Student cognitive path analysis and personalized feedback method, storage medium and device
By analyzing students' cognitive trajectories using state-space models and graph neural networks, and combining dynamic knowledge graphs and reinforcement learning, the problem of existing intelligent grading systems being unable to provide personalized feedback has been solved, enabling accurate diagnosis of students' cognitive states and personalized teaching.
Patent Information
- Application Number
- CN202610708301.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-21
- Publication Date
- 2026-07-21
Smart Images
Figure SMS_1 
Figure SMS_5 
Figure SMS_9
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent education technology, specifically to a method, storage medium, and device for student cognitive path analysis and personalized feedback. Background Technology
[0002] Currently, existing intelligent grading systems mainly rely on rule matching or shallow machine learning models to judge the correctness of students' answers. Their function is limited to identifying whether the answer is correct, and they cannot deeply analyze the cognitive biases and errors in thinking paths exhibited by students during the problem-solving process. Such systems usually organize teaching content using a static knowledge point structure, lacking the ability to model the dynamic correlation strength between knowledge points, and cannot adjust the weights of the knowledge network in real time according to the error patterns of the student group. In addition, existing technologies mostly use fixed wrong question recommendations or general learning resource pushes in the feedback mechanism, without combining the individual cognitive state evolution process to generate personalized intervention strategies, resulting in insufficient targeted tutoring suggestions. Although some systems have introduced knowledge graphs to represent the logical relationships of subjects, they still remain at the static topological level, failing to integrate graph neural networks for error propagation tracing, and unable to achieve the leap from "outcome judgment" to "process diagnosis". Summary of the Invention
[0003] To address the shortcomings of existing technologies, this invention aims to provide a method, storage medium, and device for student cognitive path analysis and personalized feedback.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: A method for analyzing student cognitive paths and providing personalized feedback includes the following steps: S1. When students practice multiple questions in a particular subject, multimodal behavioral data of the students during the problem-solving process is collected simultaneously, and the students' cognitive trajectory is constructed based on a state-space model. The cognitive trajectory is represented as a time-series cognitive state vector S. t ,satisfy:
[0005] F(S t 1)=σ(W S t 1+b) Among them, S t and S t-1 Let F() represent the cognitive state vector of the student at times t and t-1, and let ε be the nonlinear state transition function. t Let t be the random perturbation term representing carelessness or fluctuations in attention; W represents the state transition weight matrix, which quantifies the influence of each dimension of the cognitive state at the previous moment on the cognitive state at the current moment, reflecting the inherent logical dependence between knowledge points and the cognitive transfer law; b represents the bias vector, which is used to adjust the baseline level of state transition, accommodating the basic state deviation caused by individual differences or external teaching interventions; σ represents the nonlinear activation function, which is used to map the result after linear transformation to a reasonable range of cognitive state values, ensuring the boundedness and nonlinearity of cognitive state values. The multimodal behavioral data includes the shape and speed features of handwritten stroke trajectory, the rhythm and frequency features of keyboard input, the sequential logic features of answering steps, the distribution density features of mouse click hotspots, and the textual semantic features of answering subjective questions step by step. Each element in the cognitive state vector represents the multimodal quantitative features of the student's problem-solving behavior at a specific moment. S2. Construct an initial knowledge graph G=(V,E,W) based on subject knowledge points, where V is the set of knowledge point nodes, E is the set of logical dependency edges between knowledge point nodes, and W is the initial weight on the logical dependency edges; the weight W will be dynamically updated according to the error co-occurrence frequency of the student group, so that the knowledge graph evolves with the learning process. S3, the cognitive state vector S obtained in step S1 t The node features are input into the Graph Neural Network (GNN), which processes the cognitive state vector S through a message-passing mechanism. t Conduct error propagation path analysis to identify the root cause knowledge point deficiencies that lead to incorrect answers; S4. Establish a state-action mapping model Q(s,a)=E[R|s,a] based on the reinforcement learning framework, where s is the student's current knowledge mastery state and a is the recommended learning intervention action. Train the state-action mapping model using historical learning effect data to generate optimal personalized feedback suggestions. The historical learning effect data specifically includes records of the student's past learning intervention actions, changes in knowledge mastery state after intervention, persistent feature data of cognitive impairment, and statistics on the error recurrence frequency of knowledge points. S5. Through topological data analysis, extract the persistent intervals of cognitive obstacles for each knowledge point in the student's learning process, and generate persistent pairs PD={(bi,di)}, where bi and di represent the first error time and the correction completion time of the i-th cognitive point with cognitive obstacles, respectively. Based on this, construct a three-dimensional visualized learning path map to show learning bottlenecks and development trends.
