Metamulti-graph-assisted decoupling graph neurocognition diagnosis method

Abstract

The invention discloses a meta-multi-graph assisted decoupling graph neurocognition diagnosis method, which comprises the following steps of: 1, constructing a student-exercise-concept interaction graph according to answer records of students and exercise concept information labeled by experts, 2, learning a graph neural network aggregation mode and updated representation in a differentiable mode through a one-element multi-graph aggregation module; 3, stripping an interaction graph from student nodes, decoupling a new exercise-concept dependency graph, and updating exercise characterization through an attention mechanism; 4, stripping the dependency graph from the exercise nodes again, decoupling the concept-concept relation graph, and updating the concept representation through the attention mechanism; and 5, after the representations of the three types of nodes are obtained, predicting the mastering degree of the knowledge concept by the students through a novel and highly explanatory diagnosis formula. According to the method, through a decoupling thought and a neural architecture search method, cognitive diagnosis can be carried out on the ability of students, and the robustness of the model can be ensured to a certain extent.

Description

Technical Field

[0001] The present invention relates to personalized cognitive diagnosis of students in the field of smart education. Background Art

[0002] With the rapid development of artificial intelligence technology, more and more disciplines have been combined with artificial intelligence technology and have made significant progress. Smart education is an emerging industry that applies artificial intelligence technology to the education industry. This field analyzes and mines student answer data and diagnoses students' knowledge mastery through deep learning technology to assist subsequent personalized learning of students (such as course recommendations, exercise recommendations, learning routes, etc.).

[0003] As a basic technology in smart education, cognitive diagnosis originated from educational psychology and relied on statistical methods to design an efficient diagnostic formula to diagnose students' knowledge mastery. Later, scholars applied deep learning technology to cognitive diagnosis tasks, using the relationship between text information and knowledge concepts to enrich the representation of students and exercises, and designing more efficient diagnostic formulas to fit the complex interactions between students and exercises. In particular, the graph-based cognitive diagnosis method has achieved the best results by exploring the high-order relationships between students, exercises, and concepts to model them.

[0004] Although graph cognitive diagnosis methods have made significant progress, there are still some problems to be solved. On the one hand, the interaction of student representations is introduced when modeling exercise representations, which may cause some noise. Similar concept representations are also affected by exercise representations and student representations, which is obviously unreasonable. Secondly, previous graph cognitive diagnosis methods only include the representation of the last layer in the diagnosis, and do not consider that the shallow graph neural network representation is also beneficial to the final representation. Therefore, some measures need to be taken to avoid the influence of these factors in order to improve the reliability of diagnosis. Summary of the invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes a meta-multi-graph assisted decoupled graph neural diagnosis method, in order to obtain a more accurate representation of student abilities, and to a certain extent resist the influence of noise data, showing excellent robustness.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0007] The characteristic of the meta-multi-graph-assisted decoupled graph neurocognitive diagnosis method of the present invention is that it is carried out in the following steps:

[0008] Step 1: Build a student node based on the student's answer record and exercise concept information , Exercise Node}, knowledge concept node ;in, Indicates student nodes, Indicates Exercise nodes, Indicates knowledge concept nodes, , , ; M is the number of students, N is the number of exercises, and K is the number of knowledge concepts;

[0009] Step 2: Construct the student-exercise-knowledge concept interaction diagram G I =( , ),in, Indicated by , , A collection of nodes, represents the edge set, and ,in, represents the edge set between student nodes and exercise nodes, Represents the edge set between exercise node and knowledge concept node, Represents the edge set between each knowledge concept node;

[0010] Jordi Student nodes For Exercise Node If you answer Student nodes and Exercise Node There is an edge between them, denoted by ;

[0011] If the jth exercise node The content of the examination includes the kth knowledge concept node , then it means the jth exercise node and the kth knowledge concept node There is an edge between ;

[0012] If the kth knowledge concept node and Knowledge concept node If there is a similarity or successor relationship, it means and There is an edge between ;

