A student behavior simulation method based on dynamic graph retrieval enhanced generation

Abstract

The application discloses a student behavior simulation method based on dynamic graph retrieval enhanced generation, comprising the following steps: 1) inputting education information, extracting problems, knowledge points and thinking chain to construct dynamic knowledge element units, and combining transfer support to dynamically calibrate edge weights to construct a dynamic knowledge graph; 2) inputting student interaction sequences, using graph propagation aggregation and time decay mechanism to calculate and update the potential knowledge mastery state of students in real time; 3) based on the updated state, Markov walk is performed in the graph to sample candidate questions, predict response pairs, recall thinking chains or interference items, and guide the large language model to generate learning behavior sequences; 4) a multi-objective hybrid loss function including main prediction loss and contrast learning regularization loss is constructed, and model parameters are optimized in reverse propagation. The application effectively eliminates the logical illusion of the large language model generation, accurately captures the dynamic cognitive evolution, and realizes the high-fidelity student behavior simulation.

Description

Technical Field

[0001] This patent application belongs to the field of smart education and educational data mining technology, specifically a student behavior simulation method based on dynamic graph retrieval enhancement generation. Background Technology

[0002] Student behavior simulation and dynamic knowledge tracking (KT) are core technologies in Intelligent Tutoring Systems (ITS), aiming to dynamically assess the evolution of students' cognitive states and simulate and plan future learning behaviors by analyzing students' historical interaction data. However, existing technologies generally suffer from problems such as sparse offline data, insufficient capture of dynamic evolution, and "logical illusions" in generative large models, leading to inaccurate predictions and unreasonable generation in complex and ever-changing real-world teaching scenarios. These technological limitations are particularly pronounced in specific educational applications: In basic education, students' learning trajectories are often fragmented and disjointed, and models, unable to handle dynamic dependencies across knowledge points, are prone to recommending questions with unbalanced difficulty, leading to student frustration; in vocational skills and industrial training, skills learning has extremely strict prerequisite constraints (such as mastering theory and safety regulations before practical application), and existing large language models are prone to generating "reverse learning sequences" that violate these prerequisite relationships, thus causing serious compliance and security risks; in massive online learning platforms (MOOCs), due to the lack of real-time guidance from live teachers, the static graph rules relied upon by traditional models are difficult to adapt to learners' highly personalized and dynamically changing cognitive evolution, resulting in rigid recommendation paths; and in personalized and precise tutoring scenarios, because existing deep models lack fine-grained cognitive process modeling and interpretability, teachers find it difficult to accurately locate students' fundamental cognitive deficiencies (i.e., "not knowing why they are wrong") based on the system's "black box" output, thus hindering effective teaching intervention.

[0003] To alleviate the above problems, existing research mainly focuses on three aspects: (1) Sequence and memory enhancement methods, which use recurrent neural networks or memory matrices to encode historical interaction sequences to predict future performance, such as DKT, DKVMN, etc.; (2) Graph structure learning methods, which use knowledge graphs or bipartite graphs to explicitly capture the dependencies between knowledge points and questions, such as GKT, GIKT, etc.; (3) Generative simulation and retrieval enhancement methods based on large language models (LLMs), which use agents or external text libraries to generate simulated interaction data to alleviate data sparsity, such as Agent4Edu, traditional RAG-EDU, etc.

[0004] While the aforementioned methods alleviate the problems of data sparsity and sequence modeling to some extent, they still have many shortcomings: sequence models ignore complex topological dependencies; graph structure models rely heavily on static graphs and lack dynamic update mechanisms that evolve with interactions; and existing retrieval augmented generation (RAG) often focuses on surface-level "text similarity" rather than deep "cognitive dependencies." These limitations prevent existing models from completely eliminating "logical illusions" in the generation process, and their cognitive tracking fidelity, logical consistency, and adaptive generalization ability remain very limited in real-world complex educational scenarios. Summary of the Invention

[0005] The purpose of this invention is to address the technical problems in existing technologies, such as logical illusions in the generation of student interaction data by large language models, the inability of traditional knowledge tracing models to adapt to the dynamic evolution of cognitive states due to their reliance on static structures, and the resulting inaccuracies in the prediction and generation of learning behavior sequences. This invention proposes a student simulation method based on dynamic graph retrieval enhancement and cognitive state tracking. This method constructs cognitive constraints with clear semantic boundaries through a dynamic knowledge graph module and introduces a dynamic knowledge tracing module and a pattern-adaptive retrieval enhancement generation mechanism. This effectively eliminates factual and logical errors in the generation process of large models, improving the diagnostic accuracy and simulation data fidelity of the model in sparse data scenarios.