[0006] Furthermore, in step S2, the weight W of the logical dependency edge between knowledge point i and knowledge point j... ij The dynamic update formula for ∈W is:
[0007] Among them, weight and These are the weights before and after the update, respectively. Let α represent the conditional probability that making a mistake on knowledge point i will lead to failure in solving the problem on knowledge point j. α∈(0,1) is a decay factor used to preserve the strength of historical associations.
[0008] Furthermore, in step S3, the graph neural network adopts either a graph convolutional network (GCN) or a graph attention network (GAT), and its (l+1)th hidden state is represented as follows:
[0009] in, To add the adjacency matrix after adding self-loops, for The corresponding degree matrix, H (l) For the knowledge point embedding representation of layer l, Θ (l) Let be the trainable parameter matrix of the l-th layer, and σ be the activation function.
[0010] Furthermore, in step S3, the causal significance of the error propagation path output by the graph neural network (GNN) is further tested by combining the causal reasoning model to exclude knowledge points that are only correlated but have no causal relationship.
[0011] Furthermore, in step S4, the learning intervention actions include one or more of the following: pushing review micro-lesson videos, assigning special practice question sets, pointing out common thinking errors, guiding students to redo typical wrong questions, and recommending peer learning partners.
[0012] Furthermore, in step S4, the specific process of training the state-action mapping model is as follows: the state-action mapping model Q(s,a) is trained using a deep Q-network (DQN) or a proximal policy optimization (PPO) algorithm, and an exploration-exploitation mechanism is introduced to balance new policy attempts with existing effective policy executions.
[0013] Furthermore, in step S5, the three axes of the three-dimensional visualized learning path map are time dimension, knowledge point difficulty level, and cognitive stability index, respectively. Cluster analysis can be used to identify plateau periods, abrupt improvement phases, and recurring oscillation zones, assisting teachers in developing intervention strategies.
[0014] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.
[0015] The present invention also provides a computer device, including a processor and a memory, wherein the memory is used to store a computer program; and the processor is used to execute the computer program to implement the above-described method.
[0016] The beneficial effects of this invention are as follows: By constructing a cognitive trajectory analysis mechanism based on a state-space model, integrating a knowledge graph with dynamic weight updates, combining error propagation tracing with graph neural networks, and a feedback decision engine based on reinforcement learning, this invention achieves explicit expression and accurate diagnosis of students' implicit cognitive states during problem-solving; the introduced topological data analysis method can generate a three-dimensional visualized learning path map, intuitively presenting students' cognitive bottlenecks and development trends at each knowledge point.
[0017] This invention forms a closed-loop educational intervention mechanism of "perception-analysis-decision-feedback-optimization", which significantly improves the interpretability, personalization and teaching practicality of the intelligent grading system. It helps teachers to teach accurately and students to reflect independently, and has good value for promotion and application. Detailed Implementation
[0018] The present invention will be further described below. It should be noted that this embodiment is based on the present technical solution and provides detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to this embodiment.
[0019] This embodiment provides a method for analyzing student cognitive paths and providing personalized feedback, including the following steps: S1. When students practice multiple questions in a particular subject, multimodal behavioral data of the students during the problem-solving process is collected simultaneously, and the students' cognitive trajectory is constructed based on a state-space model. The cognitive trajectory is represented as a time-series cognitive state vector S. t ,satisfy:
[0020] F(S t 1)=σ(W S t 1+b) Among them, S t and S t-1 Let F() represent the cognitive state vector of the student at times t and t-1, and let ε be the nonlinear state transition function. t Let t be the random perturbation term representing carelessness or fluctuations in attention; W represents the state transition weight matrix, used to quantify the influence of each dimension of the cognitive state at the previous moment on the cognitive state at the current moment, reflecting the inherent logical dependence between knowledge points and the cognitive transfer law; b represents the bias vector, used to adjust the baseline level of state transition, accommodating the basic state deviation caused by individual differences or external teaching intervention; σ represents the nonlinear activation function, using the Sigmoid function, to map the result after linear transformation to a reasonable range of cognitive state values, ensuring the boundedness and nonlinearity of cognitive state values; Each element in the cognitive state vector represents a multimodal quantitative feature of a student's problem-solving behavior at a specific moment. Specifically, the composition of the cognitive state vector corresponds to five categories of collected multimodal behavioral data, including the morphological and speed features of handwritten stroke trajectories, the rhythm and frequency features of keyboard input, the sequential logical features of answering steps, the distribution density features of mouse click hotspots, and the textual semantic features of step-by-step answers to subjective questions. These elements together constitute an explicit numerical representation of the student's implicit cognitive state.
