A cognitive diagnosis method based on intra-stratum similarity relationship

Abstract

The application discloses a diagnostic method based on intra-layer similarity relation, comprising the following steps: 1. According to the history answer record of students, the correlation between exercises and knowledge points, the similarity among students, exercises and knowledge points is calculated to form intra-layer similarity relation; 2. The embedding representation of nodes is learned by using the intra-layer similarity relation graph convolution network, and the nodes with less exercise records or knowledge point correlation information are called student tail nodes or exercise tail nodes. The head node information is transmitted to the tail nodes to improve the sparsity of the inter-layer relation of the tail nodes; 3. The cognitive state of students is obtained by fusing the student vector and the knowledge point vector, and then the cognitive state of students and the knowledge point mastery required by exercises are obtained. The application realizes the information transmission of tail nodes by increasing the intra-layer similarity relation among students, exercises and knowledge points, so that the cognitive diagnosis precision of students with less exercises can be improved.

Description

Technical Field

[0001] This invention relates to cognitive diagnosis in the field of intelligent education, specifically a cognitive diagnosis method based on intra-layer similarity relationships. Background Technology

[0002] In the field of education, with the widespread application of intelligent education systems, how to model students' knowledge mastery and make personalized exercise recommendations has become a focus of attention.

[0003] Cognitive diagnostic models model students' mastery of different knowledge points based on their historical answer records, and then use common diagnostic models to predict students' accuracy on unanswered questions. Early work used manually designed functions to model the linear relationship between students and questions, such as traditional item response theory to model single student abilities. Most existing cognitive diagnostic models rely on ample student answer records for diagnosis. However, in reality, the relationships between student answer records, questions, and knowledge points exhibit a long-tailed distribution; that is, some students answer fewer questions, and some questions only cover a few knowledge points. This long-tailed distribution poses a challenge to model training. How to utilize students' historical answer records and the relationships between questions and knowledge points to obtain information about students with insufficient answer records' mastery of knowledge points and to predict their answers has become a worthy research question. Summary of the Invention

[0004] The present invention aims to address the shortcomings of the existing technology by proposing a cognitive diagnostic method based on intra-layer similarity relationships. This method aims to increase the intra-layer similarity relationships between students, exercises, and knowledge points to achieve information transmission at the tail node, thereby enabling more accurate prediction of students' answer performance and their mastery of specific knowledge points.

[0005] The present invention adopts the following technical solution to solve the technical problem:

[0006] The cognitive diagnostic method based on intra-layer similarity relationships of this invention is characterized by the following steps:

[0007] Step 1: Obtain students' historical answer records and construct a heterogeneous graph:

[0008] Let S denote the student set, and S = {s1,...,s...} a ,...,s N}, where s a Let represent the a-th student, N represent the total number of students, 1≤a≤N; let E represent the problem set, and E={e1,...,e i ,...,e M}, where e iLet represent the i-th exercise, M represent the total number of exercises, 1 ≤ i ≤ M; let K represent the set of knowledge points, and K = {k1, ..., k}. j ,...,k T}, where k j Let K represent the j-th knowledge point, and K represent the total number of knowledge points, where 1 ≤ j ≤ T;

[0009] Each student's chosen exercises and the knowledge points they encompass are represented in a three-part graph. The relationships between the student's answers and the relationships between the knowledge points contained in the exercises are termed inter-level relationships. The set of interaction records between students and exercises is denoted as... Let the set of relationships between exercises and knowledge points be denoted as... in, s represents the a-th student a and the i-th exercise e i Interaction records, Represents the i-th exercise e i and the kth knowledge point k j The edges corresponding to the relationship between them;

[0010] Step 2: Initialize the student embedding representation, exercise embedding representation, and knowledge point embedding representation;

[0011] Step 2.1: After initializing the student set S using the Xavier method, the student representation set E is obtained. s ={e s1 ,...,e sa ,...,e sN}, where e sa s represents the a-th student a d-dimensional embedding representation;

[0012] Step 2.2: After initializing the problem set E using the Xavier method, the problem representation set E is obtained. e ={e e1 ,...,e ei ,...,e eM}, where e ei Represents the i-th exercise e i d-dimensional embedding representation;

[0013] Step 2.3: After initializing the knowledge point set K using the Xavier method, the knowledge point representation set E is obtained. k ={e k1 ,...,e kj ,...,e kT}, where e kj Represents the j-th knowledge point k j d-dimensional embedding representation;

