Cognitive structure diagnosis system based on edge feature map attention network

Abstract

The invention discloses a cognitive structure diagnosis system based on an edge feature map attention network, and belongs to the technical field of cognitive diagnosis, and the system comprises an embedded module which is used for mapping knowledge concepts, the relation between the concepts, the state of a learner and exercises to a low-dimensional vector space, and achieving the efficient representation of features; the cognitive structure state representation module is used for dynamically modeling the knowledge state and the knowledge structure state of the learner; the prediction module is used for predicting the answer result of the learner and improving the diagnosis accuracy and interpretability in combination with a cross entropy loss function optimization model; according to the cognitive structure diagnosis system based on the edge feature map attention network, accuracy, real-time performance and interpretability of cognitive diagnosis are remarkably improved, a more accurate and comprehensive cognitive diagnosis tool is provided for an intelligent education system, and the cognitive structure diagnosis system has high universality and expansibility.

Description

Technical Field

[0001] The present invention relates to the technical field of cognitive diagnosis, and in particular to a cognitive structure diagnosis system based on an edge feature map attention network. Background Art

[0002] With the rapid development of the intelligent education field, cognitive diagnosis (CD), as a key technology, is widely applied in multiple fields such as personalized education, games, and medical diagnosis. The goal of cognitive diagnosis is to help achieve more personalized and efficient teaching through the accurate assessment of learners' cognitive states. Existing cognitive diagnosis methods mainly focus on evaluating learners' mastery of individual knowledge concepts. However, in actual teaching, learners' understanding of the cognitive structure state is equally crucial. The cognitive structure state mainly consists of two parts: the knowledge state (KS) and the knowledge structure state (KUS). Although existing technologies have considered the influence of the cognitive structure state, they often use the knowledge structure state as an implicit feature to improve and enhance the representation of the knowledge state, without modeling and evaluating the knowledge structure state, thus limiting the comprehensiveness and accuracy of cognitive diagnosis.

[0003] In the field of cognitive diagnosis, how to simultaneously consider learners' knowledge state and knowledge structure state, and achieve the dynamic modeling and updating of both, remains a technical problem to be urgently solved. The deficiencies of existing methods directly affect the accuracy and interpretability of personalized teaching, and limit the effectiveness of intelligent education systems in practical applications.

[0004] Therefore, how to effectively model learners' cognitive structure state and dynamically update it is an important direction for improving the performance of cognitive diagnosis methods and achieving precise personalized teaching. Aiming at the deficiencies of existing technologies, the present invention proposes a new cognitive structure diagnosis framework - a cognitive structure state-based cognitive diagnosis framework (CSTCD) based on an edge feature map attention network. This framework can dynamically model and update learners' knowledge state and knowledge structure state, thereby comprehensively evaluating learners' cognitive structure, improving the accuracy and interpretability of cognitive diagnosis, and providing a more accurate and comprehensive cognitive diagnosis tool for intelligent education systems. Summary of the Invention

[0005] The objective of the present invention is to provide a cognitive structure diagnosis system based on an edge feature map attention network, so as to solve the problem that existing cognitive diagnosis methods cannot comprehensively evaluate the cognitive structure state of learners, be able to accurately evaluate the cognitive structure of learners, and achieve comprehensive diagnosis.

[0006] To achieve the above objective, the present invention provides a cognitive structure diagnosis system based on an edge feature map attention network, including the following modules:

[0007] An embedding module, which is used to map knowledge concepts, relationships between concepts, learner states, and exercises into a low-dimensional vector space to achieve efficient representation of features. This module captures the semantic information of knowledge concepts and their mutual relationships, ensuring an accurate reflection of the learner's cognitive state;

[0008] A cognitive structure state representation module, which is used to dynamically model the knowledge state (KS) and knowledge structure state (KUS) of learners;

[0009] A prediction module, which is used to predict the learner's answer results and optimize the model in combination with the cross-entropy loss function to improve the diagnostic accuracy and interpretability.

[0010] Preferably, the output of the embedding module is the One-Hot encoding of learners, the predecessor-successor relationship graph (directed graph), the correlation relationship graph (undirected graph), and the One-Hot encoding of exercises. The exercises obtain the knowledge dependency vector (Knowledge Relevancy) Q e , the knowledge difficulty vector (Knowledge Difficulty) h diff and the exercise discrimination vector (Exercise Discrimination) h disc .

