CN-120543004-B - Cognitive diagnosis method and system based on heterogeneous relation graph embedding

CN120543004BCN 120543004 BCN120543004 BCN 120543004BCN-120543004-B

Abstract

The invention relates to a cognitive diagnosis method and a system based on heterogeneous relation graph embedding, which aim to improve the accuracy and efficiency of cognitive diagnosis by combining graph neural network and metric learning. The method comprises four technology implementation stages of constructing a multi-element cognition relation diagram, performing embedding processing, relational perception coding based on a graph neural network, measuring learning and answer mode distinguishing, and combined training and model optimizing. Specifically, the complex interactive relation among students, test questions and knowledge points is finely depicted by constructing a multi-element cognitive relation graph and performing embedding treatment. And carrying out relation sensing coding based on the graph neural network, updating node embedded representation, and fusing different types of relation information. And combining with a metric learning technology, accurately distinguishing correct and incorrect answering behaviors of students. Finally, through the combined training optimization model, the accuracy and the robustness of the cognitive diagnosis are improved, and the method is suitable for the fields of personalized learning evaluation, intelligent education and the like.

Inventors

HUANG TAO
GENG JING
ZUO JIANLIN
ZHANG JINHONG
CHEN YUXIA
YANG HUALI
CAI JIE
MA WENBO
GAO ZHU

Assignees

宁夏师范大学

Dates

Publication Date: 20260512
Application Date: 20250507

Claims (10)

