A student cognitive modeling method based on a class self-encoder type non-negative matrix co-factor decomposition

Abstract

The application provides a student cognitive modeling method based on a self-encoder type non-negative matrix collaborative factor decomposition, and comprises the following steps: obtaining historical answering data of students, and extracting question-knowledge point association information of test question texts; according to the answering data of the students and the question-knowledge point association information, building an encoding module to construct low-dimensional hidden space representations of the students, the questions and the knowledge points, and building a decoding module to represent the mastery degrees of the students on the knowledge points; fusing the encoding module and the decoding module to complete the building of a student cognitive model; adopting a block coordinate descent projection gradient method with constraints and combining a Lipschitz step length solving strategy to perform model training; completing a learning achievement prediction task according to the low-dimensional hidden space representations of the questions and the students, and completing a knowledge level diagnosis task according to the mastery degrees of the students on the knowledge points obtained through training. The application can realize the accurate prediction of learning achievements and the accurate diagnosis of knowledge levels of students simultaneously.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of educational data mining, and particularly relates to a student cognitive modeling method based on a non-negative matrix co-factor decomposition framework. BACKGROUND

[0002] Student cognitive modeling aims to evaluate the stage cognitive state of students in the learning process, and specifically relates to student learning achievement prediction tasks and knowledge level diagnosis tasks, modeling and evaluation scenarios such as shown in FIG. 1. Establishing an accurate and reliable student cognitive model is not only an important means to realize personalized learning, but also provides important support for the construction of a high-quality education system. Figure 2

[0003] Around the basic task of student cognitive modeling, related analysis methods can be divided into two classic routes. The first route is the educational psychology method represented by cognitive diagnosis models (CDMs), including continuous CDMs typified by item response theory and discrete CDMs typified by deterministic inputs, noisy "and" gate. Although cognitive diagnosis models can perform knowledge level diagnosis tasks, they are heavily dependent on pre-defined cognitive variables driven by expert experience, which leads to their difficulty in fully capturing the cognitive mode of students, and in turn weak performance in student achievement prediction tasks.

[0004] The second route is the educational data mining technology centered on matrix decomposition methods. Compared with cognitive diagnosis models, matrix decomposition methods can directly decompose student answer data into a series of cognitive variables and perform student achievement prediction tasks. Further, to improve the explainability of the decomposed cognitive variables for knowledge level diagnosis tasks, a scalable non-negative matrix co-factor decomposition model (SNMCF) uses a cover function to estimate the knowledge level of students. However, the SNMCF model separates student achievement prediction tasks and knowledge level diagnosis tasks, i.e., only focuses on improving the prediction effect of student achievement in the modeling process, and does not show the mastery of knowledge points by students, and in turn lacks the consistency of simulating the high and low of student knowledge level and the good and bad of answer results, resulting in weak diagnosis ability and rough diagnosis granularity.

[0005] Accordingly, the educational psychology method can provide more accurate knowledge level diagnosis results, but the student achievement prediction accuracy is not enough; on the contrary, the educational data mining technology can provide more accurate student achievement prediction results, but the diagnosis effect on student knowledge level is not good. ​SUMMARY

[0006] The present application aims to provide a student cognitive modeling method based on non-negative matrix co-factor decomposition framework, which can break through the bottleneck that the educational data mining technology cannot deeply understand the knowledge level of students while ensuring high-precision prediction of student performance, comprehensively diagnose the current knowledge point mastery of students, and predict their future performance.

[0007] To achieve the above object, the technical scheme of the present application is as follows: a student cognitive modeling method based on non-negative matrix co-factor decomposition of self-encoder type, comprising the following steps:

[0008] Step 1: Obtain the historical answer data of students, and extract the question-knowledge point association information of the test text;

[0009] Step 2: According to the answer data of students and the question-knowledge point association information, build an encoding module to construct the low-dimensional hidden space representation of students, questions, and knowledge points;

[0010] Step 3: Use the low-dimensional hidden space information of students, questions, and knowledge points to build a decoding module to represent the mastery of knowledge points by students;

[0011] Step 4: Fuse the encoding module and the decoding module to complete the construction of the student cognitive model;

[0012] Step 5: Use the block coordinate descent projection gradient method with constraints and combine the Lipschitz step size solving strategy to train the model;

[0013] Step 6: Complete the learning achievement prediction task according to the low-dimensional hidden space representation of questions and students, and complete the knowledge level diagnosis task according to the mastery of knowledge points by students obtained by training.