[0021] It should be noted that the multimodal behavioral data collection phase is conducted simultaneously while students practice multiple questions. The data collected focuses on students' multimodal behavioral data throughout the entire problem-solving process, i.e., their overall behavioral performance during the entire practice session. The cognitive state vector is defined as a time-series cognitive trajectory, reflecting the changes and transitions in students' cognitive states at different times through a state-space model. This continuous trajectory construction requires complete problem-solving process data to ensure that the dynamic evolution of cognitive states can be captured.
[0022] S2. Construct an initial knowledge graph G=(V,E,W) based on subject knowledge points, where V is the set of knowledge point nodes, E is the set of logical dependency edges between knowledge point nodes, and W is the initial weight on the logical dependency edges; the weight W will be dynamically updated according to the error co-occurrence frequency of the student group, so that the knowledge graph evolves with the learning process. S3, the cognitive state vector S obtained in step S1 t The node features are input into the Graph Neural Network (GNN), which processes the cognitive state vector S through a message-passing mechanism. t Conduct error propagation path analysis to identify the root cause knowledge point deficiencies that lead to incorrect answers; S4. Establish a state-action mapping model Q(s,a)=E[R|s,a] based on the reinforcement learning framework, where s is the student's current knowledge mastery state and a is the recommended learning intervention action. Train the state-action mapping model using historical learning effect data to generate the optimal personalized feedback suggestion. The historical learning outcome data specifically includes records of learning interventions previously used by students, changes in knowledge mastery after intervention, persistent characteristics of cognitive impairment, and statistics on the frequency of error recurrence of knowledge points.
[0023] S5. Through topological data analysis, extract the persistent intervals of cognitive obstacles for each knowledge point in the student's learning process, and generate persistent pairs PD={(bi,di)}, where bi and di represent the first error time and the correction completion time of the i-th cognitive point with cognitive obstacles, respectively. Based on this, construct a three-dimensional visualized learning path map to show learning bottlenecks and development trends.
[0024] In this embodiment, step S1 involves collecting the handwritten handwriting trajectory using a terminal device used by the student. Specifically, when the student writes in the answer area of the terminal device, the technical interface of the touchscreen or digital writing tablet is used to record in real time the sequence of coordinate points of pen contact, the writing pressure value, and the timestamp corresponding to each coordinate point. These discrete data points are connected into a continuous trajectory line in chronological order and bound to the spatial position of the answer interface, thereby forming complete handwritten handwriting trajectory data that includes writing speed, stroke order, and correction marks.
[0025] In this embodiment, in step S2, the weight W of the logical dependency edge between knowledge point i and knowledge point j ij The dynamic update formula for ∈W is:
[0026] Among them, weight and These are the weights before and after the update, respectively. Let α represent the conditional probability that making a mistake on knowledge point i will lead to failure in solving the problem on knowledge point j. α∈(0,1) is a decay factor used to preserve the strength of historical associations.
[0027] The dynamic update formula described above essentially asks, "If knowledge point j is difficult to learn, is it because knowledge point i was not learned well?" Updating the weights of logically dependent edges in the knowledge graph using this dynamic update formula essentially involves automatically adjusting the correlation strength between two knowledge points by continuously analyzing the answers to this question. α represents the decay factor; for example, setting it to 0.8 means "retaining 80% old experience and adding 20% new observations."
[0028] In this embodiment, in step S3, the graph neural network adopts either a graph convolutional network (GCN) or a graph attention network (GAT), and its (l+1)th hidden state is represented as follows:
[0029] in, To add the adjacency matrix after adding self-loops, for The corresponding degree matrix, H (l) For the knowledge point embedding representation of layer l, Θ (l) Let be the trainable parameter matrix of the l-th layer, and σ be the activation function.
[0030] It should be noted that, This is an adjacency matrix that records which knowledge points are related. If two knowledge points are related, the corresponding position is 1; otherwise, it is 0. This is used to count the number of knowledge point nodes connected to each knowledge point. This indicates weight averaging, and its function is to normalize. This enables information transfer; together, these components form a normalized adjacency matrix. Its function is to distribute information from each knowledge point evenly to its neighbors, or to receive information from its neighbors. (l) This refers to the current cognitive state. σ is the activation function, a non-linear function (such as ReLU), which introduces judgment. It determines which information is worth retaining and which can be ignored. H(l+1) represents the updated cognitive state.
[0031] In this embodiment, in step S3, the causal significance of the error propagation path output by the graph neural network (GNN) is further tested by combining the causal reasoning model to exclude the erroneous association of knowledge points that are only correlated but have no causal relationship.
