An adaptive cognitive diagnosis method based on double-layer confrontation

Abstract

The application discloses a kind of adaptive cognitive diagnosis method based on double-layer confrontation, is through the way of adaptive learning to improve the precision of student cognitive diagnosis, its steps include:1, based on vector embedding method obtains the vector embedding representation of proficiency of knowledge point, exercise difficulty and exercise difference degree, and the target vector embedding representation is adaptively learned;2, the interaction of vector embedding representation is carried out, and adaptive learning is carried out after two layers of fully connected layer;3, the loss function design of double-layer confrontation adaptive cognitive diagnosis network is carried out, and the parameter of network is updated;4, the whole double-layer confrontation adaptive cognitive diagnosis network is optimized using cross-entropy loss.This application can realize adaptive cognitive diagnosis in the scene of missing problem-solving record by using other scenes, such as different students, different exercise records, so as to realize more accurate student performance prediction.

Description

Technical Field

[0001] This invention relates to the fields of machine learning, artificial intelligence, and intelligent education technology, specifically an adaptive cognitive diagnosis method based on two-layer adversarial approaches. Background Technology

[0002] Cognitive diagnostics (CD) is considered a key task in smart education applications. This task models students' knowledge mastery, or cognitive state, based on their problem-solving records and the relationships between problems and knowledge points. Cognitive diagnostics provides a foundation for downstream tasks in smart education, such as student assessment and learning resource recommendation. Existing cognitive diagnostic methods can be broadly categorized into two types: statistical methods and deep neural network-based methods. The former, based on Item Response Theory (IRT), Multidimensional IRT (MIRT), and Deterministic Input, Noise and Gate (DINA) methods, uses hand-designed linear functions to simulate the interaction between students and problems. The latter, based on neural networks, models complex interaction patterns between students and problems while attempting to incorporate richer contextual features and prior relationships between knowledge points to enhance student and problem representation learning.

[0003] CD tasks require a large number of student response records to train diagnostic models, and they also require the training and test sets to have the same distribution. However, in application scenarios, due to various factors, it is often difficult to obtain sufficient and identically distributed logs. The differences between the source and target domains of CD tasks manifest in multiple ways; not only may the students, exercises, or knowledge points differ, but the students' cognitive states also vary. Furthermore, these differences coexist and influence each other; that is, differences in learning strategies may be due to differences in students or differences in subject matter. Summary of the Invention

[0004] The present invention aims to address the shortcomings of the existing technology by proposing an adaptive cognitive diagnosis method based on two-layer adversarial approaches. This method aims to utilize other scenarios, such as the problem-solving records of different students and different problems, to achieve adaptive cognitive diagnosis in cases where problem-solving records are missing, thereby enabling more accurate prediction of student performance.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The adaptive cognitive diagnosis method based on two-layer adversarial approach of the present invention is characterized by the following steps:

[0007] Step 1: Define the number of students in the target domain as N. t The number of exercises is M t The number of knowledge points is K t Let the j-th student, the n-th exercise, and the q-th knowledge point in the target domain be denoted as _j_, _n_, _q ... and 1≤j≤N t , 1≤n≤M t , 1≤q≤K t ;

[0008] Define the number of students in the source domain as N. s The number of exercises is M s The number of knowledge points is K s Let the i-th student, the m-th exercise, and the p-th knowledge point in the source domain be denoted as , ... and 1≤i≤N s , 1≤m≤M s 1≤p≤K s ;

[0009] Let the i-th student in the source domain For the m-th exercise The answer log is recorded as in, It is the i-th student in the source domain. For the m-th exercise The answer score, if This indicates that the answer is correct. This indicates an incorrect answer;

[0010] definition This represents the m-th exercise in the source domain. Includes the p-th knowledge point The Q matrix;

[0011] definition This represents the nth exercise in the target domain. Includes the q-th knowledge point The Q matrix;

[0012] Step 2: Construct a two-layer adversarial adaptive cognitive diagnostic model, including: source domain representation layer, source domain diagnostic layer, target domain representation layer, target domain diagnostic layer, two adaptive learning layers, and a prediction layer;