[0014] Step 3: Construct a cognitive diagnostic model based on intra-layer similarity relationships, including: an intra-layer similarity relationship module, a node representation module for fusion relationships, and a diagnostic module; wherein, the node representation module for fusion relationships includes: an L-layer relationship graph convolutional layer;

[0015] Step 3.1: Construct intra-layer similarity relationships based on the intra-layer similarity relationship module to obtain the student-student intra-layer relationship set R. s Exercises - The set of in-level relations R of exercises e Knowledge Point-Knowledge Point Hierarchical Relationship Set R k ;

[0016] Step 3.2: Based on the node representation module of the fusion relationship, message passing is performed to obtain the a-th student s updated in the L-th step. a The embedding representation is denoted as e′ sa The i-th exercise e i The embedding representation is denoted as e′ ei The j-th knowledge point k j The embedding representation is denoted as e′ kj ;

[0017] Step 3.3, Processing of the cognitive diagnosis module:

[0018] Step 3.3.1: Use equation (8) to obtain the a-th student s. a The embedded representation of e ca , used to represent the a-th student s a Cognitive state:

[0019] e ca =FC(e′) sa ·e′ k (8)

[0020] In equation (8), e′ k The matrix is ​​formed by embedding all knowledge points, where σ(·) represents the sigmoid activation function; FC represents a fully connected layer.

[0021] Step 3.3.2: Use equation (9) to predict the a-th student s. a Answer the i-th exercise correctly. i The score y ai :

[0022] y ai =L(W(e) ca -e′ ei (9)

[0023] In equation (9), W(·) is a pooling layer that reduces the dimensionality of the student's cognitive state and the exercise representation; L(·) is a scoring function using a fully connected layer.

[0024] Step 4: Construct the cross-entropy loss function Loss using equation (10):

[0025]

[0026] In equation (10), y ai It is the a-th student s a Answer the i-th exercise e i The predicted results, r ai It is the a-th student s a Answer the i-th exercise e i The actual results;

[0027] Step 5: Train the cognitive diagnostic model using the Adam optimizer and minimize the total loss function (Loss) to update the model parameters until convergence, thereby obtaining the trained cognitive diagnostic model. Based on the final embedding representation of the student nodes, the trained cognitive diagnostic model is used to determine the student's mastery of different knowledge points. The cognitive diagnostic method based on intra-layer similarity described in this invention is characterized in that step 3.1 is performed as follows:

[0028] Step 3.1.1: Using equation (1), construct the s of the a-th student who has done the exercises together. a And the bth student s b Similarity between Sim s (s a ,s b This allows us to obtain the similarity matrix Sim between each student. s :

[0029]

[0030] In equation (1), s represents the a-th student a A collection of interaction records with completed exercises. s represents the b-th student b The set of interaction records with those who have completed exercises, 1≤a≤N, 1≤b≤N, a≠b;

[0031] Step 3.1.2: Obtain the threshold λ using equation (2):

[0032] λ=α×AVE(Sim s (2)

[0033] In equation (2), α is the set coefficient; AVE indicates the averaging operation;

[0034] Step 3.1.3, when Sim s (sa ,s b When ) > λ, for the a-th student s a And the bth student s b By establishing the edges corresponding to the similarity relationships between students, we can obtain the set of relationships between students;

[0035] Step 3.1.4: Construct the p-th exercise e using equation (3). p and the qth exercise e q Similarity between Sim e (e p ,e q This allows us to obtain the similarity matrix Sim between each exercise. e ;

[0036]

[0037] In equation (3), This represents the p-th exercise e. p A collection containing knowledge points e represents the q-th exercise. q A set containing knowledge points, where 1 ≤ p ≤ M, 1 ≤ q ≤ M, and p ≠ q;

[0038] Step 3.1.5: Construct the u-th knowledge point k using equation (4). u and the vth knowledge point k v Similarity between Sim k (k u ,k v ) u,v Thus, the similarity matrix Sim between knowledge points is obtained. k ;

[0039]

[0040] In equation (4), n u,v This indicates that the answer to the u-th knowledge point k was correct. u After correctly answering the v-th knowledge point k v The number of times, 1≤u≤T, 1≤v≤T, u≠v; n u,t This indicates that the answer to the u-th knowledge point k is correct. u After correctly answering the t-th knowledge point k t The number of times, 1 < t < T;

[0041] According to Sim s Sim e Sim k We obtain the student-student hierarchical relation set R. s Exercises - The set of in-level relations R of exercises e Knowledge Point-Knowledge Point Hierarchical Relationship Set Rk .