1. The cognitive diagnosis method based on the heterogeneous relation graph embedding is characterized by comprising the following steps of: Step 1, constructing a multi-element cognition relation graph, firstly extracting physical characteristics of students, test questions and knowledge points, generating initial embedding, connecting the three to form a graph structure through different types of relations, dynamically embedding interaction relations among the entities to generate initial relation embedding; the method comprises the steps of 2, taking a multi-element cognition relation graph as input, realizing message propagation and aggregation through a relation-awareness-based graph neural network, and updating entities and relation embedding according to the aggregated messages; Step 3, embedding updated students, test questions and corresponding relations as input, embedding and projecting the students and the test questions to different hyperplanes to obtain projected student and test question characterization, and optimizing the distance between positive and negative sample triples based on measurement learning to enhance the characteristic difference of two interactive relations of correct answer and incorrect answer so as to obtain enhanced characteristics, wherein the hyperplanes comprise a correct answer relation hyperplane and an incorrect answer relation hyperplane; step 4, inputting enhanced features, capturing nonlinear interaction relation between students and test questions through a neural network structure with enhanced attention, and then outputting correct answer probability of the students through probability prediction; And 5, based on the cognitive diagnosis model formed in the step 1-4, jointly training the parameters of the optimization model, iterating the training model until convergence, and finally predicting the response of the students.
2. The cognitive diagnosis method based on heterogeneous relation graph embedding of claim 1, wherein step 1 specifically comprises the steps of extracting student, test questions and knowledge point entity characteristics from student answer data, generating initial embedding representation of an entity, and fusing side information embedding of the entity to enhance semantic information of the entity embedding.
3. The cognitive diagnostic method based on heterogeneous relational graph embedding as set forth in claim 1, wherein: Dynamically embedding 8 kinds of interaction relations among correct/incorrect answers of students, correct/incorrect feedback of test questions and students, association/guiding of test questions and knowledge and similarity/dependence among knowledge: Wherein, the Is a relationship type Is provided with an initial embedded representation of (c), Representing the type of relationship Is used for the feature vector of (a), The type of relationship is indicated and, Is a parameter matrix of a dense layer.
4. The cognitive diagnostic method based on heterogeneous relational graph embedding of claim 1, wherein the message propagation and aggregation in step 2 specifically comprises: In a multi-component cognitive relationship graph On the above, a group of entity interaction relation triples are given , Representing the source node and the source node, The edge of the relationship is represented, Generating a target node message by combining the source node and the relationship edge: Wherein, the Is the target node In the first place An embedded representation of the layer, i.e. the updated feature vector; respectively, are target nodes Source node And relationship edge In the first place An embedded representation of the layer(s), Representing a target node Including and of neighbor nodes of (a) Directly connected nodes and edges; is a trainable weight matrix, and depends on the relationship type ; Is a function that maps relationship types to indexes of the weight matrix; is a function for combining the characteristics of the source node, the target node and the relationship edge to generate a message; is a nonlinear activation function.
5. The method for cognitive diagnosis based on heterogeneous relational graph embedding as set forth in claim 4, wherein the method comprises the following steps of To represent the weight matrix in different directions, the specific calculation mode is as follows: Wherein, the , And A weight matrix respectively representing a forward relation, a reverse relation and a self-connection relation; And Is the reciprocal of the degree matrix used for weight normalization; representing a set of all forward relationships, Representing a set of all inverse relationships, Representing a set of all sub-loop relationships.
6. The method for cognitive diagnosis based on heterogeneous relational graph embedding of claim 5, wherein updating the entity and relational embedding comprises: using a relational-aware attention mechanism to target nodes Source node And relationship edge The embedding representation of (2) carries out correlation calculation, realizes message aggregation, and the specific process of updating entity embedding is as follows: Wherein, the And Respectively representing node embedding and relation embedding linear transformation matrixes; Is the target node In the first place Layer passing relationship edge The characteristics of the aggregated neighbor nodes, Representing source nodes In the first place Layer passing relationship edge Is a feature vector of (1); representation and destination node By relation edge The neighbors of the interaction are selected, Representation and destination node By relation edge A set of connected neighbor nodes that are connected, Representing a target node And source node Through relation between Is used to determine the attention score of (a), Representing a target node And source node By the relation edge between Is included in the normalized attention weights of (2); Representing relationship edges In the first place The characteristic representation of the layer is such that, Is a trainable parameter vector; for relationship edges The embedded update procedure of (1) is as follows: Is a trainable matrix for linear transformation, function For nonlinear activation.
7. The method for cognitive diagnosis based on heterogeneous relational graph embedding as set forth in claim 1, wherein in step 3, a set of answer interaction triples is given , wherein, It is shown that the student has a function of, The test questions are represented by the test questions, Representing the relationship between students and test questions, i.e. correct answer or incorrect answer, setting up the representation of student entity And characterization of test question entities And a representation of the answer relationship between the two The representation relation of the three is obtained on the vector space, namely: Will be And Projection into different hyperplanes, the hyperplane of each relationship is considered as the exclusive space of the relationship, and projection of an entity in space represents its semantics under the relationship, and for a correctly answered relationship, the projection operates as follows: Wherein, the Representing a trainable relationship normal vector, defining a correct answer relationship hyperplane, And Representing student and test question characterization projected under the right answer relation hyperplane.
8. The method for cognitive diagnosis based on heterogeneous relational graph embedding of claim 7, wherein optimizing the distances between the positive and negative sample triples by metric learning comprises: The similarity of the projected student entity and the test question entity under the correct answer relation is measured, and the formula is as follows: Wherein, the Is a function of the distance and, Feature vectors representing the correct answer relations, And similarly, calculating the similarity between the student entity and the test question entity under the wrong answer relation, wherein the process is as follows: Wherein, the Representing student entities Test question entity Similarity measure under wrong answer relation, Representing projected feature vectors of student entities in a wrong response relationship, Feature vectors representing the wrong-answer relationship, The projection characteristic vector of the test question entity under the error answer relation is represented, Trainable weight matrix representing projection of incorrect answer relation for given student-test question interaction triplet Construction of triplets So that two triples are paired positive and negative examples, the positive and negative of the triples pass through a given relation label To determine if Then describe the triplet Is a positive sample, triplet Is a negative sample; The training objectives of metric learning are as follows: Wherein, the Is a super parameter.
9. The method for cognitive diagnosis based on heterogeneous relational graph embedding of claim 8, wherein in step 4, the predictive modeling formula is as follows: Wherein, the The multi-layer perceptron consists of two full-connection layers and an activation function Sigmoid; representing the probability of predicting the student s to answer correctly on the test question q, Represents the probability of predicting the wrong answer of student s on test question q, The result prediction of the student s answering the test question q is shown; The prediction probability of student responses and the loss of actual labels are calculated by adopting a binary cross entropy loss function, and a specific calculation formula is as follows: Wherein, the And showing the result label of the student.
10. A cognitive diagnostic system based on heterogeneous relational graph embedding, comprising: A processor and a memory for storing program instructions, the processor for invoking the stored instructions in the memory to perform the cognitive diagnostic method of any of claims 1-9 based on heterogeneous relational graph embedding.

