Knowledge state prediction method and device based on modified Q matrix and attention mechanism
By constructing a knowledge point prerequisite relationship diagram and practice similarity diagram based on the method of correcting the Q matrix and attention mechanism, combining the graph attention network, the problem of unmined subjectivity and complex interaction relationships of Q matrix in the existing technology is solved, and the problem of unmined Q matrix subjectivity and complex interaction relationships is improved. The accuracy and efficiency of students' knowledge status prediction.
Patent Information
- Application Number
- CN202411302761.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-19
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-09-19
AI Technical Summary
The existing cognitive diagnostic technology relies on expert-noted Q matrix with subjective tendencies and one-sidedness, and cannot effectively explore the complex nonlinear interaction between students and exercises, resulting in inaccurate prediction results.
Using the knowledge state prediction method based on the corrected Q matrix and attention mechanism, we construct the expert Q matrix by collecting student answers, digging out the knowledge points prerequisites, establishing a knowledge point prerequisite relationship diagram, and combining the exercise similarity to build the goal exercise-knowledge point relationship Figure, input the graph attention network for training, obtain the hypergraph signal, and finally predict the student's knowledge status.
It improves the accuracy of knowledge state prediction, overcomes the problems of expert Q matrix error and manual consumption in traditional methods, enhances the distinction between practice similarity and knowledge point difficulty, and achieves more accurate diagnosis of students' knowledge state.
Smart Images

Figure CN118821927B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of cognitive diagnosis technology, and in particular to a knowledge state prediction method and device based on a modified Q matrix and an attention mechanism. Background Art
[0002] With the development of big data and artificial intelligence technology, the reform of educational informatization is constantly advancing. Online education has become an important teaching model, and many intelligent education systems have emerged at home and abroad. One of the basic tasks of intelligent education is cognitive diagnosis, and the innovation of this technology and method plays an important role in the development of intelligent education.
[0003] In 1952, Lord, an American psychometrician, first proposed the probability-based Item Response Theory (IRT), also known as the Item Characteristic Curve Theory. IRT believes that there is a functional relationship between the subject's potential ability and his response to a specific item, and this function is called the Item Response Function. IRT starts from the inside of the test in a microscopic way, using mathematical modeling and modern statistical methods to obtain the real connection between the subject's potential ability and the test items. Each test item has its own item characteristic curve, which reflects the probability that each subject with a specific potential ability can correctly answer the item.
[0004] Compared with the item response theory that only focuses on a single ability dimension, the multidimensional item response ranking (MIRT) proposed by Bock and Aitkin in 1981 is a continuous multidimensional cognitive diagnostic model with a similar logic item response function. It is extended from the one-dimensional item response theory, models the characteristics of test items in different ability dimensions, and establishes a direct connection between item response theory and factor analysis. Among them, factor analysis is a statistical method used to explore and model the potential relationship between variables. It simplifies the data structure by summarizing the observed variables into potential factors.
[0005] In 1983, based on the item response theory, Japanese-American scholar Tatsuoka proposed the Rule Space Model (RSM), which is a cognitive diagnosis model based on pattern recognition and statistical classification technology. It combines functional analysis, multivariate statistics, Boolean operations and other methods to effectively diagnose the cognitive structure and cognitive process of the subjects.
[0006] In 2001, Junker and Sijtsma proposed the Deterministic Inputs, Noisy "and" Gate Model (DINA), which is suitable for cognitive diagnosis of binary scored item tests and is one of the most widely used cognitive diagnosis models.
[0007] The measurement method of classical measurement theory only evaluates students' grades and cannot deeply explore students' cognitive states. Traditional cognitive diagnosis methods are based on artificially designed functions to mine one-dimensional linear interactions in students' answering process, but cannot mine complex nonlinear interactions between students and exercises. At the same time, traditional methods require a lot of manpower, energy and time, with high technical costs and low efficiency. Driven by big data, deep neural network models can mine students' potential characteristics from their learning behavior data, providing a prerequisite for implementing teaching students in accordance with their aptitude.
[0008] In 2020, Wang et al. proposed the Neural Cognitive Diagnosis Model (NeuralCDM), which uses deep neural networks to learn higher-order linear functions based on item response theory and multidimensional item response theory, and constructs a cognitive diagnosis model based on artificial functions using deep neural network technology to simulate the human brain's answering process, and concretize abstract information such as exercise difficulty, discrimination, and students' proficiency in answering questions.
[0009] However, this method relies on the Q matrix labeled by experts, and the process of expert labeling of the Q matrix is prone to subjective tendencies and one-sided defects, and ignores the relationship between exercises, which leads to inaccurate prediction results. Summary of the invention
[0010] The present invention provides a knowledge state prediction method and system based on a modified Q matrix and an attention mechanism, which are used to solve or at least partially solve the technical problem in the prior art that the prediction results obtained are not accurate enough.
[0011] In order to solve the above technical problems, the technical solution adopted by the present invention specifically includes:
[0012] The first aspect discloses a knowledge state prediction method based on a modified Q matrix and an attention mechanism, comprising:
[0013] Collect students' answer data, and construct an expert Q matrix based on the collected answer data to express the relationship between exercises and knowledge points through the knowledge and experience of domain experts;
[0014] The initial exercise-knowledge point relationship diagram is generated according to the expert Q matrix, and the prerequisites between knowledge points are mined from the students' historical answer data to establish a knowledge point prerequisite relationship diagram. Then, the initial exercise-knowledge point relationship diagram is merged with the knowledge point prerequisite relationship diagram to obtain an enhanced exercise-knowledge point relationship diagram. The enhanced exercise-knowledge point relationship diagram is then converted into a matrix form to obtain a revised Q matrix, in which the knowledge point prerequisite relationship diagram is used to represent the prerequisite dependency conditions between knowledge points. If a student learns a knowledge point, You can continue to learn another knowledge point ,but and There are prerequisite dependencies between them;
[0015] Calculate the similarity between different exercises, use the similarity between different exercises as the correlation between different exercises, and construct a target exercise-knowledge point relationship graph based on the modified Q matrix and the correlation between different exercises; input the target exercise-knowledge point relationship graph into the graph attention network for training to obtain a hypergraph signal;
[0016] The student's one-hot encoding, the exercise's one-hot encoding, and the hypergraph signal are embedded, and the student's knowledge status is finally obtained through information aggregation and updating.
[0017] In one embodiment, the expert Q matrix consists of a row vector and a column vector, where the row vector represents the exercise number, and the column vector represents the attribute of the knowledge point involved in the exercise. The element in the mth row and nth column of the Q matrix is Indicates whether the mth exercise question involves the nth knowledge point. If so, it is 1; if not, it is 0.
[0018] In one implementation, the prerequisites between knowledge points are mined from students' historical answer data to establish a knowledge point prerequisite relationship diagram, including:
[0019] Analyze students' historical answer data and infer the influence of one knowledge point on another from the order in which learners master skills;
[0020] Use preset evaluation indicators to measure the degree of association between knowledge points;
[0021] A knowledge point prerequisite relationship diagram is constructed based on the impact of one knowledge point on another knowledge point and the degree of association between knowledge points.