[0042] Step 3.2 is performed as follows:

[0043] Step 3.2.1: Define the current update count as l, and initialize l = 0;

[0044] Define the a-th student s a The feature vector under the l-th update The i-th exercise e i The feature vector under the l-th update The j-th knowledge point k j The feature vector under the l-th update and initialize

[0045] Step 3.2.2: The relationship graph convolution uses equations (5), (6), and (7) to obtain the a-th student s. a User feature vector under the (l+1)th update The i-th exercise e i The problem embedding representation under the (l+1)th update The j-th knowledge point k j Embedded representation of knowledge points under the (l+1)th update

[0046]

[0047]

[0048]

[0049] In equations (5), (6), and (7), s represents the a-th student a In relation r s ∈R s ∩R se The set of relationships between the current node and its neighboring nodes; For the a-th student s a The representation of neighboring nodes under the (l+1)th update, W sr (l) s represents the a-th student a The weights of neighboring nodes in the l-th update, W s0 (l) s represents the a-th student a The weight of the self-embedded representation under the l-th update; Represents the i-th exercise e i In relation r e ∈R e ∩Rse ∩R ek The set of relationships between the current node and its neighboring nodes; For the i-th exercise e i The representation of neighboring nodes under the l-th update, W er (l) Represents the i-th exercise e i The weights of neighboring nodes in the l-th update, W e0 (l) Represents the i-th exercise e i The weights of the self-embedded representation under the l-th update; Represents the j-th knowledge point k j In relation r k ∈R k ∩R ek The set of relationships between the current node and its neighboring nodes; For the j-th knowledge point k j The representation of neighboring nodes under the l-th update, W kr (l) Represents the j-th knowledge point k j The weights of neighboring nodes in the l-th update, W k0 (l) Represents the j-th knowledge point k j The weights of the self-embedding representation under the l-th update; σ represents the LeakyReLU activation function;

[0050] Step 3.2.3: After assigning l+1 to l, return to step 3.2.2 and execute sequentially until l>L, thus obtaining the a-th student s updated in the L-th iteration. a The embedding representation is denoted as e′ sa The i-th exercise e i The embedding representation is denoted as e′ ei The j-th knowledge point k j The embedding representation is denoted as e′ kj .

[0051] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the cognitive diagnostic method, and the processor is configured to execute the program stored in the memory.

[0052] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program is executed by a processor to perform the steps of the cognitive diagnostic method.

[0053] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0054] To better diagnose the cognitive status of students lacking practice problem records, this invention proposes a cognitive diagnosis method based on intra-layer similarity relationships. Addressing the long-tail distribution problem in the student-problem-knowledge point association graph in cognitive diagnosis, this method improves diagnostic performance by constructing intra-layer similarity relationships to mitigate the long tail. This allows for message passing between nodes of the same and different types, fully considering the interaction relationships between multi-source data, and is more conducive to supplementing the cognitive status of students lacking practice problem records. Simultaneously, it also effectively improves the accuracy of student performance prediction. Detailed Implementation

[0055] In this embodiment, a multi-relational cognitive diagnostic method based on a relational graph convolutional network is proposed. This method increases intra-layer similarity relationships between students, exercises, and knowledge points to achieve information transfer at the tail node. A relational graph convolutional network is used to aggregate information from the head node to the tail node and information from different nodes, generating representations of students, exercises, and knowledge points. Finally, a diagnostic function is designed to predict student performance, thereby better diagnosing students' mastery of knowledge points. The method is implemented as follows:

[0056] Step 1: Obtain students' historical answer records and construct a heterogeneous graph:

[0057] Let S denote the student set, and S = {s1,...,s...} a ,...,s N}, where s a Let represent the a-th student, N represent the total number of students, 1≤a≤N; let E represent the problem set, and E={e1,...,e i ,...,e M}, where e i Let represent the i-th exercise, M represent the total number of exercises, 1 ≤ i ≤ M; let K represent the set of knowledge points, and K = {k1, ..., k}. j ,...,k T}, where k j Let K represent the j-th knowledge point, and K represent the total number of knowledge points, where 1 ≤ j ≤ T.