Description

Cognitive diagnosis method and system based on heterogeneous relation graph embedding Technical Field The invention relates to the technical field of cognitive diagnosis, in particular to a heterogeneous relation graph embedding method based on a graph neural network. Specifically, the invention provides a cognitive diagnosis method based on heterogeneous relation graph embedding, which aims to accurately describe multi-element interaction relations among cognitive entities by constructing a directed multi-element cognitive relation attribute graph and perform joint modeling of cognitive diagnosis tasks by combining metric learning through a graph neural network technology. Background Cognitive diagnostics (Cognitive Diagnosis, CD) are increasingly used in the educational field, particularly at the heart of intelligent educational systems and personalized learning route recommendations. The purpose of cognitive diagnosis is to infer the grasping degree of students on each knowledge point by analyzing multi-modal behavior data such as answer tracks, eye movement data, interaction logs and the like of the students, so that accurate basis is provided for education evaluation. A cognitive diagnostic model (Cognitive Diagnosis Model, CDM) has become an important research direction in the field of educational measurement as a core tool to achieve this goal. Traditional cognitive diagnosis methods, such as probabilistic structural models based on term reaction theory (Item Response Theory, IRT) And deterministic input, noise AND Gate model (DETERMINISTIC INPUTS NOISY "And" Gate, DINA) And the like, mainly infer the cognitive state by analyzing the answering behaviors of students on test questions. An advantage of these traditional models is that they enable the inference of the cognitive state of the student by fitting the probability of the student's answer pattern. However, these approaches have some limitations in dealing with complex cognitive interactions and fine-grained cognitive state modeling. For example, a conventional model generally assumes that a student needs to master all knowledge points covered by a test question to answer the question, and this "all-or-nothing" assumption cannot effectively capture the degree of mastery of the student at different knowledge points, and especially in an actual education scenario, the cognitive state of the student is often a dynamic and multi-level process. Furthermore, CDM based on probabilistic structures typically relies on artificially designed Q-matrices and attribute independence assumptions, which makes them less flexible and adaptable in handling complex, real-world educational data, especially in cases of sparse or incomplete data, where the diagnostic accuracy of the model can be significantly impacted. Along with the rising of deep learning technology, the field of cognitive diagnosis starts to introduce a more complex nonlinear modeling method, in particular to a neural network (Neural Network) model, so that the expression capacity and the application range of the cognitive diagnosis model are greatly improved. For example, deep learning enhanced project response theory (Deep IRT, DIRT) proposed by Cheng et al learns the potential cognitive trait of students and the degree of discrimination of questions by using Deep neural networks (Deep Neural Network, DNN) while introducing long short term memory networks (LSTM) to model the difficulty of the questions. The method can effectively and automatically extract the characteristics from the large-scale data, and improves the performance of the model in an end-to-end training mode. However, when the cognitive diagnosis methods based on the neural network process complex multi-element interactions among cognitive entities (students, test questions and knowledge points), two significant defects still exist, namely, on the one hand, the difficulty of modeling interaction relationship homogenization is overcome, and the existing cognitive diagnosis models based on the neural network mostly use a single relationship parameter to model interaction relationships among the students, the test questions and the knowledge points. This approach, while capable of handling simple correlations, fails to adequately account for the heterogeneous semantics between "positive migration" and "negative feedback". For example, the relevance between the answering and misanswering actions of students on test questions, the feedback of the test questions to the students, and knowledge points often shows asymmetry and dynamic changes, and the information is often not accurately expressed in the existing model. In addition, the multi-relation model in the existing method is usually subjected to relation modeling through simple positive and negative sample segmentation, and the method is easy to cause poor training effect of the model under the condition of low data sparsity or interaction density, and particularly when a high-freque