[0022] In one embodiment, the method further includes setting a threshold to control the sparsity of the relationship between knowledge points in the knowledge point prerequisite relationship graph, specifically:
[0023]
[0024] represents the i-th row and j-th column element of the knowledge point prerequisite relationship matrix, represents the j-th row and i-th column element of the knowledge point prerequisite relationship matrix. The knowledge point prerequisite relationship matrix corresponds to the knowledge point prerequisite relationship graph. is the threshold used to control the sparsity of the relationship between knowledge points in the knowledge graph.
[0025] In one implementation, the modified Q matrix is calculated based on the expert Q matrix and the knowledge point prerequisite relationship matrix, and the calculation method is:
[0026]
[0027] in, represents the knowledge point prerequisite relationship matrix, represents the modified Q matrix, the number of rows of the matrix is the number of knowledge points, and the number of columns is the number of exercises. Represents the element in the i-th row and j-th column of the modified Q matrix.
[0028] In one embodiment, the similarity between different exercises is calculated as follows:
[0029]
[0030] in, Indicates practice and The similarity between Indicates practice exist When answering before, the similarity between the two answers, practise exist When answering before, the two answers were similar.
[0031] In one implementation, the student's one-hot code, the exercise's one-hot code, and the hypergraph signal are embedded, and the student's knowledge state is finally obtained through information aggregation and updating, including:
[0032] One-hot encoding of students Embedded, multiplied by the trained A matrix to obtain the student's proficiency vector , where the proficiency vector represents the students’ mastery of the knowledge points;
[0033] One-hot encoding of the exercise Embedding, multiplying with the trained D matrix to obtain the knowledge point discrimination vector , multiplied by the modified Q matrix to obtain the knowledge point relevance vector , where the knowledge point discrimination vector represents the ability of the exercise to distinguish students of different levels, and the knowledge point relevance vector represents the examination of the exercise on the knowledge point;
[0034] Hypergraph Signal X One-hot encoding with practice Embedding fusion to obtain knowledge point difficulty vector , where the knowledge point difficulty vector represents the students’ mastery of the knowledge point;
[0035] The student's proficiency vector, knowledge point discrimination vector, knowledge point relevance vector and knowledge point difficulty vector are embedded, and then information aggregation and update are performed to obtain the student's predicted score as the prediction result of the student's knowledge status.
[0036] Based on the same inventive concept, the second aspect of the present invention discloses a knowledge state prediction device based on a modified Q matrix and an attention mechanism, comprising:
[0037] Expert Q matrix acquisition module, used to collect students' answer data, and build an expert Q matrix to express the relationship between exercises and knowledge points based on the collected answer data through the knowledge and experience of domain experts; build the Q matrix through the knowledge and experience of domain experts, and experts manually fill in the Q matrix based on their understanding of each question and each skill
[0038] The modified Q matrix acquisition module is used to generate an initial exercise-knowledge point relationship diagram based on the expert Q matrix, mine the prerequisites between knowledge points from the students' historical answer data, establish a knowledge point prerequisite relationship diagram, and then fuse the initial exercise-knowledge point relationship diagram with the knowledge point prerequisite relationship diagram to obtain an enhanced exercise-knowledge point relationship diagram, and then convert the enhanced exercise-knowledge point relationship diagram into a matrix form to obtain a modified Q matrix, where the knowledge point prerequisite relationship diagram is used to represent the prerequisite dependency conditions between knowledge points. If a student learns a knowledge point You can continue to learn another knowledge point ,but and There are prerequisite dependencies between them;
[0039] The module for constructing the exercise-knowledge point relationship graph is used to calculate the similarity between different exercises, and use the similarity between different exercises as the correlation between different exercises. Based on the modified Q matrix and the correlation between different exercises, the target exercise-knowledge point relationship graph is constructed; the target exercise-knowledge point relationship graph is input into the graph attention network for training to obtain a hypergraph signal;
[0040] The knowledge status prediction module is used to embed the student's one-hot encoding, the exercise's one-hot encoding, and the hypergraph signal, and finally obtain the student's knowledge status through information aggregation and updating.
[0041] Based on the same inventive concept, the third aspect of the present invention provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the knowledge state prediction method based on the modified Q matrix and attention mechanism described in the first aspect is implemented.
[0042] Based on the same inventive concept, the fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and run on the processor. When the processor executes the program, it implements the knowledge state prediction method based on the modified Q matrix and attention mechanism described in the first aspect.
[0043] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:
[0044] The present invention proposes a knowledge state prediction method based on a modified Q matrix and a hypergraph attention mechanism. First, the students' answer data is collected through a website, and an expert Q matrix for expressing the association between exercises and knowledge points is obtained. Then, the exercise-knowledge point relationship diagram is fused with the knowledge point prerequisite relationship diagram, and finally a modified Q matrix is obtained. Among them, the "exercise-knowledge point" relationship diagram is constructed, including mining the association relationship matrix between different exercises, and obtaining the relationship diagram by calculating the similarity between different exercises. Finally, the one-hot encoding of the exercise and the one-hot encoding of the student are embedded. The multidimensional item response theory model is used for reference in the embedding process, and the parameters such as the knowledge point proficiency vector, the knowledge point relevance vector, the knowledge point difficulty vector, and the test question discrimination value are respectively mapped to the formula, and the parameters are fused and calculated. Thereby predicting the student's score and analyzing the student's knowledge state. The present invention uses a modified Q matrix based on the knowledge point prerequisite dependency diagram to diagnose the learner's knowledge points, which overcomes the problem that the expert Q matrix in the traditional cognitive diagnosis model is prone to errors and the manual consumption of annotation is too large. In order to improve the difficulty differentiation of knowledge points, the present invention proposes an exercise association measurement model (exercise-knowledge point relationship graph construction module) based on the hypergraph attention mechanism, which obtains more accurate exercise similarity. The difficulty differentiation of knowledge points is improved by improving the exercise similarity. Finally, in the prediction stage, the present invention introduces Transformer into the item response theory for information aggregation and update, predicts the optimal knowledge point proficiency, and finally obtains the student's knowledge status, thereby improving the accuracy of knowledge status prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0046] Figure 1 A schematic diagram for generating an enhanced exercise-knowledge point relationship diagram in an embodiment of the present invention;
[0047] Figure 2 It is an exercise-knowledge point relationship diagram in an embodiment of the present invention;
[0048] Figure 3 This is a schematic diagram of the encoding embedding stage in an embodiment of the present invention;
[0049] Figure 4 This is a schematic diagram of the information aggregation and update stage in an embodiment of the present invention;
[0050] Figure 5 This is an overall framework diagram of the knowledge state prediction method based on the modified Q matrix and hypergraph attention mechanism in an embodiment of the present invention. DETAILED DESCRIPTION
[0051] The present invention proposes a method and device for predicting the knowledge status of students based on a modified Q matrix and a hypergraph attention mechanism in order to improve the accuracy of predictive diagnosis of the knowledge status of students.