[0013] From GI Decoupling exercises-knowledge concept interaction diagram G R =( , ),in, Indicated by and A collection of nodes;

[0014] From G R Decouple the knowledge concept-knowledge concept relationship diagram G D =( , );

[0015] Step 3: Construct a decoupled graph neural network, including: a student node representation update module, an exercise node representation update module, a knowledge concept node representation update module, and a diagnosis module;

[0016] Step 3.1: Student node representation update module Process it and get Student nodes The final representation ;

[0017] Step 3.2: Exercise node representation update module pair Process and get the jth exercise node The final representation ;

[0018] Step 3.3: Knowledge concept node representation update module Processing to obtain the final representation of all knowledge concept nodes ;

[0019] Step 3.4: Diagnose the network according to Student nodes The final representation , the kth knowledge concept node The final representation and the jth exercise node The final representation , for Student nodes The knowledge mastery is predicted to obtain the prediction label ;

[0020] Step 4: Use formula (13) to construct the cross entropy loss function of the decoupled graph neural network :

[0021] (13)

[0022] In formula (13), represents the student-question sample pairs sampled from the answer records, Indicates Student nodes Answer Exercise Node The true label of

[0023] Step 5: Use the Adam optimizer to train the cognitive diagnosis model and minimize the cross entropy loss function To update the network parameters until The training continues until convergence, thus obtaining a trained cognitive diagnosis model, which is used to learn the degree of mastery of different knowledge points by the corresponding students.

[0024] The characteristic of the meta-multi-graph-assisted decoupled graph neurocognitive diagnosis method described in the present invention is that step 3.1 is performed as follows:

[0025] Step 3.1.1: Use formula (1) to get the initial moment Student nodes Characterization , the jth exercise node Characterization and the kth concept node Characterization :

[0026] (1)

[0027] In formula (1), It's G I The i-th student node in The one-hot vector of It's G I The jth exercise node in The one-hot vector of It's G I The jth exercise node in The one-hot vector of Represents G I Middle school students embed the initial matrix, Represents G I The initial matrix is ​​embedded in the exercise. Represents G I The knowledge concept is embedded in the initial matrix; D is a hyperparameter representing the dimension;

[0028] Step 3.1.2, define the variable u, and initialize u=1; and The representations of all nodes at the initial moment constitute the uth supernode, denoted as H (u) ;

[0029] Step 3.1.3: Use formula (3) to get the uth supernode H (u) The Pth supernode H corresponding to the end time (P) The weight threshold of the propagation path between :

[0030] (2)

[0031] In formula (3), Represents the weight value of the propagation path, is a hyperparameter;

[0032] Step 3.1.4: Let the uth supernode H (u) To the Pth supernode H (P) There are six message propagation paths: the propagation path from student nodes to exercise nodes, the propagation path from exercise nodes to student nodes, the propagation path from exercise nodes to knowledge concept nodes, the propagation path from knowledge concept nodes to exercise nodes, the propagation path of empty connections, and the propagation path of self-connections; according to , select more than six message propagation paths The message propagation path is added to the optimal message propagation path set;

[0033] Step 3.1.5: Use formula (2) to get the Pth supernode H (P) :

[0034] (3)

[0035] Formula (2): represents the uth supernode H (u) To the Pth supernode H (P) The optimal message propagation path set, f() represents the GCN aggregation function;

[0036] Step 3.1.6: From H (P) Extract the final representation of all student nodes .