[0014] Further, the step 1 is specifically:

[0015] Given a question set and a student set There are N questions and M students, and the student-question answer matrix stores all the answer records of students, where X nm =1 indicates that the student St m correctly answers the question Ex n , and if it is wrong, X nm =0;

[0016] Given K knowledge points, construct a question-knowledge point association matrix and Q nk =1 indicates that the question Ex n contains the knowledge point Kc k, and vice versa.

[0017] Further, the step two is specifically:

[0018] According to the student-exercise answer matrix X and the exercise-knowledge point correlation matrix Q, based on the non-negative matrix co-factor decomposition technology, an encoding module is constructed to represent the student, the exercise, and the knowledge point in a low-dimensional hidden space with the following optimization objective function:

[0019]

[0020] Wherein, represents the Frobenius norm, which is also called the Frobenius norm; ⊙ represents the Hadamard product, which makes two matrices with the same dimension produce another matrix with the same dimension, where each element of the new matrix is the product of the elements of the same position of the original two matrices. and respectively represent the exercise feature matrix, the student ability matrix, and the knowledge point demand matrix in the low-dimensional hidden space, wherein T represents the dimension of the (self-defined) low-dimensional hidden space, and each low-dimensional matrix is regarded as a theme skill in each dimension in the low-dimensional direction;

[0021] For the above encoding module: first, the inner product of the exercise feature matrix E and the student ability matrix U is used as the low-rank approximation of X based on the Frobenius norm; since there are missing data in the matrix X, a weight matrix is introduced to represent that the model only focuses on the part of X with numerical values. In W, the value of each element is 0 or 1, W nm = 1 if and only if X nm has an answer record; secondly, the exercise-knowledge point correlation matrix is introduced, wherein B nk represents the importance of the knowledge point Kc k to the exercise Ex n , and the larger the value is, the deeper the correlation is, the low-rank approximation of the matrix B is made by using the exercise feature matrix E and the knowledge point demand matrix V based on the Frobenius norm, and the Q matrix is used as a sparse constraint; finally, the non-negativity requirement for all variables is considered in the constraint term, which ensures that the better the student masters the knowledge point, the more helpful it is to correctly answer the related exercises.

[0022] Further, the step three is specifically:

[0023] The student-knowledge point proficiency matrix A is defined, wherein the non-negative element A km represents the proficiency of the student St m to the knowledge point Kck the specific mastery of the student, and introduce the difficulty vector of the exercises the difficulty value of the simulated exercises, wherein μ n represents the difficulty coefficient of the exercise Ex n ; based on the exercise-knowledge point correlation matrix B, the student-knowledge point proficiency matrix A, and the exercise difficulty vector M, a decoding module with the following optimization objective function is constructed to represent the mastery of the student to the knowledge points:

[0024]

[0025] γ B represents the weight factor balancing the objective function (the first term) and the regularization term (the second term), represents the L2-norm, for the first term of the above formula, l is the log-likelihood function with cross-entropy loss, which is defined as follows:

[0026]

[0027] wherein contains the index positions of the non-missing data in the matrix X; B n: and A :m respectively represent the knowledge point correlation vector of the exercise Ex n and the knowledge point proficiency vector of the student St m ; Φ(x) represents the standardized inverse link function, which is used for nonlinear mapping to simulate the answering process of the student:

[0028]

[0029] Here, is the probability density function with standard normal distribution.

[0030] Further, the construction of the student cognitive model can be transformed into solving the optimization problem with the following constraints:

[0031]

[0032] In the formula, O AF represents the objective function of the student cognitive model.

[0033] Further, the specific steps of model training in step five are as follows:

[0034] (a) initialize the exercise feature matrix E, the student ability matrix U, the knowledge point requirement matrix V, and the exercise difficulty vector M;

[0035] (b) calculate the objective function

[0036] (c) set the maximum number of external iterations;

[0037] (d) For each outer iteration, update the problem-knowledge point correlation block matrix B in turn n: (1≤n≤N), student ability block matrix U :m (1≤m≤M), problem feature matrix E, knowledge point requirement matrix V, and problem difficulty vector μ;

[0038] (e) Calculate the new objective function again, and compare the difference with the objective function of the last round. If it is less than the pre-set threshold, stop iteration and output all parameters, otherwise continue to execute (d).