[0058] Each student's chosen exercises and the knowledge points they encompass are represented in a three-part graph. The relationships between the student's answers and the relationships between the knowledge points contained in the exercises are termed inter-level relationships. The set of interaction records between students and exercises is denoted as... Let the set of relationships between exercises and knowledge points be denoted as... in, s represents the a-th student a and the i-th exercise e i Interaction records, Represents the i-th exercise e iand the kth knowledge point k j The edges corresponding to the relationship between them.

[0059] Step 2: Initialize the student embedding representation, exercise embedding representation, and knowledge point embedding representation;

[0060] Step 2.1: After initializing the student set S using the Xavier method, the student representation set E is obtained. s ={e s1 ,...,e sa ,...,e sN}, where e sa s represents the a-th student a d-dimensional embedding representation;

[0061] Step 2.2: After initializing the problem set E using the Xavier method, the problem representation set E is obtained. e ={e e1 ,...,e ei ,...,e eM}, where e ei Represents the i-th exercise e i d-dimensional embedding representation;

[0062] Step 2.3: After initializing the knowledge point set K using the Xavier method, the knowledge point representation set E is obtained. k ={e k1 ,...,e kj ,...,e kT}, where e kj Represents the j-th knowledge point k j The d-dimensional embedding representation.

[0063] Step 3: Construct a cognitive diagnostic model based on intra-layer similarity relationships, including: an intra-layer similarity relationship module, a node representation module for fusion relationships, and a diagnostic module; wherein, the node representation module for fusion relationships includes: an L-layer relationship graph convolutional layer.

[0064] Step 3.1: Construct intra-layer similarity relationships based on the intra-layer similarity relationship module:

[0065] Step 3.1.1: Using equation (1), construct the s of the a-th student who has done the exercises together. a And the bth student s b Similarity between Sim s (s a ,s b This allows us to obtain the similarity matrix Sim between each student. s :

[0066]

[0067] In equation (1), s represents the a-th student a A collection of interaction records with completed exercises. s represents the b-th student b The set of interaction records with those who have completed exercises, 1≤a≤N, 1≤b≤N, a≠b;

[0068] Step 3.1.2: Obtain the threshold λ using equation (2):

[0069] λ=α×AVE(Sim s (2)

[0070] In equation (2), α is the set coefficient; AVE indicates the averaging operation;

[0071] Step 3.1.3, when Sim s (s a ,s b When ) > λ, for the a-th student s a And the bth student s b By establishing the edges corresponding to the similarity relationships between students, we can obtain the set of relationships between students;

[0072] Step 3.1.4: Construct the p-th exercise e using equation (3). p and the qth exercise e q Similarity between Sim e (e p ,e q This allows us to obtain the similarity matrix Sim between each exercise. e ;

[0073]

[0074] In equation (3), This represents the p-th exercise e. p A collection containing knowledge points e represents the q-th exercise. q A set containing knowledge points, where 1 ≤ p ≤ M, 1 ≤ q ≤ M, and p ≠ q;

[0075] Step 3.1.5: Construct the u-th knowledge point k using equation (4). u and the vth knowledge point k v Similarity between Sim k (k u ,k v ) u,v Thus, the similarity matrix Sim between knowledge points is obtained. k ;

[0076]

[0077] In equation (4), n u,v This indicates that the answer to the u-th knowledge point k was correct. u After correctly answering the v-th knowledge point k v The number of times, 1≤u≤T, 1≤v≤T, u≠v; n u,t This indicates that the answer to the u-th knowledge point k is correct. u After correctly answering the t-th knowledge point k t The number of times, 1 < t < T;

[0078] According to Sim s Sim e Sim k We obtain the student-student hierarchical relation set R. s Exercises - The set of in-level relations R of exercises e Knowledge Point-Knowledge Point Hierarchical Relationship Set R k .

[0079] Step 3.2: Message passing is performed by the node representation module based on the fusion relationship;

[0080] Step 3.2.1: Define the current update count as l, and initialize l = 0;

[0081] Define the a-th student s a The feature vector under the l-th update The i-th exercise e i The feature vector under the l-th update The j-th knowledge point k j The feature vector under the l-th update and initialize

[0082] Step 3.2.2: The relationship graph convolution uses equations (5), (6), and (7) to obtain the a-th student s. a User feature vector under the (l+1)th update The i-th exercise e i The problem embedding representation under the (l+1)th update The j-th knowledge point k j Embedded representation of knowledge points under the (l+1)th update

[0083]

[0084]

[0085]