[0006] The technical solution to achieve the objective of this invention is:

[0007] A student behavior simulation method based on dynamic graph retrieval enhancement includes the following steps:

[0008] 1) Input educational information and construct a dynamic knowledge graph. Unstructured educational texts are used as input. Problem texts, related knowledge points and thought chains are extracted from the educational texts through text analysis. A dynamic knowledge graph module is constructed to obtain fine-grained dynamic knowledge elements with clear cognitive boundaries. The initial edges of the knowledge graph are established by using the pre-dependency relationship of knowledge points and the co-occurrence similarity of entities. The support feature of node transfer is introduced through frequent itemset mining. After the multi-dimensional features are fused, the relationship weights between nodes are dynamically calibrated, thereby constructing a dynamic knowledge graph that reflects the learning patterns of the student group.

[0009] 2) Input the target student's historical interaction sequence. After constructing the dynamic knowledge graph in step 1), use the sparse activation mechanism to extract the current question interaction node and the knowledge graph neighborhood subgraph. Aggregate the influence of the previous node and similar nodes through the graph propagation mechanism to generate context-aware cognitive proficiency features. When a new real interaction response is received, combine the time interval-based decay mechanism with the current proficiency features to calculate and update the student's potential knowledge mastery status in real time.

[0010] 3) Utilizing the knowledge mastery status updated in real time in step 2), perform Markov decision walk computation on the dynamic knowledge graph; jointly evaluate the graph topology connectivity, historical transition patterns, and revisit bias, adaptively calculate and transfer probabilities to each candidate node to sample questions; for the sampled questions, fuse state features at different levels through a gating mechanism to predict the correctness of students' responses, and use the correctness of these responses as anchor points to recall standard thought chains or error-prone interference features in the graph, guiding the large language model to generate learning behavior sequences containing correct reasoning logic or specific erroneous reasoning paths;

[0011] 4) Construct a multi-objective hybrid loss function that includes the main prediction loss of binary cross-entropy and the contrastive learning regularization loss; in different training batches, adopt the training paradigm of machine learning, aiming to optimize the potential knowledge mastery state of students in step 2) and the transition probability of candidate nodes in step 3). By jointly optimizing the hybrid loss function as an intermediary, calculate the feature distance between positive and negative samples, map the potential state of students to the feature region that matches the actual mastery ability of students, backpropagate to optimize the model parameters, and finally train an optimized model that can output a high-fidelity student learning behavior sequence, which is used to generate high-quality student data and the learning state change curve of each student under different datasets.

[0012] The construction of the dynamic graph module in step 1) includes: the construction unit of dynamic knowledge element unit, namely DKEU, and the dynamic calibration unit of edge weight;

[0013] The dynamic knowledge element unit refers to the extraction of question skills and difficulty information from the text through the large language model (Qwen3), deconstructing traditional single knowledge nodes into units containing questions. Related knowledge point set Frequent items and mind chain The quadruple representation is denoted as In order to optimize the reasoning process of question parsing obtained from large language models Generate different problem-solving branches The logical rigor of each branch is scored independently using a large language model as an evaluator. And calculate the target node based on the Softmax merging operation. The optimal fusion feature is generated based on the overall score.

[0014]

[0015] The edge weight dynamic calibration unit first establishes directed edges based on the pre-dependencies of knowledge points. Undirected edges are established based on semantic similarity calculated from entity co-occurrence, and similarity scores between nodes are calculated. Subsequently, frequent itemset mining is introduced to calculate the support score of node transitions in the real historical sequence. The edge weights are dynamically updated by weighting and fusing the node transition pattern tendency score after dynamic calibration of frequent items with the inherent semantic similarity. :

[0016] ;

[0017] The edge weight dynamic calibration unit dynamically updates the dynamic knowledge graph constructed in step 1) and is used for updating the potential knowledge mastery status of students in the subsequent step 3).

[0018] After obtaining dynamic knowledge graph information, the dynamic knowledge tracking module is used to track and update students' potential knowledge mastery status.

[0019] The dynamic knowledge tracking module includes a graph propagation aggregation unit, a time decay update unit, and a state-aware response simulation unit.