[0011] Preferably, the cognitive structure state representation module introduces an edge feature map attention network - EGAT to separately learn the learner's mastery of the knowledge state (KS) and knowledge structure state (KUS) in the predecessor-successor relationship graph (directed graph) and the correlation relationship graph (undirected graph). EGAT can adaptively adjust the weights of each knowledge point, reflect the complex dependency relationships between concepts, and thus comprehensively evaluate the learner's cognitive structure. Each EGAT layer is composed of a node attention block (Node AttentionBlock) and an edge attention block (Edge Attention Block).

[0012] Preferably, the attention factor a ij between node i and node j in the node attention block (Node Attention Block) is calculated by the following formula:

[0013]

[0014] Among them, N i represents the set of neighbor nodes of node i, represents the weight vector parameter, represents the edge feature vector between node i and node j, represents the feature vector representation of node i, represents the feature vector representation of node j, || represents the vector concatenation operation, and LeakyReLU is a non-linear activation function;

[0015] According to the attention factor a ij , the update formula of the node feature is:

[0016]

[0017] Among them, h j represents the feature vector representation of node j.

[0018] Preferably, in the Edge Attention Block, the attention factor β pq between edge p and its neighbor edge q is calculated as follows:

[0019]

[0020] Among them, N p represents the set of neighbor edges of edge p, represents the weight vector parameter, represents the feature vector representation of edge p, represents the feature vector representation of edge q, represents the node feature vector between edge p and its neighbor edge q;

[0021] According to the attention factor β pq , the update formula of the edge feature is:

[0022]

[0023] Among them, h q represents the feature vector representation of edge q.

[0024] Preferably, the EGAT layer of the cognitive structure state representation module obtains the final graph representation through the method of Graph Pooling. During the learning process of the learner, a weight is learned for the predecessor-successor relationship graph (directed graph) and the correlation relationship graph (undirected graph) respectively to measure the learning bias weight of the learner;

[0025] The specific formula is as follows:

[0026]

[0027]

[0028] Among them, h s represents the cognitive structure state representation of learner s considering the predecessor-successor relationship and related relationships; represents the cognitive structure state representation of the learner for the predecessor-successor graph; represents the cognitive structure state representation of the learner for the predecessor-successor graph; Readout represents the graph pooling operation; Ψ DRC represents the directed EGAT graph representation learning operation, Ψ URC represents the undirected EGAT graph representation learning operation; h d , e d , A d and h u , e u , A u respectively represent the inputs of directed and undirected EGAT graph representation learning: node feature vector, edge feature vector, and adjacency matrix; W d and W u respectively represent the learning bias weights of the learner for the predecessor-successor graph and the related relationship graph.

[0029] Preferably, the cognitive structure state representation module obtains the diagnostic result of the learner's cognitive structure state through the cognitive diagnosis formula as:

[0030] y = Q e ·(h s - h diff )·h dics (8)

[0031] Among them, y represents the diagnostic result of the learner's cognitive structure state.

[0032] Preferably, the prediction module selects a multi-layer perceptron model (Multi-Layer Perceptron, MLP) as a classifier to predict the learner's answer result. The multi-layer perceptron model models the learner's answer result as a binary classification problem and maps the output to a real value in the interval [0, 1] through a sigmoid activation function, representing the probability that the learner answers correctly. When the predicted value is greater than or equal to 0.5, it is determined that the learner answers correctly; when the predicted value is less than 0.5, it is determined that the learner answers incorrectly.

[0033] Preferably, the specific formula adopted in the prediction module is:

[0034]

[0035] Among them, represents the predicted value of the model, y represents the true answer result, taking values of 0 or 1, indicating the learner's answer is incorrect or correct respectively, W represents the learnable weight matrix, and b represents the learnable bias.

[0036] Preferably, the formula of the cross-entropy loss function in the prediction module is:

[0037]

[0038] Among them, l represents the loss value, which is used to measure the difference between the predicted value and the true value of the model.