[0037] Further, the step 3.2 is performed as follows:

[0038] Step 3.2.1: Use formula (4) to get the initialization graph G R The jth exercise node in Characterization , the kth concept node Characterization :

[0039] (4)

[0040] In formula (4), It is Figure G R The jth training node in The one-hot vector of It is Figure G R The kth concept node in The one-hot vector of It's G R The initial matrix is ​​embedded in the exercise. It's G R The knowledge concepts are embedded in the initial matrix;

[0041] Step 3.2.2: Define the current layer in the GAT graph neural network as layer and initialize =1; Initialize the -1 layer j-th exercise node Characterization = ; Initialize the -1st layer k-th concept node Characterization = ;

[0042] Step 3.2.3: Use equation (5) and equation (6) to get The jth exercise node in the layer GAT graph neural network Characterization and Layer k-th concept node Characterization , thus getting the jth exercise node The final representation = :

[0043] (5)

[0044] (6)

[0045] In formula (5) and formula (6), Represents the jth exercise node e j A set of adjacent knowledge concept nodes, Represents the kth concept node The set of adjacent exercise nodes, Represents the kth concept node A set of adjacent knowledge concept nodes; Indicates The j-th exercise node e in the layer GAT graph neural network j Its adjacent knowledge concept nodes The attention coefficient matrix between Indicates The kth concept node in the layer GAT graph neural network and its adjacent exercise node e j The attention coefficient between Indicates The kth concept node in the layer GAT graph neural network and its adjacent concept nodes The attention coefficient between them; and:

[0046] (7)

[0047] In formula (7), () represents the connection operation, represents the normalization function, It is -1 layer graph neural network j-th exercise node e j The representation of It is -kth concept node in 1-layer graph neural network representation.

[0048] Further, the step 3.3 is performed as follows:

[0049] Step 3.3.1: Initialize graph G using formula (8) D Concept node representation in :

[0050] (8)

[0051] In formula (8), It's G D The kth concept node in The one-hot vector of It's G D The knowledge concepts are embedded in the initial matrix;

[0052] Step 3.3.2, define G D The current layer in the GAT graph neural network is layer and initialize =1; Initialize the -The kth knowledge concept node in the 1-layer GAT graph neural network Characterization = ; Initialize the -kth concept node in 1-layer GAT graph neural network Characterization = ;

[0053] Step 3.3.3: Use formula (9) to get The kth concept node in the layer GAT graph neural network Characterization , thus obtaining the kth knowledge concept node The final representation , thus obtaining the representation vector matrix of all knowledge concept nodes :

[0054] (9)

[0055] In formula (9), Represents the kth concept node A set of adjacent knowledge concept nodes, Represents the kth concept node Adjacent knowledge concept nodes The attention coefficient between .

[0056] Further, the step 3.4 is performed as follows:

[0057] Step 3.4.1. Use formula (10) to calculate the Student nodes Capabilities for all knowledge concept nodes :

[0058] (10)

[0059] In formula (10), Fully connected layers representing student representation and knowledge concept representation;

[0060] Step 3.4.2: Use formula (11) to calculate the jth exercise node Difficulty value for all knowledge concept nodes :

[0061] (11)

[0062] In formula (11), Fully connected layers representing concept representation and knowledge concept representation;

[0063] Step 3.4.3: Use formula (12) to predict students Answer No. Exercise Node The predicted label :

[0064] (12)

[0065] In formula (12), is the sigmoid activation function, is the similarity function, Indicates Exercise Node The knowledge concept node number involved, Indicates a counting symbol; It's G I Middle Exercise Node The one-hot vector matrix of the k-th knowledge concept node involved.

[0066] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the decoupled graph neurocognitive diagnosis method, and the processor is configured to execute the program stored in the memory.

[0067] The present invention provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program, which is characterized in that when the computer program is executed by a processor, the steps of the decoupling graph neurocognitive diagnosis method are executed.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] The present invention improves the diagnostic robustness of student ability representation by introducing a decoupled graph neurocognitive diagnostic model. First, a student-exercise-concept interaction graph is constructed based on the student's answer record and the exercise concept information annotated by the expert. Then, in the student representation update module, the optimal message propagation path is determined in a differentiable way and the updated student node representation is obtained. In order to eliminate the noise influence brought by the student to the exercise representation, the student node is separated from the interaction graph to decouple the new exercise-concept dependency graph, and the neighbor information of the exercise node is aggregated through the attention mechanism to update the exercise representation. Furthermore, in order to eliminate the noise influence brought by the exercise to the concept representation, the exercise node is again separated from the dependency graph to decouple the concept-concept relationship graph, and the neighbor information of the concept node is aggregated through the attention mechanism to update the concept representation. Finally, after obtaining the representation of the three types of nodes, a novel and highly explanatory diagnostic formula is used to predict the student's mastery of knowledge concepts, thereby obtaining a more accurate representation of the student's ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The figure is an overall flow chart of the method of the present invention. DETAILED DESCRIPTION