[0039] Further, step (d) updates the problem-knowledge point correlation block matrix B n: , student ability block matrix U :m , problem feature matrix E, knowledge point requirement matrix V, and problem difficulty vector μ n according to the objective sub-function, gradient information and Lipschitz constant of the parameters B n: , student ability block matrix U :m , problem feature matrix E, knowledge point requirement matrix V, and problem difficulty value μ n ∈M, the iterative update formula is as follows:

[0040]

[0041]

[0042] Where, [x]+=max(∈,x),∈=10 -15 to ensure the non-negativity of the output result; superscripts (l) and (l+1) represent the (l)th and (l+1)th iterations, respectively, L(·) represents the Lipschitz constant, represents the gradient information.

[0043] Further, the specific calculation of the objective sub-function, gradient information and Lipschitz constant of the parameters B n: , student ability block matrix U :m , problem feature matrix E, knowledge point requirement matrix V, and problem difficulty vector μ n is as follows:

[0044] (i) Objective sub-function:

[0045]

[0046] Where,

[0047] (ii) Gradient information:

[0048]

[0049]

[0050] wherein

[0051] (iii) the Lipschitz constant of the function

[0052]

[0053] wherein, L p denotes the Lipschitz constant of the function , which can be derived as 1; Φ'(x) denotes the first derivative of Φ(x); σ1(·) denotes the maximum value of its elements; J M,N denotes a matrix of dimension M x N with all elements equal to 1.

[0054] Further, the step six is specifically: fitting the student's answering results by training the target function, obtaining the exercise feature matrix E, the student ability matrix U, and the student-knowledge point proficiency matrix A, completing the learning achievement prediction task by using E and U, and completing the knowledge level diagnosis task by using A.

[0055] Compared with the prior art, the present application has the following beneficial effects:

[0056] (a) At the knowledge level, the present application makes full use of the student's answer records to the exercises, which are regarded as the input of the encoding module and the output of the decoding module, and introduces the student-knowledge point proficiency matrix to explicitly represent the knowledge level of the students, breaking the obstacle that the traditional education data mining method cannot effectively simulate the knowledge level of the students.

[0057] (b) At the exercise level, considering that the Q matrix only roughly represents the association between the exercises and the knowledge points, and does not reflect the association strength between them, which is not only very important for the cognitive analysis of the students, but also plays a key guiding role in the personalized recommendation task in the education field. Therefore, the present application introduces the exercise-knowledge point correlation matrix to refine the association strength of different knowledge points to any one exercise.

[0058] (c) At the answer level, the present application introduces the inverse link function to make a nonlinear mapping between the knowledge feature vector and the exercise feature vector in the low-dimensional space to simulate the answering process of the students, so as to better adapt to the complexity in the cognitive process of the students, including the nonlinear change of the learning curve and the stage of the concept mastery. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A student cognitive modeling method flow chart based on the class self-encoder type non-negative matrix co-factor decomposition is provided for the embodiments of the present application;

[0060] Figure 2 to build a cognitive model for a student;

[0061] Figure 3 A student cognitive model based on a class auto-encoder type non-negative matrix co-factor decomposition is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0062] The technical solutions of the present application will be specifically described below with reference to the accompanying drawings. Figures 1-3 The technical solutions of the present application will be specifically described below with reference to the accompanying drawings.

[0063] The present application proposes a student cognitive model based on a class auto-encoder type non-negative matrix co-factor decomposition, as shown in Figure 1 , 3 The present application proposes a student cognitive model based on a class auto-encoder type non-negative matrix co-factor decomposition, as shown in

[0064] (1) Step one (data acquisition): Obtain the historical answer data (correct, wrong or score) of the students and the question-knowledge point association information (Q matrix). Given a question set and a student set (there are N questions and M students), use the student-question answer matrix to store all the answer records of the students, where X nm = 1 indicates that the student St m correctly answers the question Ex n , and if the answer is wrong, X nm = 0. In addition, given K knowledge points, construct the question-knowledge point association matrix and Q nk = 1 indicates that the question Ex n contains the knowledge point Kc k ; otherwise, it does not.