[0032] In this embodiment, in step S4, the learning intervention actions include one or more of the following: pushing review micro-lesson videos, assigning special practice question sets, pointing out common thinking errors, guiding students to redo typical wrong questions, and recommending peer learning partners.
[0033] In this embodiment, the specific process of training the state-action mapping model in step S4 is as follows: the state-action mapping model Q(s,a) is trained using a deep Q-network (DQN) or a proximal policy optimization (PPO) algorithm, and an exploration-exploitation mechanism is introduced to balance new policy attempts with existing effective policy executions.
[0034] In this embodiment, in step S5, the three coordinate axes of the three-dimensional visualized learning path map are time dimension, knowledge point difficulty level, and cognitive stability index, respectively. Cluster analysis can be used to identify plateau periods, sudden improvement phases, and repeated oscillation zones, assisting teachers in formulating intervention strategies.
[0035] Example 2 This embodiment provides an application example of the method described in Embodiment 1.
[0036] Xiao Li completed 10 practice problems on "quadratic equations in one variable" in an online assignment.
[0037] After Xiao Li submitted his answers, the system first collected multimodal behavioral data on his actions during the answering process (including the handwritten derivation trajectory at each step, the number of deletions and modifications, and skipped steps). Analysis of this multimodal behavioral data revealed that in question 3, Xiao Li skipped the "completing the form and simplifying" step and directly wrote down the result; in question 5, he repeatedly modified symbols but failed to correct the error. These behaviors were modeled as cognitive state vectors in the state space. This suggests that he exhibits cognitive characteristics such as "relying on intuition for leaps and having a weak sense of symbolism."
[0038] Next, the dynamic knowledge graph G=(V,E,W) of the current subject is invoked, where nodes contain knowledge points such as "perfect square formula," "completing the square method," "quadratic formula," and "discriminant application," and edges represent the logical dependencies between these knowledge points. Based on recent data from all students in the school, the weight W of the path "completing the square method" → "quadratic formula" has been automatically adjusted. ij This is because statistics show that students who make mistakes at this stage are highly likely to make a chain reaction of errors later on.
[0039] Subsequently, a Graph Neural Network (GNN) model was launched to analyze the error propagation path. Although Xiao Li used the quadratic formula for question 7 and the calculation process seemed independent, the GNN traced the problem through message passing and found that the initial equation for the question was not correctly standardized. The root cause was still his misunderstanding of the "essence of the completing the square method." Therefore, it was determined that "a weak grasp of the completing the square method" was the common cognitive deficiency that led to his consecutive errors in questions 3, 5, 7, and 9, rather than an isolated error.
[0040] Based on this diagnosis, the personalized feedback generation phase began. The reinforcement learning model Q(s,a) comprehensively evaluated Xiao Li's historical learning data and current state, determining that "watching a 3-minute micro-lesson + completing 3 step-by-step guided questions" was the optimal intervention strategy, superior to blindly pushing 10 similar exercises. Therefore, an animated explanation video titled "Why are there always missing items in the formula?" was pushed to Xiao Li, along with a "step-by-step fill-in-the-blank" exercise to help him rebuild his reasoning chain.
[0041] Finally, this learning record was incorporated into the student's personal growth portfolio, and new persistent pairs (bi, di) were extracted using topological data analysis. For example, the "completion method confusion" problem first appeared in week 3 and has not recurred after this intervention, forming a potential "resolution interval." All data was compiled into a three-dimensional learning path map, which was visualized on the teacher's end: the X-axis represents time, the Y-axis represents the difficulty of the knowledge point, and the Z-axis represents the error recurrence frequency. After reviewing the data, the homeroom teacher, Ms. Wang, found that 8 students in the class experienced peak oscillations in the same area, and therefore decided to dedicate the next lesson to explaining the "structure preservation principle in algebraic transformations."
[0042] Throughout the process, not only were the grading tasks completed, but a leap was also achieved from "looking at the results" to "reading the thinking," and from "providing answers" to "addressing the root causes." Students received precise guidance, and teachers gained a comprehensive understanding of the group's learning situation, truly demonstrating the personalized and scientific value of intelligent education.
[0043] For those skilled in the art, various corresponding changes and modifications can be made based on the above technical solutions and concepts, and all such changes and modifications should be included within the protection scope of the claims of this invention.