[0071] In this embodiment, a meta-multi-graph assisted decoupled graph neurocognitive diagnosis method is to use the decoupling idea to independently model the student, exercise, and concept representations in three sub-graphs (student-exercise-concept interaction graph, exercise-concept dependency graph, and concept-concept relationship graph), and at the same time introduce a novel meta-multi-graph aggregator in the modeling of student representation to dynamically search for propagation paths, and finally incorporate it into a highly explanatory diagnostic formula for diagnosis. Specifically, Figure 1 As shown, it is carried out in the following steps:

[0072] Step 1: Construct a student set based on the student's answer records and the concept-related information of the exercises annotated by experts ,in, Indicates students, N is the number of students; the exercise set ,in, represents the jth exercise, M is the number of exercises; knowledge concept set ,in, represents the kth knowledge concept, K is the number of knowledge concepts; the exercise-concept relationship matrix Q=(Q jk ∈{0,1}) k×k , where Q jk =1 indicates the kth knowledge concept In the jth exercise In, on the contrary Q jk =0.

[0073] Step 2: Graph Neural Network Construction:

[0074] Constructing the student-exercise-knowledge concept interaction diagram G I =( , ),in, Indicated by , , A collection of nodes, represents the edge set, and ,in, represents the edge set between student node and exercise node, indicating that Represents the edge set between exercise node and knowledge concept node, Represents the edge set between each knowledge concept node;

[0075] Jordi Student nodes For Exercise Node If you answer Student nodes and Exercise Node There is an edge between them, denoted by ;

[0076] If the jth exercise node The content of the examination includes the kth knowledge concept node , then it means the jth exercise node and the kth knowledge concept node There is an edge between ;

[0077] If the kth knowledge concept node and Knowledge concept node If there is a similarity or successor relationship, it means and There is an edge between ;

[0078] From G I Decoupling exercises-knowledge concept interaction diagram G R =( , ), eliminating the influence of student nodes on the update of exercise representation, where Indicated by and A collection of nodes;

[0079] Further in the exercise - concept dependency graph G R The influence of exercise nodes on concept representation is separated and the concept-concept dependency graph G is decoupled. D =( , );

[0080] Step 3: Construct a decoupled graph neural network, including: student node representation update module, exercise node representation update module, knowledge concept node representation update module, and diagnosis module:

[0081] Step 3.1: Student node representation update module Process it and get Student nodes The final representation :

[0082] Step 3.1.1: Use formula (1) to get the initial moment Student nodes Characterization , the jth exercise node Characterization and the kth concept node Characterization :

[0083] (1)

[0084] In formula (1), It's G I The i-th student node in The one-hot vector of It's G I The jth exercise node in The one-hot vector of It's G I The jth exercise node in The one-hot vector of Represents G I Middle school students embed the initial matrix, Represents G I The initial matrix is ​​embedded in the exercise. Represents G I The knowledge concepts in it are embedded in the initial matrix; D is a hyperparameter representing the dimension.

[0085] Step 3.1.2, define the variable u, and initialize u=1; and The representations of all nodes at the initial moment constitute the uth supernode, denoted as H (u) ;

[0086] Step 3.1.3: In order to obtain the optimal propagation mode and make the final student representation optimal, weights are assigned to the propagation paths between supernodes, and propagation paths greater than the threshold are selected for message propagation. Formula (3) is used to obtain the uth supernode H (u) The Pth supernode H corresponding to the end time (P) The weight threshold of the propagation path between :

[0087] (2)

[0088] In formula (3), Represents the weight value of the propagation path, is a hyperparameter.