[0039] Therefore, the present invention adopts the above-mentioned cognitive structure diagnosis system based on the edge feature map attention network, and has the following beneficial effects:

[0040] (1) Through the embedding module, the cognitive structure state representation module and the prediction module, the dynamic modeling and update of the learner's knowledge state (KS) and knowledge structure state (KUS) are realized;

[0041] (2) Through EGAT, the multiple relationship graph modeling of the predecessor-successor relationship and the correlation relationship is realized, and the adjacency matrices of the directed graph and the undirected graph are combined to represent the node connection relationship, so as to generate a comprehensive and accurate feature representation;

[0042] (3) Through the graph pooling operation and the introduction of the learner-biased weight, the dynamic nature and interpretability of the cognitive structure state representation are further enhanced.

[0043] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0044] Figure 1 is the model framework diagram of a cognitive structure diagnosis system based on the edge feature map attention network of the present invention;

[0045] Figure 2 is the framework diagram of EGAT of a cognitive structure diagnosis system based on the edge feature map attention network of the present invention. Detailed Embodiments

[0046] The following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0047] The following will further illustrate the technical solutions of the present invention through the drawings and embodiments.

[0048] Embodiment

[0049] As Figure 1 shown, the present invention provides a cognitive structure diagnosis system based on an edge feature map attention network (EGAT), that is, the CSTCD model is adopted to solve the problem that the cognitive structure state cannot be diagnosed in the previous cognitive diagnosis methods. The CSTCD model mainly consists of three parts: an embedding module, a cognitive structure state representation module, and a prediction module, and shows the training process of the model.

[0050] The embedding module is used to map knowledge concepts, the relationships between concepts, learner states, and exercises into a low-dimensional vector space to achieve efficient representation of features. This module captures the semantic information of knowledge concepts and their mutual relationships to ensure an accurate reflection of the learner's cognitive state.

[0051] The output of the embedding module is the learner's One-Hot encoding, the predecessor-successor relationship graph (directed graph), the correlation relationship graph (undirected graph), and the One-Hot encoding of the exercise. The learner initializes their initial mastery of the knowledge state (KS) and knowledge structure state (KUS) in these two graphs according to the predecessor-successor relationship graph (directed graph) and the correlation relationship graph (undirected graph).

[0052] The predecessor-successor relationship and the correlation relationship are two knowledge structure states. Specifically, the adjacency matrices of the predecessor-successor relationship graph (directed graph) and the correlation relationship graph (undirected graph) represent the connection relationships between the nodes in the graph. The nodes and edges in the graph are both vectorized to represent the learner's initial mastery of the knowledge state (KS) and knowledge structure state (KUS) in the predecessor-successor relationship graph (directed graph) and the correlation relationship graph (undirected graph).

[0053] The exercise obtains the knowledge dependency vector (KnowledgeRelevancy) Q used in the cognitive diagnosis formula through the embedding module e , the knowledge difficulty vector (Knowledge Difficulty) h diff and the exercise discrimination vector (Exercise Discrimination) h disc .

[0054] The cognitive structure state representation module is used to dynamically model the learner's knowledge state (KS) and knowledge structure state (KUS), and introduces an edge feature map attention network - EGAT to separately learn the learner's mastery of the knowledge state (KS) and knowledge structure state (KUS) in the predecessor-successor relationship graph (directed graph) and the correlation relationship graph (undirected graph). The EGAT framework diagram is as Figure 2As shown, EGAT can adaptively adjust the weights of each knowledge point, reflect the complex dependency relationships between concepts, and thus comprehensively evaluate the cognitive structure of learners.

[0055] Each EGAT layer consists of a Node Attention Block and an Edge Attention Block.

[0056] In the Node Attention Block, the attention factor a between node i and node j ij is calculated as follows:

[0057]

[0058] where, N i represents the set of neighbor nodes of node i, represents the weight vector parameter, represents the edge feature vector between node i and node j, represents the feature vector representation of node i, represents the feature vector representation of node j, || represents the vector concatenation operation, and LeakyReLU is a non-linear activation function;

[0059] According to the attention factor a ij , the update formula for node features is:

[0060]

[0061] where, h j represents the feature vector representation of node j.

[0062] In the Edge Attention Block, the attention factor β between edge p and its neighbor edge q pq is calculated as follows:

[0063]

[0064] where, N p represents the set of neighbor edges of edge p, represents the weight vector parameter, represents the feature vector representation of edge p, represents the feature vector representation of edge q, represents the node feature vector between edge p and its neighbor edge q;

[0065] According to the attention factor β pq , the update formula for edge features is:

[0066]

[0067] Among them, h q represents the eigenvector representation of edge q.

