconfiguration of connection patterns

[0144] A structural hierarchy contrast loss function is formed by maximizing structural representation similarity between positive samples and minimizing structural representation similarity between negative samples

[0145] S62, based on the student behavior sub-graph, a structural disturbance adversarial mechanism is introduced, and by slightly disturbing the node connection structure, semantic edge type and sub-graph topology, a structural disturbance sub-graph is obtained

[0146] S63, the mean square error is used to model the adversarial consistency of the student behavior sub-graph and the structural disturbance sub-graph, and an adversarial disturbance consistency loss term is constructed

[0147] S64, the structural hierarchy contrast loss and the adversarial disturbance consistency loss are jointly modeled to form an optimization objective function, and the node-level semantic fusion module parameters, edge-level semantic aggregation module parameters and sub-graph-level structure fusion module parameters in the hierarchical semantic heterogeneous graph network are jointly optimized to obtain the optimized

[0148] In this embodiment, the step S7 specifically comprises:

[0149] S71, the optimized node-level structure features, edge-level structure features and sub-graph-level structure features are fused and modeled to extract the structural semantic vector of the target student

[0150] S72, based on the structural semantic vector, the behavior performance of the target student in future learning activities is comprehensively modeled to obtain the answer correctness probability, knowledge point mastery vector, learning transfer potential score and learning path stability index

[0151] S73, the structural semantic vector and the node features of the to-be-predicted question are matched in similarity to predict the answer correctness probability of the student to the question title

[0152] S74, based on the relationship between the structural semantic vector and the knowledge point node, a multi-label classifier is used to predict the mastery degree of the student to multiple knowledge points involved in the current stage, and a knowledge point mastery score set is output

[0153] S75, the transfer deviation between the historical stage structural semantic vector and the current fusion representation is modeled, the ability of the student to transfer existing knowledge to new question types or new knowledge points in the learning process is quantified, and the learning transfer potential score is output

[0154] S76, according to the structural disturbance sub-graph, the consistency score of the knowledge tracking output before and after disturbance is calculated, and the learning path stability index is generated

[0155] Example 1

[0156] To verify the actual application effect of the present application, the present application is applied to a smart teaching platform of a key middle school in a province, and online answering data of students in a real teaching process is collected and analyzed on the platform. The experimental data covers all online answering behaviors of 316 students in five classes of ninth grade of the school in the learning process of "algebra and function" unit, including student answering records, question content, knowledge points, question type classification, answering time and correctness information, and accumulatively forms 22,578 pieces of effective answering behavior data, 342 questions in total, 25 knowledge points involved, and three types of question type distribution including selection questions, fill-in-the-blank questions and calculation questions.

[0157] In this scenario, first, the student answering behavior data is preprocessed, and invalid records are removed and fields are uniformly coded. Second, an educational semantic heterogeneous graph is constructed, students, questions, knowledge points and question types are modeled as different types of nodes, answering behaviors, knowledge associations, question type attributions and question associations are modeled as different types of edges, forming a graph structure connected by multiple semantics. Based on the answering time information, the learning process of the students is divided into five stages, a time evolution graph sequence is constructed, and the behavior subgraph of each student in each stage is extracted. Then, the student behavior subgraph is represented at the node level, edge level and subgraph level by a hierarchical semantic heterogeneous graph network, and the parameters of the graph neural network model are optimized by a structure level comparison mechanism and a structure disturbance adversarial mechanism, so that the model has stronger generalization ability and robustness in capturing structural heterogeneity and dynamic evolution characteristics. On the basis of the fused structure semantic representation, four knowledge tracking indicators are constructed respectively: student answering accuracy prediction, knowledge point mastery score, learning transfer potential score and learning path stability index.

[0158] To evaluate the effectiveness of the present application, four kinds of current mainstream knowledge tracking models are selected as comparative baseline schemes, which are: deep knowledge tracking model (DKT), attention knowledge tracking model (AKT), graph neural network knowledge tracking model (GKT) and heterogeneous graph knowledge tracking model (HGKT). The performance is evaluated in four dimensions of answering accuracy prediction, knowledge point mastery score, learning transfer consistency rate and learning path stability index, and the specific results are shown in Table 1.