[0089] Step 3.1.4: Let the uth supernode H (u) To the Pth supernode H (P) There are six message propagation paths: the propagation path from student nodes to exercise nodes, the propagation path from exercise nodes to student nodes, the propagation path from exercise nodes to knowledge concept nodes, the propagation path from knowledge concept nodes to exercise nodes, the propagation path of empty connections, and the propagation path of self-connections; according to , select more than six message propagation paths The message propagation path is added to the optimal message propagation path set;

[0090] Step 3.1.5: Use formula (2) to get the Pth supernode H (P) :

[0091] (3)

[0092] Formula (2): represents the uth supernode H (u) To the Pth supernode H (P) The optimal message propagation path set, f() represents the GCN aggregation function;

[0093] Step 3.1.6: From H (P) Extract the final representation of all student nodes .

[0094] Step 3.2: Exercise node representation update module pair Processing is performed to obtain the final representation of all exercise nodes:

[0095] Step 3.2.1: Reinitialize the problem and concept representations so that they can be updated and aggregated independently in the decoupled problem-concept dependency graph. Use formula (4) to get the initialization graph G: R The jth exercise node in Characterization , the kth concept node Characterization :

[0096] (4)

[0097] In formula (4), It is Figure G R The jth training node in The one-hot vector of It is Figure G R The kth concept node in The one-hot vector of It's G R The initial matrix is ​​embedded in the exercise. It's G R The knowledge concepts are embedded in the initial matrix;

[0098] Step 3.2.2: Define the current layer in the GAT graph neural network as layer and initialize =1; Initialize the -1 layer j-th exercise node Characterization = ; Initialize the -1st layer k-th concept node Characterization = .

[0099] Step 3.2.3: Use equation (5) and equation (6) to get The jth exercise node in the layer GAT graph neural network Characterization and Layer k-th concept node Characterization , thus getting the jth exercise node The final representation = :

[0100] (5)

[0101] (6)

[0102] In formula (5) and formula (6), Represents the jth exercise node e j A set of adjacent knowledge concept nodes, Represents the kth concept node The set of adjacent exercise nodes, Represents the kth concept node A set of adjacent knowledge concept nodes; Indicates The j-th exercise node e in the layer GAT graph neural network j Its adjacent knowledge concept nodes The attention coefficient matrix between Indicates The kth concept node in the layer GAT graph neural network and its adjacent exercise node e j The attention coefficient between Indicates The kth concept node in the layer GAT graph neural network and its adjacent concept nodes The attention coefficient between them; and:

[0103] (7)

[0104] In formula (7), () represents the connection operation, represents the normalization function, It is -1 layer graph neural network j-th exercise node e j The representation of It is -kth concept node in 1-layer graph neural network representation.

[0105] Step 3.3: Knowledge concept node representation update module Processing to obtain the final representation of all knowledge concept nodes :

[0106] Step 3.3.1: Reinitialize the concept representation so that it can be updated independently in the concept-concept relationship graph. Use formula (8) to initialize the graph G. D Concept node representation in :

[0107] (8)

[0108] In formula (8), It's G D The kth concept node in The one-hot vector of It's G D The knowledge concepts are embedded in the initial matrix;

[0109] Step 3.3.2, define G D The current layer in the GAT graph neural network is layer and initialize =1; Initialize the -The kth knowledge concept node in the 1-layer GAT graph neural network Characterization = ; Initialize the -kth concept node in 1-layer GAT graph neural network Characterization = ;

[0110] Step 3.3.3: Use formula (9) to get The kth concept node in the layer GAT graph neural network Characterization , thus obtaining the kth knowledge concept node The final representation , thus obtaining the representation vector matrix of all knowledge concept nodes :

[0111] (9)

[0112] In formula (9), Represents the kth concept node A set of adjacent knowledge concept nodes, Represents the kth concept node Adjacent knowledge concept nodes The attention coefficient between .