[0020] The graph propagation aggregation unit design employs a sparse activation mechanism, performing operations only on the current node and its k-hop neighborhood subgraph to reduce the computational overhead of the entire graph, utilizing attention weights. and Aggregate from the preceding node respectively and similar nodes The state influence, after layer normalization, generates a smooth context proficiency representation. :

[0021]

[0022] The time decay update unit design: After receiving the student's actual interactive response, a time interval-based update unit is introduced. Exponential decay factor To simulate the forgetting curve of human memory, combined with the learning rate Response correctness and similarity calculation of state increment And through explicit numerical truncation functions Force the updated state to be bound to Within the probability interval, the update formula is:

[0023] ;

[0024] After assessing students' mastery of different knowledge points through the dynamic knowledge tracking module, in order to simulate student behavior, the retrieval enhancement generation module uses an adaptive method to match the exercises that need to be completed for different students.

[0025] The retrieval enhancement generation module includes a pattern-adaptive problem sampling unit, a state-aware response simulation unit, and an enhanced content generation unit;

[0026] The mode-adaptive problem sampling unit transforms the retrieval strategy into a Markov decision process and calculates the comprehensive transition score. The score dynamically incorporates logical scores. Pattern Score and historical revisit bias weight Control indication function :

[0027]

[0028] Subsequently, through the temperature parameter The Softmax function converts the overall transition score into transition probability. The candidate questions are determined by comprehensively assessing the students' current knowledge level, calculating the probability of each type of question being selected as the next question, and selecting the most suitable candidate question based on this probability.

[0029] The state-aware response simulation unit is used to determine the virtual student's performance. This unit introduces a hierarchical perception tracking mechanism and gating mechanism, dividing the knowledge information output by the dynamic knowledge tracking module into basic hierarchical information. With high-level information When dealing with basic-level information, assign higher fixed weights to this basic-level information. This anchors and defines the initial cognitive boundaries of virtual students; for higher-level information, an adaptive weight is calculated based on the text content of the higher-level information and related skill features. This allows for the dynamic capture and modification of the student's current state. Basic and high-level information are aggregated by a gating mechanism to generate a fused state vector. Subsequently, the fused state vector The input is fed into a multilayer perceptron, and after nonlinear mapping, the final output is the accuracy of the answer. Based on the accuracy rate of the answers Determine whether the virtual student's response to the selected candidate question is right or wrong;

[0030] The enhanced content generation unit: using the selected question as the anchor point, it recalls standard thought chains and similar error-prone interference items from the vector index of the graph. If the simulated response is correct, the large language model generates the correct reasoning logic based on the contextual standard thought chain; if the simulated response is incorrect, it generates a specific erroneous reasoning path based on the characteristics of the recalled interference items.

[0031] This technical solution constructs a dynamic knowledge graph and uses state tracking to guide retrieval and generation. Compared with existing student simulation and knowledge tracking methods, it can effectively eliminate the logical illusion caused by the lack of cognitive constraints when large language models generate learning sequences, improve the model's ability to capture and predict the dynamic learning evolution process of students, and significantly improve the diagnostic accuracy and fidelity of generated data in sparse data (cold start) scenarios through multi-objective joint optimization. Attached Figure Description

[0032] Figure 1 This is a conceptual comparison diagram between the implementation method and Large Model-Assisted Education (LLM-EDU) and Retrieval Enhancement Education (RAG-EDU);

[0033] Figure 2 This is a framework diagram of an embodiment;

[0034] Figure 3 This is a flowchart of the execution process in the embodiment;

[0035] Figure 4 The knowledge graph generated by the method in the implementation examples;

[0036] Figure 5 This is a schematic diagram illustrating the student knowledge diagnostic matrix using the method described in the embodiment.

[0037] Figure 6 A comparative chart showing the progress of existing advanced knowledge tracking research;

[0038] Figure 7 This is a comparative chart generated from existing advanced student data. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0040] Example:

[0041] Reference Figure 1 The student behavior simulation and knowledge tracking task aims to dynamically assess the evolution of students' cognitive states based on their historical interactions, and thereby generate personalized learning sequences that conform to educational logic. For example, in actual teaching scenarios, it is necessary to infer students' current knowledge weaknesses based on the historical questions they have answered, and to rationally plan the practice path for the next question.