[0159] Table 1 Performance comparison of the present application system and existing models in multi-dimensional knowledge tracking tasks

[0160]

[0161] As can be seen from Table 1, in terms of the answer correctness rate prediction, the answer correctness rate of the system of the present application is 0.854, which is obviously better than 0.784 of the deep knowledge tracking method and 0.823 of the heterogeneous graph knowledge tracking method, embodying the significant advantage of the fusion of heterogeneous structure information in the prediction accuracy. In terms of the knowledge point mastery degree, the present application introduces a modeling method combining the structural semantics and the multi-label classification mechanism, and the knowledge point mastery score is 0.743, which is more than 5% higher than that of other methods, proving the high accuracy and robustness of the present application in the multi-knowledge point state recognition. In the learning transfer potential modeling, by modeling the difference between the historical semantic structure and the current representation, the learning transfer consistency rate of the present application is 0.689, which is significantly better than other graph models, indicating that the present application is better at depicting the cross-structure knowledge transfer ability of students. In terms of the learning path stability evaluation, the present application uses the structural disturbance subgraph to simulate the graph structure variation, measures the response stability of the model to the micrograph structure change, and the learning path stability index is 0.627, embodying the advantage of the present application in capturing the structural cognitive stability of students.

[0162] In summary, the present application has obvious advantages in predicting the answer correctness rate, modeling the knowledge transfer, structural perception and learning path evaluation complex tasks, and fully verifies the wide prospect and practical value of the present application in the intelligent education application.

[0163] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A knowledge tracking system based on heterogeneous graph contrastive learning, characterized in that, include: The data preprocessing module is used to collect educational data and preprocess it to form standardized educational data. The heterogeneous graph construction module is used to construct educational semantic heterogeneous graphs through a behavioral semantic edge modeling mechanism. The graph structure evolution module is used to divide the educational semantic heterogeneous graph into student behavior stages according to the time dimension and construct a time evolution sequence graph. The student behavior subgraph extraction module is used to extract the associated structure neighborhood from the time evolution sequence graph and generate the student behavior subgraph. The hierarchical structure modeling module is used to construct a hierarchical semantic heterogeneous graph network and to perform structural representation modeling of student behavior subgraph nodes, edges and subgraph structure hierarchy; The structural contrast optimization module is used to design a structural hierarchical contrast learning mechanism and a structural perturbation adversarial mechanism to jointly optimize the hierarchical semantic heterogeneous graph network. The knowledge tracking module is used to model and reason about the optimized structural semantic features, enabling multidimensional prediction and evaluation of students' knowledge status.

2. The knowledge tracking system based on heterogeneous graph contrastive learning according to claim 1, characterized in that, The modules are connected in the following way: S1. Collect educational data and preprocess it to form standardized educational data; S2. Based on standardized educational data, construct an educational semantic heterogeneous graph through a behavioral semantic edge modeling mechanism; S3. Using a time evolution modeling mechanism, the educational semantic heterogeneous graph is divided into student behavior stages to construct a time evolution sequence graph; S4. Based on the historical behavioral trajectory of the target student, extract the associated structural neighborhood from the time evolution sequence graph to generate a student behavior subgraph. S5. Using a hierarchical semantic heterogeneous graph network, the student behavior subgraph is structurally represented and modeled at the node, edge and subgraph structure levels, generating node structural features, edge-level structural features and subgraph-level structural features. S6. Design a hierarchical comparative learning mechanism and a structural perturbation adversarial mechanism to jointly optimize the parameters of hierarchical semantic heterogeneous graph network; S7. Based on the optimized node structure features, edge structure features, and subgraph structure features, students' knowledge status is predicted and evaluated in multiple dimensions.

3. The knowledge tracking system based on heterogeneous graph contrastive learning according to claim 2, characterized in that, The educational data includes student answer records, question text, question attribute information, knowledge point tags, question type classification information, and answer timestamps. The preprocessing includes: removing outliers from the educational data, filling in missing items, standardizing the format and normalizing the identifier encoding, and sorting the student answer records by time based on the answer timestamps to form standardized educational data.