[0113] Step 3.4: Diagnose the network according to Student nodes The final representation , the kth knowledge concept node The final representation and the jth exercise node The final representation , for Student nodes The knowledge mastery is predicted to obtain the prediction label ;

[0114] Step 3.4.1. Use formula (10) to calculate the Student nodes Capabilities for all knowledge concept nodes :

[0115] (10)

[0116] In formula (10), Fully connected layers representing student representation and knowledge concept representation;

[0117] Step 3.4.2: Use formula (11) to calculate the jth exercise node Difficulty value for all knowledge concept nodes :

[0118] (11)

[0119] In formula (11), Fully connected layers representing concept representation and knowledge concept representation.

[0120] Step 3.4.3: Use formula (12) to predict students Answer No. Exercise Node The predicted label ,like >0.5, the student is considered to have done the exercise correctly, otherwise it is considered to have done it wrong:

[0121] (12)

[0122] In formula (12), is the sigmoid activation function, is the similarity function, Indicates Exercise Node The knowledge concept node number involved, Indicates a counting symbol; It's G I Middle Exercise Node The one-hot vector matrix of the k-th knowledge concept node involved.

[0123] Step 4: Use formula (13) to construct the cross entropy loss function of the decoupled graph neural network :

[0124] (13)

[0125] In formula (13), represents the student-question sample pairs sampled from the answer records, Indicates Student nodes Answer Exercise Node The true label of

[0126] Step 5: Use the Adam optimizer to train the cognitive diagnosis model and minimize the cross entropy loss function To update the network parameters until The training continues until convergence, thus obtaining a trained cognitive diagnosis model, which is used to learn the degree of mastery of different knowledge points by the corresponding students.

[0127] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0128] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0129] In summary, the main technical idea of ​​the present invention is to use the idea of ​​decoupling to decouple the constructed student-exercise-concept interaction graph into two subgraphs, remove the influence of irrelevant nodes, and learn the representations of students, exercises, and concepts respectively through a fruitful meta-multi-graph aggregator and two unique graph attention networks, and feed them into a novel and explanatory diagnostic formula for knowledge mastery diagnosis. This method aims to provide students with more accurate and personalized learning suggestions and can model more accurately in a real environment full of noise interference.

[0130] Example:

[0131] In order to verify the effectiveness of this method, the present invention uses two datasets in smart education research: Assistments2009 and Math for verification. In each dataset, the answer records of students with less than 15 answer records are deleted. For the processed Assistments2009 dataset, 4163 students, 17746 exercises and 123 knowledge concepts can be obtained; for the processed Math dataset, 1976 students, 1686 exercises and 61 knowledge concepts can be obtained.

[0132] For the task of student knowledge mastery, the present invention uses three types of indicators for evaluation, namely, accuracy ACC, root mean square error RMSE and area under the receiver operating characteristic curve AUC. The higher the ACC and AUC, the better the model performance; the lower the RMSE, the better the model performance.

[0133] In this invention, five different diagnostic models, DINA, MIRT, NCD, ECD and RCD, are used as baselines for comparison, among which RCD is also a cognitive diagnostic framework based on graph neural network. As shown in Table 1:

[0134] Table 1 Experimental results

[0135]

[0136] Specifically, firstly, on all data sets, DisenGCD has a good improvement over the baseline, which shows that the graph-based method can effectively improve the diagnosis effect; secondly, DisenGCD also has an improvement over another graph-based diagnostic method RCD, which shows that the decoupling method of the present invention has a certain effect in anti-noise and can improve the diagnostic accuracy.