[0065] (2) Step two (coding module): According to the student-question answer matrix X and the question-knowledge point association matrix Q, based on the non-negative matrix co-factor decomposition technology, construct a coding module with the following optimization objective function to represent the students, questions and knowledge points in a low-dimensional hidden space:

[0066]

[0067] where, denotes the Frobenius norm, which is also called the Frobenius norm; ⊙ denotes the Hadamard product, which makes two matrices with the same dimension produce another matrix with the same dimension, where each element of the new matrix is the product of the elements of the same position of the original two matrices, and respectively represent the item feature matrix, the student ability matrix, and the knowledge point demand matrix in the low-dimensional latent space. Among them, T represents the dimension of the (self-defined) low-dimensional latent space, and for each low-dimensional matrix, each dimension in the low-dimensional direction is regarded as a topic skill.

[0068] For the above encoding module: first, the inner product of the item feature matrix E and the student ability matrix U is used as the low-rank approximation of X based on the Frobenius norm; since there are missing data in the matrix X, a weight matrix W is introduced to represent that the model only focuses on the part of X with numerical values. In W, the value of each element is either 0 or 1, W nm = 1 if and only if X nm has an answer record; second, the item-knowledge point correlation matrix B is introduced, where B nk represents the importance of the knowledge point Kc k to the item Ex n , and the larger the value, the deeper the correlation, and the item feature matrix E and the knowledge point demand matrix V are used to make a low-rank approximation of the matrix B, and the Q matrix is used as a sparse constraint. Finally, the constraint term considers the non-negativity requirement for all variables, which ensures that the better the student masters the knowledge points, the more helpful it is to correctly answer related items.

[0069] (3) Step three (decoding module): define the student-knowledge point proficiency matrix A , where the non-negative element A km represents the specific mastery of the student St m to the knowledge point Kc k , and the item difficulty vector M = μ is introduced to simulate the difficulty value of the item, where μ n represents the difficulty coefficient of the item Ex n . Based on the item-knowledge point correlation matrix B, the student-knowledge point proficiency matrix A, and the item difficulty vector M, a decoding module with the following optimization objective function is constructed to represent the mastery of knowledge points by students:

[0070]

[0071] γ B represents the weight factor that balances the objective function (the first term) and the regularization term (the second term), represents the L2-norm, for the first term of the above formula, l is the log-likelihood function with cross-entropy loss, defined as follows:

[0072]

[0073] where contains the index positions of the non-missing data in the matrix X. Bn: and A :m respectively represent the knowledge point correlation vector of the exercise Ex n and the knowledge point proficiency vector of the student St n . Φ(x) represents a normalized inverse link function, which is used for nonlinear mapping to simulate the answering process of the student:

[0074]

[0075] Here, is a probability density function with a standard normal distribution.

[0076] (4) Step four (module fusion): fuse the encoding module and the decoding module, and the constructed student cognitive model can be converted into an optimization problem with the following constraints:

[0077]

[0078] In the formula, indicates the objective function of the student cognitive model.

[0079] (5) Step five (model training): considering that the model to be solved is not a joint convex function of parameters E, U, V, and has a non-negative constraint; the exercise-knowledge point correlation matrix B has a block attribute, so the block coordinate descent projection gradient method with constraints is adopted in the present application. At the same time, considering the influence of the iteration speed, the present application proposes a step length solving strategy based on the Lipschitz constant. Specifically, regarding the objective sub-function of parameters B n: , U :m , E, V, and μ n , the gradient information and the Lipschitz constant are as follows:

[0080] (i) Objective sub-function:

[0081]

[0082] Wherein,

[0083] (ii) Gradient information:

[0084]

[0085] Wherein

[0086] (iii) Lipschitz constant:

[0087]

[0088] Wherein, L p represents the function Lipshitz constant, whose value can be derived as 1; Φ'(x) represents the first derivative of Φ(x); σ1(·) represents taking the maximum value of its elements; J M,N represents a matrix with dimension M x N and all elements equal to 1;

[0089] Based on the gradient information and the Lipshitz constant, the B n: For example, the iterative update formula can be described as follows (the rest of the parameters are similar) :

[0090]

[0091] Here [x] + = max(∈, x), ∈ = 10 -15 to ensure the non-negativity of the output result, the superscripts (l) and (l+1) represent the (l)th and (l+1)th iterations, respectively, L(·) represents the Lipshitz constant, represents the gradient information.