[0052] Based on the neurocognitive diagnosis model, this method improves the problem of insufficient accuracy of expert-annotated Q matrix in student factors and neglect of the correlation between exercises in exercise factors, and proposes a solution to correct the Q matrix and construct an "exercise-knowledge point" relationship graph containing exercise similarity, which effectively improves the diagnostic accuracy of students' knowledge status level. The characteristics are as follows: first, for student factors, the model mines the prerequisites between knowledge points from the learners' historical answer data and establishes a knowledge point relationship graph. The corrected Q matrix is obtained based on the expert-annotated Q matrix, the knowledge point prerequisite relationship matrix, and the question-knowledge point matrix; secondly, for exercise factors, the model considers the hypergraph relationship between different exercises, and uses the graph attention network (GAT) to construct an "exercise-knowledge point" relationship graph containing exercise similarity, fully extracting the complex relationship between knowledge points and between exercises and knowledge points; finally, in order to mine students' long-term and short-term semantic information, the one-hot encoding of the exercise and the one-hot encoding of the student are embedded together and used as the input of the Transformer to train the model, analyze the students' knowledge status level, and then predict the students' scores.
[0053] The present invention proposes a knowledge state prediction method based on a modified Q matrix and an attention mechanism, which mainly includes the following steps:
[0054] Step 1, data collection, including using the website to collect students' answer data and obtain the expert Q matrix used to express the relationship between exercises and knowledge points;
[0055] Step 2, modifying the Q matrix, including using the skill conversion method and the seven indicators proposed by Wenbin Gan to explore the potential knowledge graph, and finally obtaining the modified Q matrix;
[0056] Step 3, constructing an "exercise-knowledge point" relationship graph, including mining the associations between different exercises and obtaining a relationship graph by calculating the similarities between different exercises;
[0057] Step 4: embed the one-hot code of the exercises and the one-hot code of the students. The embedding process draws on the multidimensional item response theory model, maps the knowledge point proficiency vector, knowledge point relevance vector, knowledge point difficulty vector, and test question discrimination value into the formula, and performs fusion calculation on the parameters, thereby predicting the student's score and analyzing the student's knowledge status.
[0058] The technical solution of the present invention first collects students' answer data through the website, and constructs an expert Q matrix expressing the relationship between exercises and knowledge points through the knowledge and experience of domain experts. Then, the skill conversion method and the seven indicators proposed by Wenbin Gan are used to explore the potential knowledge graph, and finally the revised Q matrix is obtained. Secondly, the "exercise-knowledge point" relationship diagram is constructed, including mining the correlation relationship matrix between different exercises, and obtaining the relationship diagram by calculating the similarity between different exercises. Finally, the one-hot encoding of the exercises and the one-hot encoding of the students are embedded. The multidimensional item response theory model is used for reference in the embedding process, and the parameters such as the knowledge point proficiency vector, the knowledge point relevance vector, the knowledge point difficulty vector, and the test question discrimination value are respectively mapped to the formula, and the parameters are fused and calculated. In this way, the students' scores are predicted and the students' knowledge status is analyzed.
[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0060] Embodiment 1
[0061] The present invention discloses a knowledge state prediction method based on a modified Q matrix and an attention mechanism, comprising:
[0062] S1: Collect students’ answer data and construct an expert Q matrix based on the collected answer data to express the relationship between exercises and knowledge points through the knowledge and experience of domain experts;
[0063] S2: Generate an initial exercise-knowledge point relationship diagram based on the expert Q matrix, mine the prerequisites between knowledge points from the students' historical answer data, establish a knowledge point prerequisite relationship diagram, and then fuse the initial exercise-knowledge point relationship diagram with the knowledge point prerequisite relationship diagram to obtain an enhanced exercise-knowledge point relationship diagram, and then convert the enhanced exercise-knowledge point relationship diagram into a matrix form to obtain a revised Q matrix, where the knowledge point prerequisite relationship diagram is used to represent the prerequisite dependency conditions between knowledge points. If a student learns a knowledge point You can continue to learn another knowledge point ,but and There are prerequisite dependencies between them;
[0064] S3: Calculate the similarity between different exercises, use the similarity between different exercises as the correlation between different exercises, and construct a target exercise-knowledge point relationship graph based on the modified Q matrix and the correlation between different exercises; input the target exercise-knowledge point relationship graph into the graph attention network for training to obtain a hypergraph signal;
[0065] S4: Embed the student’s one-hot code, the exercise’s one-hot code, and the hypergraph signal, and finally obtain the student’s knowledge status through information aggregation and updating.
[0066] Specifically, S1 is the acquisition of the expert Q matrix, S2 fuses the initial exercise-knowledge point relationship graph with the knowledge point prerequisite relationship graph to obtain the enhanced exercise-knowledge point relationship graph, and then converts it into a revised Q matrix. Specifically, it is to establish a knowledge point prerequisite relationship graph by mining the prerequisites between knowledge points, and then fuse the prerequisites between knowledge points into the initial exercise-knowledge point relationship graph, and enhance the initial exercise-knowledge point relationship graph to obtain the enhanced exercise-knowledge point relationship graph. S3 is to calculate the similarity between different exercises to express the association between different exercises. On the basis of the enhanced exercise-knowledge point relationship graph, the similarity between different exercises is introduced to obtain the target exercise-knowledge point relationship graph, and then the target exercise-knowledge point relationship graph is input into the graph attention network for training to obtain a hypergraph signal. S4 is to aggregate and update information by introducing Transformer into the item response theory, predict the optimal knowledge point proficiency, and finally obtain the student's knowledge status.
[0067] The following is an explanation of the specific implementation steps:
[0068] 1. Collect students’ answer data and construct a Q matrix through the knowledge and experience of domain experts. Experts manually fill in the Q matrix based on their understanding of each question and each skill.
[0069] The website is used to collect students’ answer data and obtain the expert Q matrix of the relationship between expression exercises and knowledge points.
[0070] During the specific implementation process, the students' answer records are obtained, including the students' names, exercise numbers, and whether different students answer different exercises correctly or not, where a student's correct answer to the question is recorded as 1, otherwise it is recorded as 0; an expert Q matrix that can measure the correspondence between exercises and knowledge points is obtained. The Q matrix is also called a "0-1 matrix" or a Boolean matrix. It consists of row vectors and column vectors, where the row vectors represent the exercise numbers, and the column vectors represent the attributes of the knowledge points involved in the exercises. The Q matrix is a bridge between knowledge states and measurable and observable ideal response patterns.
[0071] This embodiment follows the following principles in the process of establishing the Q matrix.
[0072] First, the Q matrix must ensure that the item attributes contained in the test practice questions are covered; second, in order to ensure the accuracy of the measurement results, it is also necessary to ensure that each item attribute is examined by at least three test items, that is, each cognitive attribute must be examined at least three times by the practice questions; finally, it is necessary to check the degree of normalization of the Q matrix and eliminate invalid cognitive attributes.