Claims

1. A meta-graph-assisted decoupled graph neurocognitive diagnosis method, characterized in that: The steps are as follows: Step 1: Build a student node based on the student's answer record and exercise concept information , Exercise Node }, knowledge concept node ;in, Indicates student nodes, Indicates Exercise nodes, Indicates knowledge concept nodes, , , ; M is the number of students, N is the number of exercises, and K is the number of knowledge concepts; Step 2: Construct the student-exercise-knowledge concept interaction diagram G I =( , ),in, Indicated by , , A collection of nodes, represents the edge set, and ,in, represents the edge set between student nodes and exercise nodes, Represents the edge set between exercise node and knowledge concept node, Represents the edge set between each knowledge concept node; Jordi Student nodes For Exercise Node If you answer Student nodes and Exercise Node There is an edge between them, denoted by ; If the jth exercise node The content of the examination includes the kth knowledge concept node , then it means the jth exercise node and the kth knowledge concept node There is an edge between ; If the kth knowledge concept node and Knowledge concept node If there is a similarity or successor relationship, it means and There is an edge between ; From G I Decoupling exercises-knowledge concept interaction diagram G R =( , ),in, Indicated by and A collection of nodes; From G R Decouple the knowledge concept-knowledge concept relationship diagram G D =( , ); Step 3: Construct a decoupled graph neural network, including: a student node representation update module, an exercise node representation update module, a knowledge concept node representation update module, and a diagnosis module; Step 3.1: Student node representation update module Process it and get Student nodes The final representation ; Step 3.2: Exercise node representation update module pair Process and get the jth exercise node The final representation ; Step 3.3: Knowledge concept node representation update module Processing to obtain the final representation of all knowledge concept nodes ; Step 3.4: Diagnose the network according to Student nodes The final representation , the kth knowledge concept node The final representation and the jth exercise node The final representation , for Student nodes The knowledge mastery is predicted to obtain the prediction label ; Step 4: Use formula (13) to construct the cross entropy loss function of the decoupled graph neural network : (13) In formula (13), represents the student-question sample pairs sampled from the answer records, Indicates Student nodes Answer Exercise Node The true label of Step 5: Use the Adam optimizer to train the cognitive diagnosis model and minimize the cross entropy loss function To update the network parameters until The training continues until convergence, thus obtaining a trained cognitive diagnosis model, which is used to learn the degree of mastery of different knowledge points by the corresponding students.

2. The method of neurocognitive diagnosis based on decoupled graph assisted by meta-graphs according to claim 1, characterized in that: The step 3.1 is performed as follows: Step 3.1.1: Use formula (1) to get the initial moment Student nodes Characterization , the jth exercise node Characterization and the kth concept node Characterization : (1) In formula (1), It's G I The i-th student node in The one-hot vector of It's G I The jth exercise node in The one-hot vector of It's G I The jth exercise node in The one-hot vector of Represents G I Middle school students embed the initial matrix, Represents G I The initial matrix is ​​embedded in the exercise. Represents G I The knowledge concepts are embedded in the initial matrix; D is a hyperparameter representing the dimension; Step 3.1.2, define the variable u, and initialize u=1; and The representations of all nodes at the initial moment constitute the uth supernode, denoted as H (u) ; Step 3.1.3: Use formula (3) to get the uth supernode H (u) The Pth supernode H corresponding to the end time (P) The weight threshold of the propagation path between : (2) In formula (3), Represents the weight value of the propagation path, is a hyperparameter; Step 3.1.4: Let the uth supernode H (u) To the Pth supernode H (P) There are six message propagation paths: the propagation path from student nodes to exercise nodes, the propagation path from exercise nodes to student nodes, the propagation path from exercise nodes to knowledge concept nodes, the propagation path from knowledge concept nodes to exercise nodes, the propagation path of empty connections, and the propagation path of self-connections; according to , select more than six message propagation paths The message propagation path is added to the optimal message propagation path set; Step 3.1.5: Use formula (2) to get the Pth supernode H (P) : (3) Formula (2): represents the uth supernode H (u) To the Pth supernode H (P) The optimal message propagation path set, f() represents the GCN aggregation function; Step 3.1.6: From H (P) Extract the final representation of all student nodes .