[0092] Specifically, the complete iteration process is as follows:

[0093] (a) Initialize the problem feature matrix E, the student ability matrix U, the knowledge point requirement matrix V, and the problem difficulty vector M;

[0094] (b) Calculate the objective function

[0095] (c) Set the maximum number of external iterations;

[0096] (d) For each round of external iteration, update the problem-knowledge point correlation block matrix B n: (1≤n≤N), the student ability block matrix U :m (1≤m≤M), the problem feature matrix E, the knowledge point requirement matrix V, and the problem difficulty vector M in turn;

[0097] (e) Calculate the new objective function again and compare the difference with the objective function of the last round. If it is less than the pre-set threshold, stop iteration and output all parameters, otherwise continue to execute (d).

[0098] (6) Step six (cognitive analysis): fit the student's answer results through the training objective function, obtain the problem feature matrix E, the student ability matrix U, and the student-knowledge point proficiency matrix A, use E and U to complete the learning achievement prediction task, and use A to complete the knowledge level diagnosis task.

[0099] The above is the preferred embodiment of the present application. Any changes made in accordance with the technical solutions of the present application, as long as the resulting functional effects do not exceed the scope of the technical solutions of the present application, are within the scope of protection of the present application.

Claims

1. A student cognitive modeling method based on autoencoder-like non-negative matrix co-factorization, characterized by: The following steps are involved: Step 1: Obtain students' historical answer data and extract the exercise-knowledge point association information in the test text; Step 2: Based on the students’ answer data and the relationship between exercises and knowledge points, an encoding module is built to construct a low-dimensional latent space representation of students, exercises, and knowledge points; Step 3: Use the low-dimensional latent space information of students, exercises, and knowledge points to build a decoding module to represent the students' mastery of the knowledge points; Step 4: Integrate the encoding module and the decoding module to complete the construction of the student cognitive model; Step 5: Use the constrained block coordinate descent projected gradient method and the Lipschitz step size solution strategy to train the model; Step 6: Complete the learning performance prediction task based on the exercises and the low-dimensional latent space representation of the students, and complete the knowledge level diagnosis task based on the students' mastery of the knowledge points obtained through training.

2. A student cognitive modeling method based on autoencoder-like non-negative matrix co-factorization according to claim 1, characterized in that: The step 1 is specifically as follows: Given a set of exercises Gather with students There are N exercises and M students, using the student-exercise answer matrix Store all students' answer records, where X nm =1 means student St m Correctly answer the Exercises n If the answer is wrong, then X nm =0; Given K knowledge points, construct the exercise-knowledge point association matrix And Q nk =1 means Exercise Ex n Contains knowledge point Kc k , otherwise it is not included.

3. A student cognitive modeling method based on autoencoder-like non-negative matrix co-factorization according to claim 2, characterized in that: The step 2 is specifically as follows: Based on the student-exercise answer matrix X and the exercise-knowledge point association matrix Q, and using non-negative matrix co-factorization technology, an encoding module with the following optimization objective function is constructed to represent students, exercises, and knowledge points in a low-dimensional latent space: stB≥0,E≥0,U≥0,V≥0, in, represents the Frobenius norm; ⊙ represents the Hadamard product, which allows two matrices with the same dimension to produce another matrix with the same dimension, where each element of the new matrix is ​​the product of the elements at the same position of the original two matrices; as well as They represent the exercise feature matrix, student ability matrix, and knowledge point requirement matrix in the low-dimensional latent space respectively. T represents the dimension of the customized low-dimensional latent space, and each low-dimensional matrix regards each dimension in its low-dimensional direction as a subject skill. For the above encoding module: First, based on the Frobenius norm, the inner product of the exercise feature matrix E and the student ability matrix U is used as a low-rank approximation of X. Since there are missing data in the matrix X, the weight matrix is ​​introduced To indicate that the model only focuses on the numerical part of X, in W, the value of each element is either 0 or 1, W nm =1 if and only if X nm There is a record of answering questions; secondly, introduce the exercise-knowledge point correlation matrix Among them B nk Represents knowledge point Kc k For Exercise Ex n The importance of the exercise is expressed in terms of the exercise feature matrix E and the knowledge point requirement matrix V. The Frobenius norm is used to make a low-rank approximation of the matrix B, and the Q matrix is ​​used as a sparse constraint. Finally, the constraint terms take into account the non-negativity requirements for all variables to ensure that the better the students master the knowledge points, the more helpful it is to correctly answer related exercises.