[0073] 2. Obtain the revised Q matrix, including using the skill conversion method and the seven indicators proposed by Wenbin Gan to explore the knowledge point prerequisite dependency graph, and then integrate the prerequisites between the knowledge points into the initial practice-knowledge point relationship graph, enhance the initial practice-knowledge point relationship graph, thereby obtaining the enhanced practice-knowledge point relationship graph, and finally obtaining the revised Q matrix.
[0074] The modified Q matrix can be obtained by using the expert Q matrix and the learner's historical answer data, the knowledge point prerequisite relationship matrix and the question-knowledge point matrix. The basic process is as follows:
[0075] Phase 1: Obtain the initial exercise-knowledge point relationship diagram through the expert Q matrix.
[0076] First, determine the set of knowledge points that need to be evaluated or trained. Knowledge points can be specific skills or concepts in a certain field. Design exercises related to each knowledge point. Each question should test one or more knowledge points. Question design should be completed by domain experts to ensure the relevance and effectiveness of the questions and knowledge points. Construct an initial expert Q matrix. The rows of the Q matrix represent exercise questions and the columns represent knowledge points. The elements in the mth row and nth column of the Q matrix are Indicates whether the m-th exercise question involves the n-th knowledge point, using a binary value (0 for not involved, 1 for involved).
[0077] Generate an initial exercise-knowledge point relationship graph based on the Q matrix. The initial exercise-knowledge point relationship graph can graphically display the relationship between exercise questions and knowledge points. The nodes of the graph represent exercise questions and knowledge points, and the edges represent the relationship between questions and knowledge points.
[0078] (2) The second stage: mining the prerequisites between knowledge points from learners’ historical answer data and establishing a knowledge point prerequisite relationship diagram.
[0079] For example, when students learn knowledge points After that, you can continue to learn knowledge points , and There is a dependency relationship between them, that is, there are prerequisite dependencies between knowledge points. This relationship can be intuitively described by drawing a topological diagram, which is the knowledge point prerequisite relationship diagram.
[0080] The theoretical basis is the skill transfer method and the seven evaluation indicators proposed by Wenbin Gan. The skill transfer method is to infer the prerequisites from the order in which learners master skills by observing the students' historical answer data logs; and the seven evaluation indicators proposed by Wenbin Gan can measure the prerequisite relationship between a pair of knowledge points. These indicators are widely used to measure the degree of association between two variables.
[0081] ①Skill conversion method
[0082] In the process of student learning, different knowledge points will influence each other. In order to mine the influence of one knowledge point mastered by the learner on another knowledge point from the learning data, this implementation method uses the learner's correct or incorrect answer as the potential skill mastery indicator for a given question and and The interactive table summarizes the learner's knowledge and The impact is shown in the table.
[0083] Table 1 Knowledge points and Interaction table
[0084]
[0085] In Table 1, you need to master the knowledge points The problem arises when you need to master the knowledge points Before the question, when the same question pair appears multiple times in the learning data, this embodiment only considers the first question pair.
[0086] in, Indicates that the learner has mastered the knowledge point , and mastered the knowledge points situation; Indicates that the learner has mastered the knowledge point But I didn't master the knowledge points c means the learner has mastered the knowledge point But I didn't master the knowledge points d means that the learner has not mastered the knowledge point , and no knowledge points situation, Indicates that the learner has mastered the knowledge point The overall situation, and so on.
[0087] When mapping knowledge points, this embodiment finds that the evaluation indicators proposed by Wenbin Gan can measure the prerequisite relationship between a pair of knowledge points. These indicators are widely used to measure the degree of association between two variables. and The mutual influence table is constructed based on the evaluation index proposed by Wenbin Gan.
[0088] ② Seven evaluation indicators proposed by Wenbin Gan
[0089] These indices measure whether the premise relationship between a pair of skills is consistent. These indices are widely used to measure the association between two variables. The evaluation indicators are shown in Table 2.
[0090] Table 2 Use knowledge points and The mutual influence table obtains the evaluation index of the knowledge graph
[0091]
[0092] In Table 2, Cohen's Kappa is used to measure the agreement between two classifiers or evaluators, especially after adjusting for the effect of chance agreement.
[0093] Adjusted Kappa is the adjusted Kappa coefficient, which is used to control the impact of the number of categories and category distribution on the Kappa value and more accurately reflect the consistency between classifiers.
[0094] Phi Coefficient is used to measure the association between two binary variables.
[0095] Yule Coefficient is used to analyze the association between binary variables and is particularly suitable for binary tables.
[0096] The Ochiai Coefficient and Sokal Coefficient measure the similarity between two sets, emphasizing the proportion of common elements.
[0097] Jaccard Coefficient is also used to measure the similarity between two sets, especially when the data is sparse.
[0098] In the formula, a represents the number of samples that have mastered both knowledge points, b represents the number of samples that have mastered the first knowledge point but not the second knowledge point, c represents the number of samples that have mastered the second knowledge point but not the first knowledge point, and d represents the number of samples that have not mastered both knowledge points.
[0099] Since the constructed knowledge graph is a unidirectional graph, this implementation method simplifies it by setting a threshold to control the sparsity of the relationship between knowledge points in the knowledge graph. The final knowledge point prerequisite relationship matrix is expressed as ,in It represents a collection of skill conversion methods and seven evaluation index methods, among which The elements on the diagonal are set to 1, and the expressions involved are shown in Formula (1) and Formula (2).
[0100] (1)
[0101] (2)
[0102] The i, j in the above text are consistent with the knowledge points and The i and j in correspondence. represents the j-th row and i-th column element of the knowledge point prerequisite relationship matrix. The knowledge point prerequisite relationship matrix corresponds to the knowledge point prerequisite relationship graph. is the threshold used to control the sparsity of the relationship between knowledge points in the knowledge graph.
[0103] The third stage: The initial exercise-knowledge point relationship diagram is merged with the knowledge point prerequisite relationship diagram to obtain the enhanced exercise-knowledge point relationship diagram, which is converted into a matrix form, namely the modified Q matrix .
[0104] like Figure 1 As shown, this embodiment expresses the constructed initial exercise-knowledge point matrix based on the Q matrix annotated by the expert as , that is, the number of rows in the matrix is the number of knowledge points, and the number of columns is the number of exercises. According to the knowledge point prerequisite relationship matrix The question-knowledge matrix is corrected The matrix is represented as The mathematical expressions are shown in formula (3) and formula (4).
[0105] (3)
[0106] (4)
[0107] In formula (3), the knowledge point prerequisite relationship matrix The transposed matrix of explains the knowledge point prerequisite relationship in the current problem. Specifically, a question containing a specific knowledge point is related to the knowledge point in the prerequisite relationship.