[0042] Traditional work such as Figure 1As illustrated by "LLM-EDU (Large Model-Assisted Education)," this approach typically utilizes large language models (LLMs) to capture latent skill connections and cognitive states from massive amounts of student learning data, attempting to directly generate simulated student question-and-answer sequences. However, this purely generative method lacks strict constraints on pedagogical logic, often leading to inaccuracies in model predictions and generation, producing serious "logic illusions" (such as generating reverse learning paths that violate pre-teaching relationships), and failing to accurately capture students' dynamic cognitive evolution.

[0043] In addition, such as Figure 1 As shown in "RAG-EDU (Retrieval Enhanced Education)," existing retrieval enhancement methods attempt to plan learning paths based on static knowledge graphs. While this approach provides an external reference for large models to some extent, its heavy reliance on static graph structures makes it difficult to adapt to the dynamic changes in knowledge states during real teaching processes. Consequently, its anti-illusion capabilities remain weak, and it completely lacks dynamic adaptability to the evolution of students' cognition.

[0044] The method in this example (DyRAG-KT) is as follows: Figure 1 As shown, from the perspective of dynamic graph structure construction and real-time cognitive state interaction, a solution based on dynamic graph retrieval enhancement is proposed. This method effectively extracts fine-grained knowledge features with clear cognitive boundaries from unstructured text by dynamically calibrating the relational edge weights between knowledge graph nodes during student learning. Furthermore, by introducing a dynamic knowledge tracking mechanism, the retrieval process on the graph is guided by the real-time updated potential knowledge mastery status of students. This mechanism ensures that the final generated learning behavior sequence strictly follows the laws of cognitive development, possesses strong logical consistency and strong anti-illusion properties, and provides a solid guarantee for outputting high-fidelity dynamic student behavior simulation data.

[0045] Reference Figure 2 , Figure 3 A student behavior simulation method based on dynamic graph retrieval enhancement includes the following steps:

[0046] 1) Input educational information and construct a dynamic knowledge graph. Unstructured educational texts are used as input. Problem texts, related knowledge points and thought chains are extracted from the educational texts through text analysis. A dynamic knowledge graph module is constructed to obtain fine-grained dynamic knowledge elements with clear cognitive boundaries. The initial edges of the knowledge graph are established by using the pre-dependency relationship of knowledge points and the co-occurrence similarity of entities. The support feature of node transfer is introduced through frequent itemset mining. After the multi-dimensional features are fused, the relationship weights between nodes are dynamically calibrated, thereby constructing a dynamic knowledge graph that reflects the learning patterns of the student group.

[0047] The module for constructing a dynamic knowledge graph includes: the dynamic knowledge element unit, namely the construction unit of DKEU, and the dynamic calibration unit of edge weights;

[0048] The dynamic knowledge element unit refers to the extraction of question skills and difficulty information from the text through a large language model (Qwen3), and the deconstruction of traditional single knowledge nodes into units containing questions. Related knowledge point set Frequent items and mind chain The quadruple representation is denoted as In order to optimize the reasoning process of the question analysis obtained from the large model. Different problem-solving approaches can be generated through analytical skills. The logical rigor of each branch is scored independently using a large language model as an evaluator. The target node is calculated through a Softmax-based merging operation. The optimal fusion feature is generated based on the overall score.

[0049]

[0050] The edge weight dynamic calibration unit first establishes directed edges based on the pre-dependencies of knowledge points. Undirected edges are constructed based on semantic similarity calculated from entity co-occurrence, and the similarity score between nodes is calculated. Subsequently, frequent itemset mining is introduced to calculate the support score of node transitions in the real historical sequence. The edge weights are dynamically updated by weighting and fusing the node transition pattern tendency score after dynamic calibration of frequent items with the inherent semantic similarity. :

[0051] .

[0052] The weighted dynamic calibration unit dynamically updates the dynamic knowledge graph under construction and is used to update the potential knowledge mastery status of students in subsequent step 3).

[0053] After obtaining dynamic knowledge graph information, the dynamic knowledge tracking module is used to track and update students' potential knowledge mastery status.

[0054] The dynamic knowledge tracking module includes a graph propagation aggregation unit, a time decay update unit, and a state-aware response simulation unit;

[0055] The graph propagation aggregation unit design employs a sparse activation mechanism, performing operations only on the current node and its k-hop neighborhood subgraph to reduce the computational overhead of the entire graph, utilizing attention weights. and Aggregate from the preceding node respectively and similar nodes The state influence, after layer normalization, generates a smooth context proficiency representation. :

[0056]

[0057] The time decay update unit design: After receiving the student's actual interactive response, a time interval-based update unit is introduced. Exponential decay factor To simulate the forgetting curve of human memory, combined with the learning rate Response correctness and similarity calculation of state increment And through explicit numerical truncation functions Force the updated state to be bound to Within the probability interval, the update formula is:

[0058] ;

[0059] After assessing students' mastery of different knowledge points through the dynamic knowledge tracking module, in order to simulate student behavior, the retrieval enhancement generation module uses an adaptive method to match the exercises that need to be completed for different students.