4. The knowledge tracking system based on heterogeneous graph contrastive learning according to claim 2, characterized in that, Step S2 specifically includes: S21. Based on standardized educational data, extract student entities, question entities, knowledge point entities, and question type entities through field mapping and attribute label parsing methods, and construct a node set, which includes a student node set, a question node set, a knowledge point node set, and a question type node set. S22. Construct a semantic edge set through a behavioral semantic edge modeling mechanism. The behavioral semantic edge modeling mechanism combines behavioral categories and structural semantics, and generates an edge set containing multiple semantic relationships based on student answering behavior, the mapping relationship between questions and knowledge points, the relationship between questions and question types, and the dependency relationship between questions. The semantic edge set includes answering behavior edges, knowledge point association edges, question type association edges, and question dependency edges. S23. Construct an educational semantic heterogeneous graph from the set of nodes and the set of semantic edges; S24. Set an initial attribute vector for each type of node, where student nodes are set as student attribute vectors, question nodes are set as question information vectors, knowledge point nodes are set as knowledge point feature vectors, and question type nodes are set as question type category vectors. Combine the attribute vectors of each type of node to form an initial node feature set. S25. Record the semantic type and connection relationship of each edge.

5. The knowledge tracking system based on heterogeneous graph contrastive learning according to claim 2, characterized in that, The aforementioned temporal evolution modeling mechanism divides the educational semantic heterogeneous graph into multiple stage graphs according to the response time, and constructs a graph structure sequence arranged in chronological order. Step S3 specifically includes: S31. Based on the answer timestamps in the standardized educational data, divide each student's answering behavior into several consecutive time stages in chronological order to form a set of time stages; S32. For each time stage, extract the active nodes and related semantic edges in the current stage from the educational semantic heterogeneous graph, and construct the corresponding time slice graph. S33. Arrange the time slices in chronological order to construct a time evolution diagram sequence; S34. Construct structural evolution pairs between any consecutive time slice graphs. The structural evolution pairs establish temporal correspondences between graph structures by comparing the changes in node states, the adjustment of edge semantic types, and the differences in local connection structures between adjacent time slice graphs.

6. The knowledge tracking system based on heterogeneous graph contrastive learning according to claim 2, characterized in that, Step S4 includes generating a student behavior sub-graph based on the target student's behavior trajectory, specifically including the following steps: S41. Based on the unique identification information of the target students, extract the historical answer behavior sequence of the target students from the standardized education data, and combine it with the time evolution graph sequence to select multiple time slices that match the time period of the students' answering questions. S42. Set the neighborhood radius r, take the student node corresponding to the target student as the center, and combine the structural depth and edge type to limit the node selection process. Select the r-hop semantic neighbor nodes connected to the target student from the time slice graph, construct the semantic neighbor node set, extract the corresponding semantic edge set, and generate the local structure subgraph of the target student. S43. For each candidate node in the local structure subgraph, define a metapath set, wherein the metapath is composed of a predefined sequence of node types and a sequence of edge types. S44. Based on metapath, the embedding vectors of candidate nodes and student nodes under metapath semantics are generated by graph embedding method, and the structural correlation between the two is obtained by normalized dot product similarity. S45. Weighted fusion of structural correlations on all paths to obtain weighted structural correlations; S46. Set the association strength threshold, filter the semantic edges of candidate nodes and candidate nodes according to the weighted structural association value, remove nodes and edges whose structural association with the target student node is lower than the threshold, retain only nodes and edges whose weighted structural association is greater than or equal to the association strength threshold, and generate a staged filtering graph. S47. Merge all the stage-by-stage filter diagrams in chronological order to form a sub-diagram of the behavioral structure for the target students.