4. A student cognitive modeling method based on autoencoder-like non-negative matrix co-factorization according to claim 3, characterized in that: The step three is specifically as follows: Defining the Student-Knowledge Point Proficiency Matrix Among them, the non-negative elements A km Indicates student St m Knowledge point Kc k Specific mastery of the problem, and introduce the difficulty vector of the problem The difficulty value of the simulation exercise, where μ n Representation Exercises n The difficulty coefficient is: Based on the exercise-knowledge point association matrix B, the student-knowledge point proficiency matrix A, and the exercise difficulty vector M, a decoding module with the following optimization objective function is constructed to represent the student's mastery of the knowledge point: γ B represents the weight factor that balances the objective function and the regularization term, represents the L2-norm. For the first term in the above formula, is the log-likelihood function with cross-entropy loss, defined as follows: in Contains the index position of the non-missing data in matrix X; B n: and A :m Respectively represent Exercise Ex n The knowledge point correlation vector and student St m The knowledge point proficiency vector of ; Φ(x) represents the standardized inverse link function, which is used as a nonlinear mapping to simulate the student's answering process: here, is the probability density function with a standard normal distribution.

5. The student cognitive modeling method based on autoencoder-like non-negative matrix co-factorization according to claim 4 is characterized in that: The construction of the student cognitive model can be transformed into solving an optimization problem with the following constraints: stB≥0,E≥0,U≥0,V≥0. Where, Represents the objective function of the student cognitive model.

6. A student cognitive modeling method based on autoencoder-like non-negative matrix co-factorization according to claim 5, characterized in that: The specific steps of model training in step 5 are as follows: (a) Initialize the exercise feature matrix E, the student ability matrix U, the knowledge point requirement matrix V, and the exercise difficulty vector M; (b) Calculate the objective function (c) Set the external maximum number of iterations; (d) For each round of external iteration, update the exercise-knowledge point association block matrix B in sequence n: (1≤n≤N), student ability block matrix U :m (1≤m≤M), exercise feature matrix E, knowledge point requirement matrix V, and exercise difficulty vector M; (e) Calculate the new objective function again and compare the difference with the objective function of the previous round. If it is less than the preset threshold, stop the iteration and output all parameters. Otherwise, continue to execute (d).

7. The student cognitive modeling method based on autoencoder-like non-negative matrix co-factorization according to claim 6 is characterized in that: Step (d) is based on parameter B n: 、U :m , E, V, and μ n ∈M's target subfunction, gradient information and Lipschitz constant, set the exercise-knowledge point association block matrix B n: , student ability block matrix U :m , exercise feature matrix E, knowledge point requirement matrix V, and exercise difficulty value μ n The iterative update formula of ∈M is as follows: Where [x] + =max(∈,x),∈=10 -15 To ensure the non-negativity of the output results; the superscripts (l) and (l+1) represent the (l)th and (l+1)th iterations respectively, and L(·) represents the Lipschitz constant. Represents gradient information.

8. The student cognitive modeling method based on autoencoder-like non-negative matrix co-factorization according to claim 7 is characterized in that: The parameter B n: 、U :m , E, V, and μ n The specific calculation of the objective subfunction, gradient information and Lipschitz constant is: (i) Target subfunction: in, (ii) Gradient information: in (iii) Lipschitz constant: Among them, L p Representation function The Lipschitz constant, whose value can be derived as 1; Φ'(x) represents the first-order derivative of Φ(x); σ1(·) represents the maximum value of its elements; J M,N Represents a matrix of dimension M×N with all elements set to 1.

9. The student cognitive modeling method based on autoencoder-like non-negative matrix co-factorization according to claim 6 is characterized in that: The specific steps of step six are as follows: fitting the students' answer results through the training objective function to obtain the exercise feature matrix E, the student ability matrix U, and the student-knowledge point proficiency matrix A, using E and U to complete the learning performance prediction task, and using A to complete the knowledge level diagnosis task.

Images