[0086] In equations (5), (6), and (7), s represents the a-th student a In relation r s ∈R s ∩R se The set of relationships between the current node and its neighboring nodes; For the a-th student s a The representation of neighboring nodes under the (l+1)th update, W sr (l) s represents the a-th student a The weights of neighboring nodes in the l-th update, W s0 (l) s represents the a-th student a The weight of the self-embedded representation under the l-th update; Represents the i-th exercise e i In relation r e ∈R e ∩R se ∩R ek The set of relationships between the current node and its neighboring nodes; For the i-th exercise e i The representation of neighboring nodes under the l-th update, W er (l) Represents the i-th exercise e i The weights of neighboring nodes in the l-th update, W e0 (l) Represents the i-th exercise e i The weights of the self-embedded representation under the l-th update; Represents the j-th knowledge point k j In relation r k ∈R k ∩R ek The set of relationships between the current node and its neighboring nodes; For the j-th knowledge point k j The representation of neighboring nodes under the l-th update, W kr (l) Represents the j-th knowledge point k j The weights of neighboring nodes in the l-th update, W k0 (l) Represents the j-th knowledge point k j The weights of the self-embedding representation under the l-th update; σ represents the LeakyReLU activation function;

[0087] Step 3.2.3: After assigning l+1 to l, return to step 3.2.2 and execute sequentially until l>L, thus obtaining the a-th student s updated in the L-th iteration. a The embedding representation is denoted as e′ sa The i-th exercise ei The embedding representation is denoted as e′ ei The j-th knowledge point k j The embedding representation is denoted as e′ kj .

[0088] Step 3.3, Processing of the cognitive diagnosis module:

[0089] Step 3.3.1: Use equation (8) to obtain the a-th student s. a The embedded representation of e ca , used to represent the a-th student s a Cognitive state:

[0090] e ca =FC(e′) sa ·e′ k (8)

[0091] In equation (8), e′ k The matrix is ​​formed by embedding all knowledge points, where σ(·) represents the sigmoid activation function; FC represents a fully connected layer.

[0092] Step 3.3.2: Use equation (9) to predict the a-th student s. a Answer the i-th exercise correctly. i The score y ai :

[0093] y ai =L(W(e) ca -e′ ei (9)

[0094] In equation (9), W(·) is a pooling layer that reduces the dimensionality of the student’s cognitive state and the exercise representation; L(·) is a scoring function that uses a fully connected layer.

[0095] Step 4: Construct the cross-entropy loss function Loss using equation (10):

[0096]

[0097] In equation (10), y ai It is the a-th student s a Answer the i-th exercise e i The predicted results, r ai It is the a-th student s a Answer the i-th exercise e i The actual results.

[0098] Step 5: Use the Adam optimizer to train the cognitive diagnostic model and minimize the total loss function Loss to update the model parameters until convergence, thereby obtaining the trained cognitive diagnostic model. Based on the final embedding representation of the student nodes, use the trained cognitive diagnostic model to determine the corresponding student's mastery of different knowledge points.

[0099] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0100] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0101] Example:

[0102] To verify the effectiveness of this method, this invention uses three publicly available datasets commonly used in intelligent education research: ASSISTMents, Mooper, and JunYi. Exercises that do not contain knowledge point information are removed from each dataset.

[0103] For the student performance prediction task, this invention uses ACC and AUC as evaluation criteria. Five methods were selected for performance comparison: IRT, DINA, NeuralCDM, RCD, and KSCD. Specifically, Table 1 shows the experimental results on the above datasets, demonstrating that the proposed method outperforms the compared methods in both ACC and AUC. In the table, ISCD-S indicates that only intra-layer similarity relationships are constructed for students, while no intra-layer similarity relationships are constructed for exercises and knowledge points; ISCD-E and ISCD-K indicate that only intra-layer similarity relationships are constructed for exercises and knowledge points, respectively.

[0104] Table 1 shows the results of multiple algorithms on three datasets.

[0105]

[0106] To further verify the improvement of the proposed method in tail node diagnosis performance, Table 2 shows the diagnosis results of the model with and without student intra-layer relations on the ASSIST dataset for each interval of the number of questions answered.