[0108] 3. Construct the target "exercise-knowledge point" relationship diagram, including mining the associations between different exercises, calculating the similarities between different exercises, and constructing the target exercise-knowledge point relationship diagram based on the modified Q matrix and the associations between different exercises;
[0109] The modified Q matrix only mines the prerequisite dependencies between knowledge points, and does not take into account the hypergraph relationships between different exercises. In fact, not only are there associations between knowledge points, but there are also associations between exercises. One exercise can have an association relationship with another exercise, or with two or more other exercises, which are represented as hypergraph relationships in the graph structure.
[0110] The correlation matrix between different exercises can be mined by calculating the similarity between different exercises. According to this matrix, an "exercise-knowledge point" relationship graph (i.e., target exercise-knowledge point relationship graph) containing exercise similarity is constructed, and the relationship graph is placed in the graph attention network for training. The hypergraph signal X is embedded together with the one-hot encoding of the exercise, and the knowledge point difficulty vector is obtained by fusion.
[0111] In order to obtain the final hypergraph signal, the basic process is as follows:
[0112] (1) Phase 1: Construct an exercise-knowledge point relationship diagram.
[0113] When students always obtain similar answer results (correct / wrong) on two exercises during the answering process, it is assumed that there may be a close relationship between the two exercises. When these exercises have a similar relationship, the answer results obtained by the same student answering these exercises will be similar. The F1 score is an indicator used in statistics to measure the accuracy of binary models, and the degree of similarity between the predicted value and the target value can be inferred. Since the F1 score is an indicator that can well measure the similarity between two sets of binary data, the present invention uses the actual answer result of the previous exercise to predict the answer result of the next exercise, and calculates the F1 score between them as the similarity between them.
[0114] The similarity calculated by the F1 score is shown in formula (5).
[0115] (5)
[0116] in Indicates practice exist When answering before, the two answers were similar. The calculation method of is shown in formula (6).
[0117] (6)
[0118] in, and It is used to calculate part. and They are calculated by formula (7) and formula (8) respectively.
[0119] (7)
[0120] (8)
[0121] in Indicates the answer is adopted Answer Exercise In adopting the answer Answer Exercise In addition, in order to prevent the denominator from being too small, the present invention introduces the Laplace smoothing parameter λ into the equation, where the value of λ is 0.01. After generating similarity between two exercises, when When it is greater than the hyperparameter ω, the present invention will practice The edges between them are added to the exercise-knowledge point relationship graph. So far, the present embodiment has constructed the exercise-knowledge point relationship graph G and mapped it into the priority association matrix H of the exercise questions.
[0122] like Figure 2 As shown in , the relationship diagram between “exercise-knowledge point” is composed of exercise nodes, knowledge point nodes, the association relationship between exercises and knowledge points, and the similarity relationship between exercises, such as Figure 2 As shown, square nodes represent exercises, circular nodes represent knowledge points, the relationship between exercises and knowledge points is represented by a dotted line connection in the figure, and the association relationship between exercises (the present invention refers to the similarity relationship) is represented by a solid line in the figure.
[0123] Stage 2: Calculate hypergraph convolution.
[0124] Hypergraph convolution is a generalization of traditional graph convolution and is suitable for processing hypergraphs with complex relational structures. In a hypergraph, a hyperedge can connect multiple vertices, which is different from the binary edges in ordinary graphs. The computation process of hypergraph convolution involves propagating and aggregating hypergraph signals.
[0125] The exercise-knowledge point hypergraph G is placed into the graph attention network for training. The calculation process of the hypergraph convolution is described as follows.
[0126] Given Exercises - Knowledge Point Hypergraph , a set of vertices Include vertices, initial vertex feature matrix The size is ,in is the dimension of the initial features.
[0127] The hypergraph convolution operation is performed through the hypergraph Laplacian matrix The vertex features are propagated and aggregated. The convolution operation can be expressed as:
[0128] (9)
[0129] in is the initial vertex feature matrix, is the weight matrix to be learned, is an activation function (such as ReLU), and finally obtains the hypergraph signal .
[0130] Compute the hypergraph Laplacian matrix The method is shown in formula (10):
[0131] (10)
[0132] in is the identity matrix, is the vertex degree matrix, the diagonal elements Represents a vertex The degree. Reason Exercise-Knowledge Point Relationship Diagram Mapped exercise priority association matrix . is the hyperedge degree matrix, with diagonal elements Represents a hyperedge The degree.
[0133] Through the hypergraph convolution operation , get the updated vertex features , that is, the hypergraph signal.
[0134] After obtaining the hypergraph signal X, the hypergraph signal X can be embedded together with the one-hot encoding of the exercise to obtain the knowledge point difficulty vector.
[0135] Compared with traditional neural networks, the graph attention network (GAT) can better extract the complex relationship between exercises and knowledge points in students' answer data, and between different exercises.
[0136] 4. Embedding of students and exercises one-hot encoding, including information aggregation and updating, conducts model training, and finally obtains the student knowledge status.
[0137] like Figure 3 As shown in the model, the knowledge point proficiency vector is used Represents students and indicates the degree of their mastery of knowledge points; selects knowledge point correlation vector It represents the examination status of the knowledge points by the exercises, that is, whether they are examined or not, and is mapped to a row of the Q matrix.
[0138] The model uses the knowledge point difficulty vector of the exercise Indicates the difficulty of each knowledge point tested by the exercise, which is extended from the exercise difficulty used in item response theory. The difficulty of the exercise is evaluated based on the overall performance of the students. If most students can answer an exercise correctly, the knowledge difficulty of the exercise is low; conversely, if most students answer incorrectly, the knowledge difficulty of the exercise is high.
[0139] The knowledge point discrimination vector of this model It represents the ability of the exercise to distinguish between students of different levels. It is a numerical variable. The discrimination parameter reflects the difference in responses between students of different ability levels.
[0140] The knowledge point proficiency vector, knowledge point relevance vector, knowledge point difficulty vector, and knowledge point discrimination vector are obtained by one-hot encoding students and exercises, respectively, and inputting them into the deep neural network for training. is the student one-hot vector, is the one-hot vector of the exercise, and A and D represent trainable matrices. Considering that the exercises involve multi-dimensional skill features, this model also draws on the Multidimensional Item Response Theory (MIRT) model, maps the knowledge point proficiency vector, knowledge point relevance vector, knowledge point difficulty vector, and exercise discrimination value into the formula, and performs fusion calculation on these parameters.
[0141] In the embedding stage, for the student factor, the student's one-hot encoding Embedded, multiplied by the trained A matrix to obtain the student's proficiency vector , Each entry in is a continuous vector with values in the interval [0,1]. =[0.7,0.1] indicates that students have a high level of mastery of the first knowledge concept, but a low level of mastery of the second knowledge concept.
[0142] For the training factor, one-hot encoding of the training Embedding, multiplying with the trained D matrix and the modified Q matrix to obtain the training discrimination vector And the knowledge point relevance vector .