[0068] The EGAT model realizes the efficient learning of node and edge features through the parallel update of the Node Attention Block and the Edge Attention Block, generating a more comprehensive and accurate feature representation. This collaborative optimization mechanism not only effectively integrates the feature information of nodes and edges in the graph but also significantly improves the model's processing ability for complex graph data. Through the design of a multi-layer attention mechanism, EGAT can capture multi-scale feature information, providing strong support for the representation learning of graph data and downstream tasks.

[0069] We write EGAT as a functional function Ψ(·). The EGAT layer obtains the final graph representation through the method of Graph Pooling. During the learning process of the learner, a weight is learned for the predecessor-successor relationship graph (directed graph) and the correlation relationship graph (undirected graph) respectively to measure the learning bias weight of the learner;

[0070] The specific formula is as follows:

[0071]

[0072] Among them, h s represents the cognitive structure state representation of learner s considering the predecessor-successor relationship and the correlation relationship; represents the cognitive structure state representation of the learner for the predecessor-successor graph; represents the cognitive structure state representation of the learner for the predecessor-successor graph; Readout represents the graph pooling operation; Ψ DRC represents the directed EGAT graph representation learning operation, Ψ URC represents the undirected EGAT graph representation learning operation; h d , e d , A d and h u , e u , A u respectively represent the inputs of the directed and undirected EGAT graph representation learning: node feature vector, edge feature vector, and adjacency matrix; W d and W u respectively represent the learning bias weights of the learner for the predecessor-successor graph and the correlation relationship graph.

[0073] After obtaining the cognitive structure state representation h s of learner s, the cognitive structure state diagnosis result of the learner is obtained through the cognitive diagnosis formula:

[0074] y = Q e ·(h s -h diff )·h dics (8)

[0075] Among them, y represents the diagnostic result of the learner's cognitive structure state.

[0076] The prediction module is used to predict the learner's answer result, and optimize the model in combination with the cross-entropy loss function to improve the diagnostic accuracy and interpretability.

[0077] The prediction module selects the multi-layer perceptron model (Multi-Layer Perceptron, MLP) as the classifier to predict the learner's answer result. This classifier divides the instances into different categories through the input feature vectors, and learns the feature patterns and their internal laws based on the training data, so as to infer the learner's answer result. The advantages of using the classifier for prediction are mainly reflected in the following two aspects: First, it can use supervised learning technology to efficiently train the model; Second, it can effectively process large-scale and high-dimensional feature spaces, thus significantly improving the prediction accuracy.

[0078] The multi-layer perceptron model models the learner's answer result as a binary classification problem, and maps the output to a real value in the interval [0, 1] through the sigmoid activation function, representing the probability that the learner answers correctly. When the predicted value is greater than or equal to 0.5, it is determined that the learner answers correctly; when the predicted value is less than 0.5, it is determined that the learner answers wrongly. This modeling method not only simplifies the problem statement, but also can make full use of the information in the feature space, thus significantly improving the classification performance.

[0079] The specific formula used in the prediction module is:

[0080]

[0081] Among them, represents the predicted value of the model, y represents the real answer result, taking values of 0 or 1, representing the learner's answer error or correct respectively, W represents the learnable weight matrix, and b represents the learnable bias.

[0082] The formula of the cross-entropy loss function in the prediction module is:

[0083]

[0084] Among them, l represents the loss value, which is used to measure the difference between the predicted value and the real value of the model.

[0085] By minimizing the cross-entropy loss function, the parameters of the model will be optimized to improve the prediction accuracy of the model on the training data, so as to more effectively predict the correctness of the learners' answers.

[0086] Therefore, the present invention adopts the above-mentioned cognitive structure diagnosis system based on the edge feature map attention network, which significantly improves the accuracy, real-time performance and interpretability of cognitive diagnosis, provides a more accurate and comprehensive cognitive diagnosis tool for the intelligent education system, and has strong generality and scalability.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cognitive structure diagnosis system based on edge feature graph attention network, characterized in that: include: Embedding module, which is used to map knowledge concepts, relations between concepts, learner states and exercises into low-dimensional vector space to achieve efficient representation of features; The cognitive structure state representation module is used to dynamically model the learner's knowledge state and knowledge structure state; The prediction module is used to predict the learners' answer results and optimize the model in combination with the cross entropy loss function to improve the diagnostic accuracy and interpretability.