[0107] Table 2. Distribution of Diagnostic Accuracy Rates for Students with Different Number of Questions Completed on the ASSIST Dataset

[0108]

[0109] As can be seen from Table 3:

[0110] 1) As the number of questions completed increases, the student diagnostic accuracy of NCD, RCD, and the model of this invention gradually improves. In the tail nodes [0-40] where the number of questions completed is relatively small, the average diagnostic ability of the three methods is only 60.78%, while for the head nodes [500-1000] where the number of questions completed is relatively large, the average student diagnostic accuracy of the three models is 74.68%, indicating that the diagnostic performance of the tail nodes is indeed lower than that of the head nodes.

[0111] 2) Compared to NCD and RCD, ISCD-S showed superior diagnostic performance for tail nodes in middle school students, improving accuracy by 3-6%. These results indicate that utilizing intra-layer relationships among students can effectively improve diagnostic accuracy, especially for a small number of students taking the test. It is evident that introducing intra-layer relationships among students can obtain more information about their test-taking process, thereby improving the modeling of students' cognitive states. However, for head nodes, the performance of the ISCD-S method is lower than that of NCD and RCD. This is mainly because head nodes themselves contain relatively abundant information, and some noise is introduced during the construction of intra-layer relationships.

[0112] To further verify the necessity of simultaneously adding hierarchical relationships among students, exercises, and knowledge points, the experiment shown in Table 3 was designed in this embodiment. Here, Mooper-S indicates that in the original Mooper dataset, only students have a long-tail distribution, while exercises and knowledge points do not; Mooper-SK indicates that some exercises without knowledge point associations were added to the original Mooper dataset, thus making both students and knowledge points have long-tail distributions; ASSIST-SE indicates that in the original ASSIST dataset, both students and exercises have long-tail distributions.

[0113] Table 3. Diagnostic accuracy with different relational models

[0114]

[0115]

[0116] 1) Adding intra-layer similarity relationships to student nodes is effective when they have tail nodes. A) In the Mooper-S dataset, only students exhibit a long-tail distribution; exercises and knowledge points do not. Compared to ISCD, ISCD-S improves accuracy by 2% on the Mooper-S dataset and by 2.9% in tail node diagnosis, indicating that adding similarity relationships between student head nodes and tail nodes can transfer problem-solving information from head nodes to tail nodes, thereby improving accuracy. B) In the ASSIST-SE dataset, both students and exercises have long tails. Compared to ISCD, ISCD-S improves accuracy by 2.7% for all students on the ASSIST dataset and by 6.7% in tail node diagnosis, indicating that adding intra-layer similarity relationships to student nodes is effective when students have a long-tail distribution.

[0117] 2) When only student-related long-tailed nodes exist, and exercise knowledge points do not have long-tailed nodes, simply increasing the intra-layer relationships between exercises can improve performance, but not as significantly as simply increasing the intra-layer relationships between students. On the Mooper-S dataset, compared to ISCD, ISCD-E improved by 1.6%, while ISCD-S improved by 3%. This is because, during the RGCN process, the intra-layer relationships between exercises aggregate more exercise information onto the student nodes, thereby alleviating the sparsity of student tail nodes. Table 3 yields a similar conclusion: when student exercises have tail nodes but knowledge points do not, simply increasing the intra-layer relationships between knowledge points can also improve overall diagnostic accuracy.

[0118] 3) When only the student knowledge points are long-tailed, and the problem knowledge points are not, adding multiple intra-layer relationships has little effect on improvement. Compared with ISCD, ISCD-S improved the diagnostic accuracy of all students on the Mooper-S dataset by 2.1%, while ISCD-SE only improved the diagnostic accuracy of all students on ISCD-S by 0.6%, and the diagnostic accuracy of tail nodes improved by only 0.3%, which is basically the same. This indicates that there is information overlap in the superposition of multiple intra-layer relationships, making it difficult to obtain significant improvement.

[0119] Table 3 summarizes the experiments using the long-tail distribution of students as an example. Similar conclusions were reached for exercises and knowledge points, and will not be elaborated upon due to space limitations. It is worth noting that the long tail of knowledge points weakens the influence of intra-level relationships within knowledge points. When knowledge points have a long tail, increasing intra-level relationships has a relatively small improvement on diagnostic results. Mooper-SK exhibits both student and knowledge point long tails, and ISCD-K shows only a 0.2% improvement compared to ISCD. This may be because knowledge points are relatively independent compared to students and exercises, making it difficult to construct accurate similarity relationships. In summary, the method proposed in this invention significantly outperforms many compared methods, thus proving the feasibility of the proposed method.