[0013] Step 2.1: The source domain representation layer obtains the i-th student using equations (1), (2), (3), and (4) respectively. Knowledge point proficiency vector The m-th exercise Difficulty vector The m-th exercise Discrimination vector The m-th exercise Includes the p-th knowledge point Q matrix vector

[0014]

[0015]

[0016]

[0017]

[0018] In equations (1)-(4), They represent the i-th student. and the m-th exercise One-hot encoded vector; These represent three matrices to be trained;

[0019] Step 2.2: The target domain representation layer uses equations (5), (6), (7), and (8) to obtain the j-th student respectively. Knowledge point proficiency vector The nth exercise Difficulty vector The nth exercise Discrimination vector The nth exercise Includes the q-th knowledge point Q matrix vector

[0020]

[0021]

[0022]

[0023]

[0024] In equations (5)-(8), Let each represent the j-th student. and the nth exercise One-hot encoded vector; These represent three matrices to be trained;

[0025] Step 2.3: The first adaptive learning layer includes a first generator and a first discriminator;

[0026] Step 2.3.1: The first generator uses equation (9) to generate the knowledge point proficiency vectors in the two domains. and Encode:

[0027]

[0028] In equation (9), and These are the feature representations of the knowledge proficiency vectors in the source and target domains, respectively, where u1, u2, d1, and d2 represent the training parameters of the hidden layer. These are the transposes of u1 and u2, respectively;

[0029] Step 2.3.2: The first generator uses equation (10) to calculate the difficulty vectors of the exercises in the two domains. and Encode:

[0030]

[0031] In equation (10), and , , and , respectively, represent the feature representations of the problem difficulty vectors in the source and target domains, and u3, u4, d3, and d4 represent the training parameters of the hidden layer. These are the transposes of u3 and u4, respectively;

[0032] Step 2.3.3: The first generator uses equation (11) to evaluate the discriminant vectors of exercises in the two domains. and Encode:

[0033]

[0034] In equation (11), and , which are the feature representations of the exercise discrimination vectors in the source and target domains, respectively, and u5, u6, d5, and d6 represent the training parameters of the hidden layer. These are the transposes of u5 and u6, respectively;

[0035] Step 2.3.4: Construct the objective function of the knowledge point proficiency of the first discriminator using equations (12)-(14). Objective function of problem difficulty Objective function of discrimination between exercises

[0036]

[0037]

[0038]

[0039] Step 2.4: The source domain diagnostic layer includes a source domain interaction layer and two source domain fully connected layers;

[0040] Step 2.4.1: The source domain interaction layer obtains the i-th student using equation (15). Do the m-th exercise that contains the p-th knowledge point. Interaction vectors

[0041]

[0042] Step 2.4.2: The first source domain fully connected layer and the second source domain fully connected layer use equation (16) to obtain the i-th student. Do the m-th exercise that contains the p-th knowledge point. Feature representation and dimensionality reduction feature representation

[0043]

[0044] In equation (16), φ is the activation function Sigmoid, W1 and W2 are the weight matrices of the source domain fully connected layer, and b1 and b2 are the bias vectors of the source domain fully connected layer.

[0045] Step 2.5: The target domain diagnostic layer includes a target domain interaction layer and two target domain fully connected layers;

[0046] Step 2.5.1: The target domain interaction layer uses equation (17) to establish the j-th student. Do the nth exercise that includes the qth knowledge point. Interaction vectors

[0047]

[0048] Step 2.5.2: The first target domain fully connected layer and the second target domain fully connected layer use equation (18) to obtain the j-th student. Do the nth exercise that includes the qth knowledge point. Feature vector representation of the degree of mastery of knowledge points Dimensionally reduced feature vector representation of the degree of mastery of knowledge points

[0049]

[0050] In equation (18), W3 and W4 are the weight matrices of the target domain fully connected layer, and b3 and b4 are the bias vectors of the target domain fully connected layer.