[0143] The mathematical expressions are shown in formulas (11) to (13):
[0144] (11)
[0145] (12)
[0146] (13)
[0147] Embed the hypergraph vector X with the one-hot encoding of the exercise Embedding fusion to obtain knowledge point difficulty vector , its mathematical expression is shown in formula (14):
[0148] (14)
[0149] Among them, A and D are both trainable matrices, whose values can be updated through optimization algorithms during the training process. By learning, useful features or relationships can be extracted from the data to improve the performance of the model. Sigmoid (·) is an activation function that provides nonlinear transformation, compresses the input value to the range of (0, 1), and provides a smooth gradient for easy training and optimization.
[0150] Then information aggregation and updating are performed, such as Figure 4As shown, based on Item Response Theory (IRT), this model is inspired by Multidimensional IRT (MIRT) and obtains the input of Transformer .
[0151] Its mathematical expression is shown in formula (15):
[0152] (15)
[0153] After training, the final model can get the student's predicted score y. The parameters in the model are The value of is the prediction result obtained by the present invention, which represents the knowledge status of the student.
[0154] It should be noted that Figure 4 In the figure, FeedForward represents the feedforward layer, Cross Attention represents cross attention, and Self Attention represents self attention.
[0155] See also Figure 5 , which is an overall framework diagram of the knowledge state prediction method based on the modified Q matrix and hypergraph attention mechanism in an embodiment of the present invention.
[0156] The method and model of the present invention are explained and verified through specific experiments below.
[0157] 1. Data processing
[0158] In this implementation, the data set is divided into a training set and a test set, of which 80% is used for training and 20% is used for testing. The model public data set is used for experiments and compared with multiple existing cognitive diagnosis models. These public data sets are Assist0910, Assist1213 and EdNet.
[0159] The Assist0910 and Assist1213 datasets were collected in the ASSISTments intelligent tutoring system, which is a computer-based intelligent tutoring system provided by Worcester Polytechnic Institute in the United States. The system can analyze learners' learning characteristics, establish adaptive learning scaffolds for students, provide a variety of learning diagnostic reports, and provide personalized guidance for students, with both "support" and "evaluation" functions. The Assist0910 dataset was collected using the ASSISTments online tutoring system from 2009 to 2010, and contains the interactions of 3,002 learners and 17,705 projects involving 123 skills, forming 277,540 interaction records. Assist1213 dataset updates Assist0910 and was collected throughout 2012. In this dataset, 22,339 learners in the system attempted 52,825 math problems requiring 265 skills. This dataset contains the largest number of learners, problems, and interaction entries (nearly 2.7 million entries). It is a dataset with only one skill for each problem, which means that each problem requires one skill. It provides students' answer logs and knowledge concepts and exercise-concept correlations.
[0160] The EdNet dataset is a dataset of all student system interactions collected by Santa, a multi-platform artificial intelligence tutoring service company. The dataset is large-scale and hierarchical, and Santa can be used through Android, iOS, and the Internet. The EdNet selected in this experiment randomly selected 5,000 students from the original dataset. They answered 12,372 questions and required 188 skills, thus obtaining 347,866 interaction logs. The statistics of the three datasets are shown in Table 3.
[0161] Table 3 Dataset statistics
[0162]
[0163] 2. Baseline model selection
[0164] In order to verify the effect of the HGCDQ+ proposed in this invention on students' cognitive diagnosis (knowledge status prediction), this experiment compared HGCDQ+ with the classic cognitive diagnosis methods IRT, MIRT, DINA and the popular deep learning cognitive diagnosis method NeuralCDM.
[0165] IRT: is a mathematical model used to quantify individual characteristics (such as abilities, aptitudes, or attitudes) in a test, questionnaire, or quiz. It estimates the level of an individual's characteristic based on their responses to items (such as questions, tasks, or statements).
[0166] MIRT: is a model used to evaluate multiple dimensions in an exam or questionnaire. Compared with the traditional IRT model, MIRT considers more than one ability dimension. It assumes that the subjects have different potential ability levels on different ability dimensions, and each exercise item is also associated with multiple dimensions.
[0167] DINA: The DINA model is a binary multidimensional discrete cognitive diagnostic model that introduces guessing parameters and error parameters to simulate students' answers to objective questions.
[0168] NeuralCDM: A cognitive diagnosis method based on artificial neural networks, which combines student factors and exercise factors into a forward fully connected layer to predict learners' answers to objective and subjective questions.
[0169] 3. Evaluation Metric Selection
[0170] This experiment uses common evaluation indicators in the field of cognitive diagnosis, including Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The evaluation indicators for objective questions also include Accuracy (ACC).
[0171] (1) ACC: represents the ratio of the answer results correctly predicted by the model to the total true answer results. The calculation method is shown in formula (16).
[0172] (16)
[0173] The meaning of each letter is:
[0174] TP (True Positive): True positive, which means the number of samples correctly classified as positive.
[0175] TN (True Negative): True negative examples, which means the number of samples correctly classified as negative classes.
[0176] FP (False Positive): False positive examples, which indicates the number of negative samples that are mistakenly classified as positive samples.
[0177] FN (False Negative): False negative examples, which indicates the number of positive samples that are mistakenly classified as negative samples.
[0178] The numerator (TP+TN) represents the total number of correctly classified samples, including correctly predicted positive and negative samples. The denominator (TP+TN+FP+FN) represents the total number of all samples, including all correctly predicted and incorrectly predicted samples.
[0179] (2) RMSE: represents the average value of the squared error between the predicted answer result and the actual answer result. The calculation method is shown in formula (2). Although RMSE and MAE both describe the difference between the predicted answer result and the actual answer result, RMSE focuses more on abnormal points and amplifies the abnormal differences. Therefore, this experiment selects RMSE as the evaluation indicator.
[0180] (17)
[0181] Where p represents the total number of samples, Indicates The true value of the sample, Indicates The predicted value of samples, Indicates The square of the prediction error for each sample.
[0182] (3) AUC: As a model evaluation indicator, it is often used to evaluate the performance of binary classification models. The closer the AUC value is to 1, the stronger the predictive ability of the model. If the AUC is equal to 0.5, the predictive ability of the model is equivalent to random guessing. It is defined as the area under the ROC curve, where the full name of the ROC curve is the receiver operating characteristic curve. It is a curve drawn based on a series of different binary classification methods (cutoff values or decision thresholds) with the true positive rate (sensitivity) as the vertical axis and the false positive rate (1-specificity) as the horizontal axis.
[0183] 4. Comparative experiment and result analysis
[0184] This experiment divides the data set into a training set and a test set, with 80% of the data set used for training and 20% for testing. The model is continuously optimized to obtain better parameters and improve the accuracy of diagnosis. The results are compared with benchmark models (such as IRT, MIRT, etc.) and existing neurocognitive diagnosis models (such as NeuralCDM, etc.). Each model is tested 10 times on three data sets, and the experimental results with the best accuracy are recorded.
[0185] According to the data in Tables 4 to 6, the HGCDQ+ model performs better than the baseline model in cognitive diagnosis on different data sets.