3. The method of neurocognitive diagnosis based on decoupled graph assisted by meta-graphs according to claim 2, characterized in that: The step 3.2 is performed as follows: Step 3.2.1: Use formula (4) to get the initialization graph G R The jth exercise node in Characterization , the kth concept node Characterization : (4) In formula (4), It is Figure G R The jth training node in The one-hot vector of It is Figure G R The kth concept node in The one-hot vector of It's G R The initial matrix is ​​embedded in the exercise. It's G R The knowledge concepts are embedded in the initial matrix; Step 3.2.2: Define the current layer in the GAT graph neural network as layer and initialize =1; Initialize the -1 layer j-th exercise node Characterization = ; Initialize the -1st layer k-th concept node Characterization = ; Step 3.2.3: Use equation (5) and equation (6) to get The jth exercise node in the layer GAT graph neural network Characterization and Layer k-th concept node Characterization , thus getting the jth exercise node The final representation = : (5) (6) In formula (5) and formula (6), Represents the jth exercise node e j A set of adjacent knowledge concept nodes, Represents the kth concept node The set of adjacent exercise nodes, Represents the kth concept node A set of adjacent knowledge concept nodes; Indicates The jth exercise node e in the layer GAT graph neural network j Its adjacent knowledge concept nodes The attention coefficient matrix between Indicates The kth concept node in the layer GAT graph neural network and its adjacent exercise node e j The attention coefficient between Indicates The kth concept node in the layer GAT graph neural network and its adjacent concept nodes The attention coefficient between them; and: (7) In formula (7), () represents the connection operation, represents the normalization function, It is -1 layer graph neural network j-th exercise node e j The representation of It is -kth concept node in 1-layer graph neural network representation.

4. The method of neurocognitive diagnosis assisted by decoupled graphs according to claim 3, characterized in that: The step 3.3 is performed as follows: Step 3.3.1: Initialize graph G using formula (8) D Concept node representation in : (8) In formula (8), It's G D The kth concept node in The one-hot vector of It's G D The knowledge concepts are embedded in the initial matrix; Step 3.3.2, define G D The current layer in the GAT graph neural network is layer and initialize =1; Initialize the -kth knowledge concept node in 1-layer GAT graph neural network Characterization = ; Initialize the -kth concept node in 1-layer GAT graph neural network Characterization = ; Step 3.3.3: Use formula (9) to get The kth concept node in the layer GAT graph neural network Characterization , thus obtaining the kth knowledge concept node The final representation , thus obtaining the representation vector matrix of all knowledge concept nodes : (9) In formula (9), Represents the kth concept node A set of adjacent knowledge concept nodes, Represents the kth concept node Adjacent knowledge concept nodes The attention coefficient between .

5. The method of neurocognitive diagnosis assisted by decoupled graphs according to claim 4, characterized in that: The step 3.4 is performed as follows: Step 3.4.

1. Use formula (10) to calculate the Student nodes Capabilities for all knowledge concept nodes : (10) In formula (10), Fully connected layers representing student representation and knowledge concept representation; Step 3.4.2: Use formula (11) to calculate the jth exercise node Difficulty value for all knowledge concept nodes : (11) In formula (11), Fully connected layers representing concept representation and knowledge concept representation; Step 3.4.3: Use formula (12) to predict students Answer No. Exercise Node The predicted label : (12) In formula (12), is the sigmoid activation function, is the similarity function, Indicates Exercise Node The knowledge concept node number involved, Indicates a counting symbol; It's G I Middle Exercise Node The one-hot vector matrix of the k-th knowledge concept node involved.

6. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the decoupled graph neurocognitive diagnosis method described in any one of claims 1 to 5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the decoupled graph neurocognitive diagnosis method described in any one of claims 1 to 5 are executed.

Images