[0060] The retrieval enhancement generation module includes a pattern-adaptive problem sampling unit, a state-aware response simulation unit, and an enhanced content generation unit;

[0061] The mode-adaptive problem sampling unit transforms the retrieval strategy into a Markov decision process and calculates the comprehensive transition score. The score dynamically incorporates logical scores. Pattern Score and historical revisit bias weight Control indication function :

[0062]

[0063] Subsequently, through the temperature parameter The Softmax function converts the overall transition score into transition probability. The candidate questions are determined by comprehensively assessing the students' current knowledge level, calculating the probability of each type of question being selected as the next question, and selecting the most suitable candidate question based on this probability.

[0064] The state-aware response simulation unit is used to determine the virtual student's performance. This unit introduces a hierarchical perception tracking mechanism and gating mechanism, dividing the knowledge information output by the dynamic knowledge tracking module into basic hierarchical information. With high-level information When dealing with basic-level information, assign higher fixed weights to this basic-level information. This anchors and defines the initial cognitive boundaries of virtual students; for higher-level information, an adaptive weight is calculated based on the text content of the higher-level information and related skill features. This allows for the dynamic capture and modification of the student's current state. Basic and high-level information are aggregated through a gating mechanism to generate a fused state vector. Subsequently, the fused state vector The input is fed into a multilayer perceptron, and after nonlinear mapping, the final output is the accuracy of the answer. Based on the accuracy rate of the answers Determine whether the virtual student's response to the selected candidate question is right or wrong;

[0065] The enhanced content generation unit: using the selected question as the anchor point, it recalls standard thought chains and similar error-prone interference items from the vector index of the graph. If the simulated response is correct, the large language model generates the correct reasoning logic based on the contextual standard thought chain; if the simulated response is incorrect, it generates a specific erroneous reasoning path based on the characteristics of the recalled interference items.

[0066] 2) Input the target student's historical interaction sequence. After constructing the dynamic knowledge graph in step 1), use the sparse activation mechanism to extract the current question interaction node and the knowledge graph neighborhood subgraph. Aggregate the influence of the previous node and similar nodes through the graph propagation mechanism to generate context-aware cognitive proficiency features. When a new real interaction response is received, combine the time interval-based decay mechanism with the current proficiency features to calculate and update the student's potential knowledge mastery status in real time.

[0067] 3) Utilizing the knowledge mastery status updated in real time in step 2), perform Markov decision walk computation on the dynamic knowledge graph; jointly evaluate the graph topology connectivity, historical transition patterns, and revisit bias, adaptively calculate and transfer probabilities to each candidate node to sample questions; for the sampled questions, fuse state features at different levels through a gating mechanism to predict the correctness of students' responses, and use the correctness of these responses as anchor points to recall standard thought chains or error-prone interference features in the graph, guiding the large language model to generate learning behavior sequences containing correct reasoning logic or specific erroneous reasoning paths;

[0068] 4) Construct a multi-objective hybrid loss function that includes the main prediction loss of binary cross-entropy and the contrastive learning regularization loss; in different training batches, adopt the training paradigm of machine learning, and optimize the potential knowledge mastery state of students in 2) and the transition probability of candidate nodes in 3). By jointly optimizing the hybrid loss function as an intermediary, calculate the feature distance between positive and negative samples, map the potential state of students to the feature region that matches the actual mastery ability of students, backpropagate to optimize the model parameters, and finally train an optimized model that can output a high-fidelity student learning behavior sequence, which is used to generate high-quality student data and the learning state change curve of each student under different datasets.

[0069] (II) Visualization and Results Analysis

[0070] Figure 4 This example demonstrates in detail the initialization and construction process of the Dynamic Knowledge Graph (DyKG) and its internal hierarchical topology. For example... Figure 4 As shown, in the educational information input and graph construction stage, for each specific question (Qs) in the question bank, the semantic analysis engine of the large language model is used to accurately identify and map the underlying "skill set" or core knowledge concept set (KCs). For example, a question might be parsed to include fine-grained knowledge points from multiple dimensions such as "arithmetic operations," "negative numbers," or "problem solving"; while another question might be mapped to a skill combination of "algebraic expressions" and "equation solving."