[0120] In summary, the methods proposed in this invention are significantly superior to the numerous comparative methods, thus proving the feasibility of the proposed methods and verifying the necessity of constructing multiple inter-layer similarity relationships when various long-tail distributions of students, exercises, and knowledge point nodes coexist.

Claims

1. A cognitive diagnostic method based on intra-layer similarity relationships, characterized in that, The procedure is as follows: Step 1: Obtain students' historical answer records and construct a heterogeneous graph: Let S denote the student set, and S = {s1,...,s...} a ,...,s N }, where s a Let represent the a-th student, N represent the total number of students, 1≤a≤N; let E represent the problem set, and E={e1,...,e i ,...,e M }, where e i Let represent the i-th exercise, M represent the total number of exercises, 1 ≤ i ≤ M; let K represent the set of knowledge points, and K = {k1, ..., k}. j ,...,k T }, where k j Let K represent the j-th knowledge point, and K represent the total number of knowledge points, where 1 ≤ j ≤ T; Each student's chosen exercises and the knowledge points they encompass are represented in a three-part graph. The relationships between the student's answers and the relationships between the knowledge points contained in the exercises are termed inter-level relationships. The set of interaction records between students and exercises is denoted as... Let the set of relationships between exercises and knowledge points be denoted as . in, s represents the a-th student a and the i-th exercise e i Interaction records, Represents the i-th exercise e i and the kth knowledge point k j The edges corresponding to the relationship between them; Step 2: Initialize the student embedding representation, exercise embedding representation, and knowledge point embedding representation; Step 2.1: After initializing the student set S using the Xavier method, the student representation set E is obtained. s ={e s1 ,...,e sa ,...,e sN }, where e sa s represents the a-th student a d-dimensional embedding representation; Step 2.2: After initializing the problem set E using the Xavier method, the problem representation set E is obtained. e ={e e1 ,...,e ei ,...,e eM }, where e ei Represents the i-th exercise e i d-dimensional embedding representation; Step 2.3: After initializing the knowledge point set K using the Xavier method, the knowledge point representation set E is obtained. k ={e k1 ,...,e kj ,...,e kT }, where e kj Represents the j-th knowledge point k j d-dimensional embedding representation; Step 3: Construct a cognitive diagnostic model based on intra-layer similarity relationships, including: an intra-layer similarity relationship module, a node representation module for fusion relationships, and a diagnostic module; wherein, the node representation module for fusion relationships includes: an L-layer relationship graph convolutional layer; Step 3.1: Construct intra-layer similarity relationships based on the intra-layer similarity relationship module to obtain the student-student intra-layer relationship set R. s Exercises - The set of in-level relations R of exercises e Knowledge Point-Knowledge Point Hierarchical Relationship Set R k ; Step 3.2: Based on the node representation module of the fusion relationship, message passing is performed to obtain the a-th student s updated in the L-th step. a The embedding representation is denoted as e′ sa The i-th exercise e i The embedding representation is denoted as e′ ei The j-th knowledge point k j The embedding representation is denoted as e′ kj ; Step 3.3, Processing of the cognitive diagnosis module: Step 3.3.1: Use equation (8) to obtain the a-th student s. a The embedded representation of e ca , used to represent the a-th student s a Cognitive state: e ca =FC(e′ sa ·e′ k ) (8) In equation (8), e′ k The matrix is ​​formed by embedding all knowledge points, where σ(·) represents the sigmoid activation function; FC represents a fully connected layer. Step 3.3.2: Use equation (9) to predict the a-th student s. a Answer the i-th exercise correctly. i The score y ai : y ai =L(W(e ca -have been' ei )) (9) In equation (9), W(·) is a pooling layer that reduces the dimensionality of the student's cognitive state and the exercise representation; L(·) is a scoring function using a fully connected layer. Step 4: Construct the cross-entropy loss function Loss using equation (10): In equation (10), y ai It is the a-th student s a Answer the i-th exercise e i The predicted results, r ai It is the a-th student s a Answer the i-th exercise e i The actual results; Step 5: Use the Adam optimizer to train the cognitive diagnostic model and minimize the total loss function Loss to update the model parameters until convergence, thereby obtaining the trained cognitive diagnostic model. Based on the final embedding representation of the student nodes, use the trained cognitive diagnostic model to determine the corresponding student's mastery of different knowledge points.