2. According to claim 1, a cognitive structure diagnosis system based on edge feature graph attention network is characterized in that: The embedding module outputs the learner's One-Hot code, the predecessor-successor relationship diagram, the correlation relationship diagram, and the One-Hot code of the exercise. The exercise obtains the knowledge dependency vector Q through the embedding module. e , knowledge difficulty vector h diff and the discrimination vector h of the exercise disc .

3. The cognitive structure diagnosis system based on edge feature graph attention network according to claim 2, characterized in that: The cognitive structure state representation module introduces an attention network based on edge feature graph - EGAT to learn the learner's mastery of knowledge state and knowledge structure state in the predecessor-successor relationship graph and the correlation relationship graph respectively. Each EGAT layer consists of a node attention block and an edge attention block.

4. A cognitive structure diagnosis system based on edge feature graph attention network according to claim 3, characterized in that: The attention factor a between node i and node j in the node attention block ij The calculation formula is: Among them, N i represents the set of neighbor nodes of node i, represents the weight vector parameter, represents the edge feature vector between node i and node j, represents the feature vector representation of node i, represents the feature vector representation of node j, || represents the vector concatenation operation, and LeakyReLU is a nonlinear activation function; According to the attention factor a ij , the update formula of node features is: Among them, h j Represents the feature vector representation of node j.

5. A cognitive structure diagnosis system based on edge feature graph attention network according to claim 3, characterized in that: In the edge attention block, the attention factor β between edge p and its neighbor edge q pq The calculation formula is: Among them, N p represents the set of neighbor edges of edge p, represents the weight vector parameter, The eigenvector representation of edge p is, The feature vector representation of edge q is, Represents the node feature vector between edge p and its neighbor edge q; According to the attention factor β pq , the update formula of edge features is: Among them, h q Denotes the feature vector representation of edge q.

6. The cognitive structure diagnosis system based on edge feature graph attention network according to claim 5, characterized in that: The cognitive structure state representation module EGAT layer obtains the final graph representation through the graph pooling method. During the learner's learning process, a weight is learned for the predecessor-successor relationship graph and the correlation relationship graph respectively to measure the learner's learning bias weight; The specific formula is as follows: Among them, h s It represents the learner's cognitive structure state considering predecessor and successor relations and related relations; It represents the learner's representation of the cognitive structure state with predecessor and successor graph; Represents the learner's cognitive structure state representation of the predecessor and successor graph; Readout represents the graph pooling operation; Ψ DRC represents the directed EGAT graph representing the learning operation, Ψ URC represents the undirected EGAT graph representing the learning operation; h d ,e d ,A d and h u ,e u ,A u Respectively represent the inputs of directed and undirected EGAT graph representation learning: node feature vector, edge feature vector and adjacency matrix; W d and W u They respectively represent the learner's learning bias weights for the predecessor and successor graph and the correlation graph.

7. A cognitive structure diagnosis system based on edge feature graph attention network according to claim 6, characterized in that: The cognitive structure state representation module obtains the cognitive structure state diagnosis result of the learner through the cognitive diagnosis formula: y=Q e ·(h s -h diff )·h dics (8) Among them, y represents the diagnostic result of the learner's cognitive structure state.

8. The cognitive structure diagnosis system based on edge feature graph attention network according to claim 1, characterized in that: The prediction module uses a multi-layer perceptron model as a classifier to predict the learner's answer results. The model models the learner's answer results as a binary classification problem, and maps the output to a real value in the interval [0, 1] through a sigmoid activation function, indicating the probability of the learner's correct answer. When the predicted value is greater than or equal to 0.5, the learner is judged to have answered correctly; When the predicted value is less than 0.5, the learner is judged to have answered incorrectly.

9. A cognitive structure diagnosis system based on edge feature graph attention network according to claim 8, characterized in that: The specific formula used in the prediction module is: in, represents the predicted value of the model, y represents the actual answer result, which takes the value of 0 or 1, indicating that the learner's answer is wrong or correct, respectively, W represents the learnable weight matrix, and b represents the learnable bias.

10. A cognitive structure diagnosis system based on edge feature graph attention network according to claim 9, characterized in that: The formula of the cross entropy loss function in the prediction module is: Among them, l represents the loss value, which is used to measure the difference between the predicted value of the model and the true value.

Images