[0071] Building upon this foundation, this example does not store these skills as isolated nodes, but rather delves into the inherent logic and progressive connections between skills to construct a... Figure 4 The right side shows the "hierarchical relationships in the knowledge graph." This topological structure intuitively and rigorously demonstrates the coherent hierarchical dependencies (i.e., prerequisite relationships) in teaching. For example, analysis shows that students must first master higher-level macro concepts (such as "arithmetic operations" or "number theory") before they can move down to master more specific, fine-grained micro skills (such as "order of operations" or "factors and multiples").

[0072] Figure 5 The document demonstrates the workflow and beneficial effects of DyRAG-KT by showcasing specific geometric learning scenarios and simulation experiments on real educational datasets such as Edi, EdNet, and ASSIST09.

[0073] 1. Error Root Cause Diagnosis and Graph Propagation Update: At time T6, the student answered incorrectly when solving for the "Area of ​​Composite Figure (Concept2)". Traditional knowledge tracing models would only reduce the mastery of Concept2; however, in this embodiment, the dynamic knowledge tracing module triggers the graph propagation aggregation unit to trace back along the topology of the dynamic graph, automatically diagnosing that the root cause of the error lies in the weak mastery of the prerequisite basic knowledge "Geometric Decomposition (Concept6)" and the related knowledge "Area Formula (Concept3)". Subsequently, the time decay update unit, combined with response weights, synchronously and smoothly adjusts the mastery status of these related nodes.

[0074] 2. Upon entering the next generation step, the retrieval enhancement generation module initiates adaptive question sampling based on the updated cognitive state. By calculating the comprehensive transfer score, it identifies weak points in prior knowledge and retrieves and locates DKEU units that explicitly contain the standard thought chain of "decomposing into rectangles." Subsequently, the state-aware response simulation unit uses hierarchical perception to track and process the knowledge information of this node. Through a gating mechanism, it assigns extremely high weights to basic geometric concepts and adaptively adjusts the weights of higher-order decomposition skills. The fused state is then used for MLP inference to predict that the probability of the student answering incorrectly is extremely high. Finally, after receiving the signal of this failed answer and the retrieved features of similar interference items, the large language model generates a simulated interaction log that accurately reflects the "logical defects in geometric decomposition."

[0075] 3. Dynamic Compensation and Weight Calibration: As targeted learning progresses, at time T10, students complete a series of exercises targeting the basic decomposition. Positive feedback from the dynamic knowledge tracking module leads to an increase in mastery of previous nodes, and the model accurately predicts that students' performance on future complex problems (Concept2) will improve. Simultaneously, the dynamic edge weight calibration mechanism, by exploiting students' frequent jumps between these specific geometric concepts, increases the edge weights from Concept6 to Concept2, causing a dynamic evolution of the graph structure (from sparse to dense interconnections).

[0076] (III) Conclusion Analysis

[0077] Comparing visualization and simulation results reveals that the proposed model significantly outperforms the baseline model in terms of accuracy in cognitive state tracking, logical consistency in generated learning paths, and robustness in predictions with sparse data. This result validates the technical advantages of this proposed model in combining dynamic knowledge graphs with retrieval-enhanced generation to eliminate the "logical illusion" of large language models. Through multi-stage cognitive simulation and feature processing, the model achieves a precise understanding of students' complex and dynamically evolving learning trajectories, thereby significantly improving the overall performance and data fidelity of knowledge tracking and behavior sequence generation tasks in intelligent education systems.

[0078] (iv) Performance Comparison

[0079] like Figure 6 As shown, experiments were conducted on four real-world educational datasets (ASSIST09, ASSIST12, EdNet, and Edi), comprehensively comparing our method (DyRAG-KT) with several mainstream state-of-the-art methods. These baseline methods can be broadly categorized into three types: sequence- and memory-enhanced models (such as DKT, AKT, and MIKT), graph-based learning models (such as GKT and GIKT), and generative simulation models (such as Agent4Edu). In the knowledge tracing (KT) prediction task, our method achieved the highest scores on all four datasets. For example, the AUCs on the ASSIST09 and Edi datasets reached 81.12% and 81.29%, respectively, significantly outperforming all sequence and graph model baselines. Particularly when facing the highly fragmented and self-learning EdNet dataset, our method demonstrated a significant advantage over other state-of-the-art models, indicating that our model has greatly enhanced dynamic reasoning and state tracking capabilities in complex, leap-based learning scenarios.