Claims
1. A method for analyzing student cognitive paths and providing personalized feedback, characterized in that, Includes the following steps: S1. When students practice multiple questions in a particular subject, multimodal behavioral data of the students during the problem-solving process is collected simultaneously, and the students' cognitive trajectory is constructed based on a state-space model. The cognitive trajectory is represented as a time-series cognitive state vector S. t ,satisfy: F(S t 1)=σ(W S t 1+b) Among them, S t and S t-1 Let F() represent the cognitive state vector of the student at times t and t-1, and let ε be the nonlinear state transition function. t Let t be the random perturbation term representing carelessness or fluctuations in attention; W represents the state transition weight matrix, which quantifies the influence of each dimension of the cognitive state at the previous moment on the cognitive state at the current moment, reflecting the inherent logical dependence between knowledge points and the cognitive transfer law; b represents the bias vector, which is used to adjust the baseline level of state transition, accommodating the basic state deviation caused by individual differences or external teaching interventions; σ represents the nonlinear activation function, which is used to map the result after linear transformation to a reasonable range of cognitive state values, ensuring the boundedness and nonlinearity of cognitive state values. The multimodal behavioral data includes the shape and speed features of handwritten stroke trajectory, the rhythm and frequency features of keyboard input, the sequential logic features of answering steps, the distribution density features of mouse click hotspots, and the textual semantic features of answering subjective questions step by step. Each element in the cognitive state vector represents the multimodal quantitative features of the student's problem-solving behavior at a specific moment. S2. Construct an initial knowledge graph G=(V,E,W) based on subject knowledge points, where V is the set of knowledge point nodes, E is the set of logical dependency edges between knowledge point nodes, and W is the initial weight on the logical dependency edges; the weight W will be dynamically updated according to the error co-occurrence frequency of the student group, so that the knowledge graph evolves with the learning process. S3, the cognitive state vector S obtained in step S1 t The node features are input into the Graph Neural Network (GNN), which processes the cognitive state vector S through a message-passing mechanism. t Conduct error propagation path analysis to identify the root cause knowledge point deficiencies that lead to incorrect answers; S4. Establish a state-action mapping model Q(s,a)=E[R|s,a] based on the reinforcement learning framework, where s is the student's current knowledge mastery state and a is the recommended learning intervention action. Train the state-action mapping model using historical learning effect data to generate optimal personalized feedback suggestions. The historical learning effect data specifically includes records of the student's past learning intervention actions, changes in knowledge mastery state after intervention, persistent feature data of cognitive impairment, and statistics on the error recurrence frequency of knowledge points. S5. Through topological data analysis, extract the persistent intervals of cognitive obstacles for each knowledge point in the student's learning process, and generate persistent pairs PD={(bi,di)}, where bi and di represent the first error time and the correction completion time of the i-th cognitive point with cognitive obstacles, respectively. Based on this, construct a three-dimensional visualized learning path map to show learning bottlenecks and development trends.
2. The method according to claim 1, characterized in that, In step S2, the weight W of the logical dependency edge between knowledge point i and knowledge point j ij The dynamic update formula for ∈W is: Among them, weight and These are the weights before and after the update, respectively. Let α represent the conditional probability that making a mistake on knowledge point i will lead to failure in solving the problem on knowledge point j. α∈(0,1) is a decay factor used to preserve the strength of historical associations.
3. The method according to claim 1, characterized in that, In step S3, the graph neural network uses either a graph convolutional network (GCN) or a graph attention network (GAT), and its (l+1)th hidden state is represented as follows: in, To add the adjacency matrix after adding self-loops, for The corresponding degree matrix, H (l) For the knowledge point embedding representation of layer l, Θ (l) Let be the trainable parameter matrix of the l-th layer, and σ be the activation function.
4. The method according to claim 1, characterized in that, In step S3, the causal significance of the error propagation path output by the graph neural network (GNN) is further tested by combining the causal reasoning model to exclude knowledge points that are only correlated but have no causal relationship.
5. The method according to claim 1, characterized in that, In step S4, the learning intervention actions include one or more of the following: pushing review micro-lesson videos, assigning special practice question sets, pointing out common thinking errors, guiding students to redo typical wrong questions, and recommending peer learning partners.
6. The method according to claim 1, characterized in that, In step S4, the specific process of training the state-action mapping model is as follows: the state-action mapping model Q(s,a) is trained using a deep Q-network (DQN) or a proximal policy optimization (PPO) algorithm, and an exploration-exploitation mechanism is introduced to balance new policy attempts with existing effective policy executions.
7. The method according to claim 1, characterized in that, In step S5, the three axes of the three-dimensional visualized learning path map are time dimension, knowledge point difficulty level, and cognitive stability index. Cluster analysis can be used to identify plateau periods, abrupt improvement phases, and recurring oscillation zones, assisting teachers in developing intervention strategies.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.
9. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer program; the processor being used to execute the computer program to implement the method of any one of claims 1-7.