2. The cognitive diagnosis method based on intra-layer similarity relationships according to claim 1, characterized in that, Step 3.1 is performed as follows: Step 3.1.1: Using equation (1), construct the s of the a-th student who has done the exercises together. a And the bth student s b Similarity between Sim s (s a ,s b This allows us to obtain the similarity matrix Sim between each student. s : In equation (1), s represents the a-th student a A collection of interaction records with completed exercises. s represents the b-th student b The set of interaction records with those who have completed exercises, 1≤a≤N, 1≤b≤N, a≠b; Step 3.1.2: Obtain the threshold λ using equation (2): λ=α×AVE(Yes s ) (2) In equation (2), α is the set coefficient; AVE indicates the averaging operation; Step 3.1.3, when Sim s (s a ,s b When ) > λ, for the a-th student s a And the bth student s b By establishing the edges corresponding to the similarity relationships between students, we can obtain the set of relationships between students; Step 3.1.4: Construct the p-th exercise e using equation (3). p and the qth exercise e q Similarity between Sim e (e p ,e q This allows us to obtain the similarity matrix Sim between each exercise. e ; In equation (3), This represents the p-th exercise e. p A collection containing knowledge points e represents the q-th exercise. q A set containing knowledge points, where 1 ≤ p ≤ M, 1 ≤ q ≤ M, and p ≠ q; Step 3.1.5: Construct the u-th knowledge point k using equation (4). u and the vth knowledge point k v Similarity between Sim k (k u ,k v ) u,v Thus, the similarity matrix Sim between knowledge points is obtained. k ; In equation (4), n u,v This indicates that the answer to the u-th knowledge point k was correct. u After correctly answering the v-th knowledge point k v The number of times, 1≤u≤T, 1≤v≤T, u≠v; n u,t This indicates that the answer to the u-th knowledge point k is correct. u After correctly answering the t-th knowledge point k t The number of times, 1 < t < T; According to Sim s Sim e Sim k We obtain the student-student hierarchical relation set R. s Exercises - The set of in-level relations R of exercises e Knowledge Point-Knowledge Point Hierarchical Relationship Set R k .

3. The cognitive diagnosis method based on intra-layer similarity relationships according to claim 2, characterized in that, Step 3.2 is performed as follows: Step 3.2.1: Define the current update count as l, and initialize l = 0; Define the a-th student s a The feature vector under the l-th update The i-th exercise e i The feature vector under the l-th update The j-th knowledge point k j The feature vector under the l-th update and initialize Step 3.2.2: The relationship graph convolution uses equations (5), (6), and (7) to obtain the a-th student s. a User feature vector under the (l+1)th update The i-th exercise e i The problem embedding representation under the (l+1)th update The j-th knowledge point k j Embedded representation of knowledge points under the (l+1)th update In equations (5), (6), and (7), s represents the a-th student a In relation r s ∈R s ∩R se The set of relationships between the current node and its neighboring nodes; For the a-th student s a The representation of neighboring nodes under the (l+1)th update, W sr (l) s represents the a-th student a The weights of neighboring nodes in the l-th update, W s0 (l) s represents the a-th student a The weight of the self-embedded representation under the l-th update; Represents the i-th exercise e i In relation r e ∈R e ∩R se ∩R ek The set of relationships between the current node and its neighboring nodes; For the i-th exercise e i The representation of neighboring nodes under the l-th update, W er (l) Represents the i-th exercise e i The weights of neighboring nodes in the l-th update, W e0 (l) Represents the i-th exercise e i The weights of the self-embedded representation under the l-th update; Represents the j-th knowledge point k j In relation r k ∈R k ∩R ek The set of relationships between the current node and its neighboring nodes; For the j-th knowledge point k j The representation of neighboring nodes under the l-th update, W kr (l) Represents the j-th knowledge point k j The weights of neighboring nodes in the l-th update, W k0 (l) Represents the j-th knowledge point k j The weights of the self-embedding representation under the l-th update; σ represents the LeakyReLU activation function; Step 3.2.3: After assigning l+1 to l, return to step 3.2.2 and execute sequentially until l>L, thus obtaining the a-th student s updated in the L-th iteration. a The embedding representation is denoted as e′ sa The i-th exercise e i The embedding representation is denoted as e′ ei The j-th knowledge point k j The embedding representation is denoted as e′ kj .

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the cognitive diagnostic methods of claims 1-3, and the processor is configured to execute the programs stored in the memory.

5. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of any one of the cognitive diagnostic methods described in claims 1-3.

Images