[0080] like Figure 7 As shown, in student behavior generation and simulation tasks, the method presented in this example demonstrates strong competitiveness compared to traditional supervised simulation baselines (such as DAISIM and KES) in metrics such as ACC and F1-score, and even surpasses the Agent4Edu large model designed specifically for generation in several metrics. This advantage strongly validates that the high-fidelity simulation data generated by combining a large language model with dynamic graph retrieval enhancement not only avoids producing "logical illusions" that violate cognitive laws, but also significantly improves the performance of graph models in knowledge tracing tasks. In scenarios such as "cold start" (i.e., extremely short student historical interaction sequences), traditional models often experience a precipitous drop in performance due to a lack of prior knowledge. However, the method presented in this example uses high-quality synthetic data to fill the sparse gaps in real data, highlighting the crucial role of synthetic data in enhancing model generalization ability and maintaining diagnostic stability.

[0081] Regarding data augmentation and generalization mechanisms, many existing studies attempt to generate additional samples by introducing adversarial networks or randomly combining graph nodes to counteract the model's dependence on static distributions. However, such methods often require extremely high-quality generated samples. If the constructed sample pairs (such as inverse learning sequences) fail to effectively follow the teaching logic, not only will the improvement be limited, but new noise will also be introduced. Furthermore, blindly constructing data detached from students' real-time states significantly increases computation time and cost. This example method, without relying on extensive external manual annotation and complex adversarial mechanisms, achieves reasonable question sampling and path generation based on real-time cognitive states by establishing a two-way feedback loop between dynamic knowledge tracking and retrieval generation. This not only demonstrates that this method no longer excessively relies on the static distribution of a specific dataset, but also showcases its outstanding efficiency and practical feasibility in the implementation of real-world smart education systems.

Claims

1. A student behavior simulation method based on dynamic graph retrieval enhancement, characterized in that, Includes the following steps: 1) Input educational information and construct a dynamic knowledge graph. Unstructured educational texts are used as input. Problem texts, related knowledge points and thought chains are extracted from the educational texts through text analysis. A dynamic knowledge graph module is constructed to obtain fine-grained dynamic knowledge elements with clear cognitive boundaries. The initial edges of the knowledge graph are established by using the pre-dependency relationship of knowledge points and the co-occurrence similarity of entities. The support feature of node transfer is introduced through frequent itemset mining. After the multi-dimensional features are fused, the relationship weights between nodes are dynamically calibrated, thereby constructing a dynamic knowledge graph that reflects the learning patterns of the student group. 2) Input the target student's historical interaction sequence. After the dynamic knowledge graph constructed in step 1), the sparse activation mechanism is used to extract the current question interaction node and the knowledge graph neighborhood subgraph. The influence of the previous node and similar nodes is aggregated through the graph propagation mechanism to generate context-aware cognitive proficiency features. Upon receiving a new real interactive response, the system combines a time-interval-based decay mechanism with the student's current proficiency characteristics to calculate and update the student's potential knowledge mastery status in real time. 3) Utilizing the knowledge mastery status updated in real time in step 2), perform Markov decision walk computation on the dynamic knowledge graph; jointly evaluate the graph topology connectivity, historical transition patterns, and revisit bias, adaptively calculate and sample questions based on the transition probabilities of each candidate node; for the sampled questions, fuse state features at different levels through a gating mechanism to predict the correctness of the student's response, and use the correctness of the response as an anchor point to recall standard thinking chains or error-prone interference features in the dynamic knowledge graph, guiding the large language model to generate learning behavior sequences containing correct reasoning logic or specific erroneous reasoning paths; 4) Construct a multi-objective hybrid loss function that includes the main prediction loss of binary cross-entropy and the contrastive learning regularization loss; in different training batches, adopt the machine learning training paradigm, aim to optimize the student's potential knowledge mastery state in step 2) and the candidate node transition probability in step 3), calculate the feature distance between positive and negative samples by jointly optimizing the hybrid loss function as an intermediary, map the student's potential state to the feature region that matches the student's actual mastery ability, backpropagate to optimize the model parameters, and finally train an optimized model that can output a high-fidelity student learning behavior sequence, which is used to generate high-quality student data and the learning state change curve of each student under different datasets.