[0186] Table 4 Test results of ASSIST0910 dataset
[0187]
[0188] The experimental results on the ASSIST0910 dataset are shown in Table 4. Compared with the other models, compared with NeuralCDM, QNCD has an Accuracy index of 0.83 percentage points higher, RMSE index decreased by 6.1 percentage points, and AUC index increased by 10.65 percentage points on the ASSIST0910 dataset. Compared with QNCD, the hypergraph convolution-based cognitive diagnosis model HGCD has an Accuracy index of 0.85 percentage points higher, RMSE index decreased by 7.66 percentage points, and AUC index increased by 1.92 percentage points in the ASSIST0910 dataset. On the basis of HGCD, HGCDQ+, which adds a modified Q matrix, has an Accuracy index of 1.84 percentage points higher, RMSE index decreased by 9.11 percentage points, and AUC index increased by 2.52 percentage points in the ASSIST0910 dataset.
[0189] Table 5 Test results of ASSIST1213 dataset
[0190]
[0191] The experimental results on the ASSIST1213 dataset are shown in Table 5. Compared with the other models, compared with NeuralCDM, QNCD has an increase of 1.08 percentage points in the Accuracy index, a decrease of 6.62 percentage points in the ASSIST1213 dataset, and an increase of 10.41 percentage points in the AUC index. Compared with QNCD, the hypergraph convolution-based cognitive diagnosis model HGCD has an increase of 1.12 percentage points in the Accuracy index, a decrease of 2.7 percentage points in the ASSIST1213 dataset, and an increase of 0.85 percentage points in the AUC index. On the basis of HGCD, HGCDQ+, which adds a modified Q matrix, has an increase of 1.77 percentage points in the Accuracy index, a decrease of 6.15 percentage points in the RMSE index, and an increase of 2.29 percentage points in the AUC index.
[0192] Table 6 EdNet dataset test results
[0193]
[0194] The experimental results on the EdNet dataset are shown in Table 6. Compared with the other models, compared with NeuralCDM, QNCD has an increase of 1.13 percentage points in the Accuracy index, a decrease of 1.6 percentage points in the RMSE index, and an increase of 1.32 percentage points in the AUC index on the EdNet dataset. Compared with QNCD, the hypergraph convolution-based cognitive diagnosis model has an increase of 1.98 percentage points in the Accuracy index, a decrease of 2.92 percentage points in the RMSE index, and an increase of 1.07 percentage points in the AUC index on the EdNet dataset. Based on HGCD, HGCDQ+, which adds a modified Q matrix, has an increase of 2.48 percentage points in the Accuracy index, a decrease of 5.23 percentage points in the RMSE index, and an increase of 2.38 percentage points in the AUC index on the EdNet dataset.
[0195] In summary, considering the dependency relationship between different knowledge points and the Q matrix, and the hyperdimensional relationship between exercises can help improve the performance of the neurocognitive diagnosis model, and further improve the model performance of neurocognitive diagnosis. Among them, HGCDQ+, which comprehensively considers the two, is better than QNCD, which considers the dependency relationship between different knowledge points and the Q matrix, and HGCD, which considers the hyperdimensional relationship between exercises.
[0196] Embodiment 2
[0197] Based on the same inventive concept, this embodiment discloses a knowledge state prediction device based on a modified Q matrix and an attention mechanism, including:
[0198] An expert Q matrix acquisition module is used to collect students' answer data and construct an expert Q matrix for expressing the relationship between exercises and knowledge points based on the collected answer data through the knowledge and experience of domain experts;
[0199] Generate an exercise-knowledge point relationship diagram based on the expert Q matrix, mine the prerequisites between knowledge points from the students' historical answer data, establish a knowledge point prerequisite relationship diagram, and then merge the exercise-knowledge point relationship diagram with the knowledge point prerequisite relationship diagram to obtain an enhanced exercise-knowledge point relationship diagram, and then convert the enhanced exercise-knowledge point relationship diagram into a matrix form to obtain a revised Q matrix, in which the knowledge point prerequisite relationship diagram is used to represent the prerequisite dependency conditions between knowledge points. If a student learns a knowledge point You can continue to learn another knowledge point ,but and There are prerequisite dependencies between them;
[0200] The exercise-knowledge point relationship graph construction module is used to calculate the similarity between different exercises, and use the similarity between different exercises as the correlation between different exercises. Based on the modified Q matrix and the correlation between different exercises, the exercise-knowledge point relationship graph is constructed; the enhanced exercise-knowledge point relationship graph is input into the graph attention network for training to obtain a hypergraph signal;
[0201] The knowledge status prediction module is used to embed the student's one-hot encoding, the exercise's one-hot encoding, and the hypergraph signal, and finally obtain the student's knowledge status through information aggregation and updating.
[0202] Specifically, the knowledge state prediction device based on the modified Q matrix and the hypergraph attention mechanism in this embodiment refers to the knowledge state prediction model based on the modified Q matrix and the hypergraph attention mechanism, that is, HGCDQ+ in Example 1.
[0203] Since the device introduced in the second embodiment of the present invention is a device used to implement the knowledge state prediction method based on the modified Q matrix and the attention mechanism in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, the person skilled in the art can understand the specific structure and deformation of the device, so it is not repeated here. All devices used in the method in the first embodiment of the present invention belong to the scope of protection of the present invention.
[0204] Embodiment 3
[0205] Based on the same inventive concept, the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method described in the first embodiment is implemented.
[0206] Since the computer-readable storage medium introduced in the third embodiment of the present invention is a computer-readable storage medium used to implement the knowledge state prediction method based on the modified Q matrix and the attention mechanism in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, the person skilled in the art can understand the specific structure and deformation of the computer-readable storage medium, so it is not repeated here. All computer-readable storage media used in the method of the first embodiment of the present invention belong to the scope of protection of the present invention.
[0207] Embodiment 4
[0208] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in Embodiment 1 when executing the program.
[0209] Since the computer device introduced in the fourth embodiment of the present invention is a computer device used to implement the knowledge state prediction method based on the modified Q matrix and the attention mechanism in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, the person skilled in the art can understand the specific structure and deformation of the computer device, so it is not repeated here. All computer devices used in the method of the first embodiment of the present invention belong to the scope of protection of the present invention.