7. The knowledge tracking system based on heterogeneous graph contrastive learning according to claim 2, characterized in that, Step S5 includes: S51. Construct a hierarchical semantic heterogeneous graph network, and perform structural representation modeling of student behavior subgraphs at the node level, edge level and subgraph level respectively. The hierarchical semantic heterogeneous graph network includes a node-level semantic fusion module, an edge-level semantic aggregation module and a subgraph-level structural fusion module. S52. Input the student behavior subgraph into the node-level semantic fusion module to extract features and fuse multiple semantic information for each type of node in the student behavior subgraph. By inheriting the initial attribute vectors set in the educational semantic heterogeneous graph for each node in the student behavior subgraph as input features, and setting a type mapping function, the node features are encoded in a type-specific manner to obtain typified features. Based on the multiple meta-path contexts involved in the student behavior subgraph, structural context features are calculated using the path attention mechanism, and global behavioral statistical features of the target student are introduced to supplement role information. The node structure features are generated by fusing typified features, structural context features, and global behavioral statistical features. S53. Input the student behavior subgraph into the edge-level semantic aggregation module to model the semantic edges; An edge type awareness mechanism is introduced. For different types of semantic edges in the student behavior subgraph, an edge semantic gating function is set to calculate the edge-level information transmission weight. The edge type awareness mechanism dynamically adjusts the information transmission intensity of each edge based on the structural feature differences between the starting node and the target node connected by the edge and in combination with the edge semantic type. By weighting the neighbor node information of the target node through edge-level information transmission, the edge-level structure features are generated by weighting and aggregating the multi-type neighbor node information of the target node. S54. The node structure feature vector and the edge structure feature vector are fused at the structural level through the subgraph-level structure fusion module. By fusing the node structure feature vector and the edge structure feature vector through residual connection, multi-layer substructure features are generated. Design a path semantic learning function to perform semantic modeling on multi-level substructure features, assign adaptive weights based on path type, and weightedly combine multi-level substructure features to generate subgraph-level structural features.

8. The knowledge tracking system based on heterogeneous graph contrastive learning according to claim 2, characterized in that, Step S6 specifically includes: S61. Construct a hierarchical comparative learning mechanism, and establish comparative learning processes for student behavior subgraphs at the node level, edge level and subgraph level respectively. The node level constructs positive and negative sample nodes for each target node based on the time evolution graph sequence. The edge level constructs positive and negative semantic edges by building an edge-level structure based on semantic edges. The subgraph level constructs positive and negative subgraphs by building positive and negative subgraphs based on local structural subgraphs. By maximizing the structural representation similarity between positive samples and minimizing the structural representation similarity between negative samples, a structural hierarchy contrast loss function is formed. S62. Based on the student behavior subgraph, a structural perturbation adversarial mechanism is introduced. By performing small perturbations on the node connection structure, semantic edge type and subgraph topology, a structural perturbation subgraph is obtained. S63. Use mean squared error to model the adversarial consistency of the student behavior subgraph and the structural perturbation subgraph, and construct an adversarial perturbation consistency loss term; S64. Jointly model the structural hierarchy contrast loss and the adversarial perturbation consistency loss to form an optimization objective function. Jointly optimize the node-level semantic fusion module parameters, edge-level semantic aggregation module parameters, and subgraph-level structural fusion module parameters in the hierarchical semantic heterogeneous graph network to obtain the optimized result.

9. The knowledge tracking system based on heterogeneous graph contrastive learning according to claim 2, characterized in that, Step S7 specifically includes: S71. The optimized node-level structural features, edge-level structural features and subgraph-level structural features are fused and modeled to extract the structural semantic vector of the target student. S72. Based on structural semantic vectors, comprehensively model the behavioral performance of target students in future learning activities to obtain the probability of correct answers, knowledge point mastery vector, learning transfer potential score, and learning path stability index. S73. By performing similarity matching between the structural semantic vector and the node features of the question to be predicted, the probability of the student answering the target question correctly is predicted. S74. Based on the relationship between structural semantic vectors and knowledge point nodes, a multi-label classifier is used to predict the students' mastery of multiple knowledge points involved in the current stage, and output a set of knowledge point mastery scores. S75. By comparing the transfer bias model between the historical stage structural semantic vector and the current fusion representation, students' ability to transfer existing knowledge to new question types or new knowledge points during the learning process is quantified, and a learning transfer potential score is output. S76. Based on the structural perturbation subgraph, calculate the consistency score of knowledge tracking output before and after the perturbation, and generate the learning path stability index.