2. The student behavior simulation method based on dynamic graph retrieval enhancement according to claim 1, characterized in that, The dynamic knowledge graph module constructed in step 1) includes: dynamic knowledge element unit, namely the construction unit of DKEU, and edge weight dynamic calibration unit; The dynamic knowledge element unit extracts question skills and difficulty information from educational texts through a large language model, deconstructing traditional single knowledge nodes into units containing questions. Related knowledge point set Frequent items and mind chain The quadruple representation is denoted as Optimize the reasoning process of question parsing obtained from large language models. To generate different branches of problem-solving approaches. A large language model is used as the evaluator to independently score the logical rigor of each branch. And calculate the target node based on the Softmax merging operation. The optimal fusion feature is generated based on the overall score. The edge weight dynamic calibration unit first establishes directed edges based on the pre-dependencies of knowledge points. Undirected edges are established based on semantic similarity calculated from entity co-occurrence, and similarity scores between nodes are calculated. Subsequently, frequent itemset mining is introduced to calculate the support score of node transitions in the real historical sequence. The edge weights are dynamically updated by weighting and fusing the node transition pattern tendency score after dynamic calibration of frequent items with the inherent semantic similarity. : ； The edge weight dynamic calibration unit dynamically updates the dynamic knowledge graph constructed in step 1) and is used for updating the potential knowledge mastery status of students in the subsequent step 3). After obtaining dynamic knowledge graph information, the dynamic knowledge tracking module is used to track and update students' potential knowledge mastery status.

3. The student behavior simulation method based on dynamic graph retrieval enhancement according to claim 2, characterized in that, The dynamic knowledge tracking module includes a graph propagation aggregation unit, a time decay update unit, and a state-aware response simulation unit. The graph propagation aggregation unit design employs a sparse activation mechanism, performing operations only on the current node and its k-hop neighborhood subgraph to reduce the computational overhead of the entire graph, utilizing attention weights. and Aggregate from the preceding node respectively and similar nodes The state influence, after layer normalization, generates a smooth context proficiency representation. : The time decay update unit design: After receiving the student's actual interactive response, a time interval-based update unit is introduced. Exponential decay factor To simulate the forgetting curve of human memory, combined with the learning rate Response correctness and similarity calculation of state increment And through explicit numerical truncation functions Force the updated state to be bound to Within the probability interval, the update formula is: ； After assessing students' mastery of different knowledge points through the dynamic knowledge tracking module, in order to simulate student behavior, the retrieval enhancement generation module uses an adaptive method to match the exercises that need to be completed for different students.

4. The student behavior simulation method based on dynamic graph retrieval enhancement according to claim 3, characterized in that, The retrieval enhancement generation module includes a pattern-adaptive problem sampling unit, a state-aware response simulation unit, and an enhanced content generation unit; The mode-adaptive problem sampling unit transforms the retrieval strategy into a Markov decision process and calculates the comprehensive transition score. The score dynamically incorporates logical scores. Pattern Score and historical revisit bias weight Control indication function : Subsequently, through the temperature parameter The Softmax function converts the overall transition score into transition probability. The selection of candidate questions is determined by comprehensively evaluating the student's current knowledge level, calculating the probability of each type of question being selected as the next question, and selecting the most suitable candidate question based on this probability. The state-aware response simulation unit is used to determine the virtual student's performance. This unit introduces a hierarchical perception tracking mechanism and gating mechanism, dividing the knowledge information output by the dynamic knowledge tracking module into basic hierarchical information. With high-level information When dealing with basic-level information, assign higher fixed weights to this basic-level information. This anchors and defines the initial cognitive boundaries of the virtual student; for higher-level information, an adaptive weight is calculated based on the text content and relevant skill features of the higher-level information. This allows for the dynamic capture and modification of the student's current state. Basic and high-level information are aggregated by a gating mechanism to generate a fused state vector. ; Subsequently, the fused state vector The input is fed into a multilayer perceptron, and after nonlinear mapping, the final output is the accuracy of the answer. Based on the accuracy rate of the answers Determine whether the virtual student's response to the selected candidate question is right or wrong; The enhanced content generation unit: using the selected question as the anchor point, it recalls standard thought chains and similar error-prone interference items from the vector index of the graph. If the simulated response is correct, the large language model generates the correct reasoning logic based on the contextual standard thought chain; if the simulated response is incorrect, it generates a specific erroneous reasoning path based on the characteristics of the recalled interference items.

Images