[0210] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0211] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0212] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic creative concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. A knowledge state prediction method based on a modified Q matrix and an attention mechanism, characterized in that: include: Collect students' answer data, and build an expert Q matrix based on the collected answer data to express the relationship between exercises and knowledge points through the knowledge and experience of domain experts; The initial exercise-knowledge point relationship diagram is generated according to the expert Q matrix, and the prerequisites between knowledge points are mined from the students' historical answer data to establish a knowledge point prerequisite relationship diagram. Then, the initial exercise-knowledge point relationship diagram is merged with the knowledge point prerequisite relationship diagram to obtain an enhanced exercise-knowledge point relationship diagram. The enhanced exercise-knowledge point relationship diagram is then converted into a matrix form to obtain a revised Q matrix, in which the knowledge point prerequisite relationship diagram is used to represent the prerequisite dependency conditions between knowledge points. If a student learns a knowledge point, You can continue to learn another knowledge point ,but and There are prerequisite dependencies between them; Calculate the similarity between different exercises, use the similarity between different exercises as the correlation between different exercises, and construct a target exercise-knowledge point relationship graph based on the modified Q matrix and the correlation between different exercises; input the target exercise-knowledge point relationship graph into the graph attention network for training to obtain a hypergraph signal; The student's one-hot code, the exercise's one-hot code, and the hypergraph signal are embedded, and the student's knowledge status is finally obtained through information aggregation and updating; Among them, the modified Q matrix is calculated based on the expert Q matrix and the knowledge point prerequisite relationship matrix, and the calculation method is: in, represents the knowledge point prerequisite relationship matrix, represents the modified Q matrix, the number of rows of the matrix is the number of knowledge points, and the number of columns is the number of exercises. represents the element in the i-th row and j-th column of the modified Q matrix; The student’s one-hot code, the exercise’s one-hot code, and the hypergraph signal are embedded. Through information aggregation and updating, the student’s knowledge status is finally obtained, including: One-hot encoding of students Embedded, multiplied by the trained A matrix to obtain the student's proficiency vector , where the proficiency vector represents the students’ mastery of the knowledge points; One-hot encoding of the exercise Embedding, multiplying with the trained D matrix to obtain the knowledge point discrimination vector , multiplied by the modified Q matrix to obtain the knowledge point relevance vector , where the knowledge point discrimination vector represents the ability of the exercise to distinguish students of different levels, and the knowledge point relevance vector represents the examination of the exercise on the knowledge point; Hypergraph Signal X One-hot encoding with practice Embedding fusion to obtain knowledge point difficulty vector , where the knowledge point difficulty vector represents the students’ mastery of the knowledge point; The student's proficiency vector, knowledge point discrimination vector, knowledge point relevance vector and knowledge point difficulty vector are embedded, and then information aggregation and update are performed to obtain the student's predicted score as the prediction result of the student's knowledge status.
2. The knowledge state prediction method based on the modified Q matrix and attention mechanism as claimed in claim 1, characterized in that: The expert Q matrix consists of row vectors and column vectors, where the row vectors represent the exercise labels and the column vectors represent the attributes of the knowledge points involved in the exercise. The element in the mth row and nth column of the Q matrix is Indicates whether the mth exercise question involves the nth knowledge point. If so, it is 1; if not, it is 0.
3. The knowledge state prediction method based on the modified Q matrix and attention mechanism as claimed in claim 1, characterized in that: Mining the prerequisites between knowledge points from students’ historical answer data, and establishing a knowledge point prerequisite relationship diagram, including: Analyze students' historical answer data and infer the influence of one knowledge point on another from the order in which learners master skills; Use preset evaluation indicators to measure the degree of association between knowledge points; A knowledge point prerequisite relationship diagram is constructed based on the impact of one knowledge point on another knowledge point and the degree of association between knowledge points.
4. The knowledge state prediction method based on the modified Q matrix and attention mechanism as claimed in claim 3, characterized in that: The method further includes setting a threshold to control the sparsity of the relationship between knowledge points in the knowledge point prerequisite relationship graph, specifically: represents the i-th row and j-th column element of the knowledge point prerequisite relationship matrix, represents the j-th row and i-th column element of the knowledge point prerequisite relationship matrix. The knowledge point prerequisite relationship matrix corresponds to the knowledge point prerequisite relationship graph. is the threshold used to control the sparsity of the relationship between knowledge points in the knowledge graph.
5. The knowledge state prediction method based on the modified Q matrix and attention mechanism as claimed in claim 1, characterized in that: The way to calculate the similarity between different exercises is: in, Indicates practice and The similarity between Indicates practice exist When answering before, the similarity between the two answers, practise exist When answering before, the two answers were similar.
6. A knowledge state prediction device based on a modified Q matrix and an attention mechanism, characterized in that: include: Expert Q matrix acquisition module, which is used to collect students' answer data and build an expert Q matrix based on the collected answer data to express the relationship between exercises and knowledge points through the knowledge and experience of domain experts The modified Q matrix acquisition module is used to generate an initial exercise-knowledge point relationship diagram based on the expert Q matrix, mine the prerequisites between knowledge points from the students' historical answer data, establish a knowledge point prerequisite relationship diagram, and then fuse the initial exercise-knowledge point relationship diagram with the knowledge point prerequisite relationship diagram to obtain an enhanced exercise-knowledge point relationship diagram, and then convert the enhanced exercise-knowledge point relationship diagram into a matrix form to obtain a modified Q matrix, where the knowledge point prerequisite relationship diagram is used to represent the prerequisite dependency conditions between knowledge points. If a student learns a knowledge point You can continue to learn another knowledge point ,but and There are prerequisite dependencies between them; The module for constructing the exercise-knowledge point relationship graph is used to calculate the similarity between different exercises, and use the similarity between different exercises as the correlation between different exercises. Based on the modified Q matrix and the correlation between different exercises, the target exercise-knowledge point relationship graph is constructed; the target exercise-knowledge point relationship graph is input into the graph attention network for training to obtain a hypergraph signal; The knowledge status prediction module is used to embed the student's one-hot code, the exercise's one-hot code, and the hypergraph signal, and finally obtain the student's knowledge status through information aggregation and updating; Among them, the modified Q matrix is calculated based on the expert Q matrix and the knowledge point prerequisite relationship matrix, and the calculation method is: in, represents the knowledge point prerequisite relationship matrix, represents the modified Q matrix, the number of rows of the matrix is the number of knowledge points, and the number of columns is the number of exercises. represents the element in the i-th row and j-th column of the modified Q matrix; The student’s one-hot code, the exercise’s one-hot code, and the hypergraph signal are embedded. Through information aggregation and updating, the student’s knowledge status is finally obtained, including: One-hot encoding of students Embedded, multiplied by the trained A matrix to obtain the student's proficiency vector , where the proficiency vector represents the students’ mastery of the knowledge points; One-hot encoding of the exercise Embedding, multiplying with the trained D matrix to obtain the knowledge point discrimination vector , multiplied by the modified Q matrix to obtain the knowledge point relevance vector , where the knowledge point discrimination vector represents the ability of the exercise to distinguish students of different levels, and the knowledge point relevance vector represents the examination of the exercise on the knowledge point; Hypergraph Signal X One-hot encoding with practice Embedding fusion to obtain knowledge point difficulty vector , where the knowledge point difficulty vector represents the students’ mastery of the knowledge point; The student's proficiency vector, knowledge point discrimination vector, knowledge point relevance vector and knowledge point difficulty vector are embedded, and then information aggregation and update are performed to obtain the student's predicted score as the prediction result of the student's knowledge status.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the knowledge state prediction method based on the modified Q matrix and the attention mechanism is implemented as described in any one of claims 1 to 5.
8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the knowledge state prediction method based on the modified Q matrix and attention mechanism is implemented as described in any one of claims 1 to 5.