[0051] Step 2.6, the second adaptive learning layer includes: a second generator and a second discriminator:

[0052] Step 2.6.1: The second generator uses equation (19) to represent the dimensionality-reduced feature vector of the mastery level of knowledge points in the two domains. and Encode:

[0053]

[0054] In equation (19), and , , and d7 are the feature representations of the degree of knowledge mastery in the source and target domains, respectively, and u7, u8, d7, and d8 represent the training parameters of the hidden layer. These are the transposes of u7 and u8, respectively;

[0055] Step 2.6.2: Use equation (20) to establish the objective function of the second discriminator.

[0056]

[0057] Step 2.7: The prediction layer uses equation (21) to predict the i-th student. Do the m-th exercise that contains the p-th knowledge point. Score y imp :

[0058]

[0059] In equation (21), W5 is the weight matrix of the source domain fully connected layer, and b5 is the bias vector of the source domain fully connected layer.

[0060] Step 3: Training the two-layer adversarial adaptive cognitive diagnostic network:

[0061] Step 3.1: Initialize all network parameters using the Xavier initialization method;

[0062] Step 3.2: Use formula (22) to calculate the s of the i-th student. i s Total loss for answering all exercises i,j :

[0063]

[0064] In equation (22), λ1, λ2, λ3 and λ4 are four weight parameters;

[0065] Step 4.3: Train the two-layer adversarial adaptive cognitive diagnostic network using the Adam optimizer and minimize the loss. i,j Until Loss i,j The optimal cognitive diagnostic network is obtained by converging until the network reaches convergence. This network is used to diagnose students' mastery of knowledge points using a two-layer adversarial adaptive cognitive diagnostic network.

[0066] 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 two-layer adversarial adaptive cognitive diagnostic method, and the processor is configured to execute the program stored in the memory.

[0067] 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 two-layer adversarial adaptive cognitive diagnostic method.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] 1. This invention proposes an adaptive cognitive diagnosis method that can effectively transfer source domain information to the target domain, helping the target domain to perform cognitive diagnosis tasks. It overcomes the limitation of traditional cognitive diagnosis methods that are limited by the small number of target domain samples, and enables cognitive diagnosis of data without question-answering records in different scenarios.

[0070] 2. This invention designs a two-layer adversarial adaptive method. The first layer of adversarial interaction is in the representation layer, which is used to adaptively learn the representation of knowledge point proficiency, exercise difficulty, and exercise discrimination. The second layer of adversarial interaction is in the diagnostic layer, which is used for adaptive learning of cognitive state. The combination of the two layers of adversarial interaction can better obtain representation, thereby improving the accuracy of cognitive diagnosis.

[0071] 3. The two-layer adversarial adaptive cognitive diagnosis method proposed in this invention has good versatility and can be well integrated into the neural network model of cognitive diagnosis. This invention incorporates the two-layer adversarial adaptive method into three neural network models of cognitive diagnosis, which improves the accuracy of cognitive diagnosis. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the adaptive cognitive diagnosis of the present invention;

[0073] Figure 2 This is a model diagram of the adaptive cognitive diagnosis method based on two-layer adversarial approaches of the present invention;

[0074] Figure 3 This is a flowchart of the adaptive learning layer of the present invention. Detailed Implementation

[0075] In this embodiment, as Figure 1 As shown, in adaptive cognitive diagnosis, since the target domain lacks response records, the cognitive diagnosis model cannot provide training for students learning physics. However, response records from students learning mathematics in the source domain can be used to assist in training the cognitive diagnosis model for physics. Specifically, a two-layer adversarial adaptive cognitive diagnosis method proceeds as follows:

[0076] Step 1: Define the number of students in the target domain as N. t The number of exercises is M t The number of knowledge points is K t Let the j-th student, the n-th exercise, and the q-th knowledge point in the target domain be denoted as _j_, _n_, _q ... and 1≤j≤N t , 1≤n≤M t , 1≤q≤K t ;

[0077] Define the number of students in the source domain as N. s The number of exercises is M s The number of knowledge points is K s Let the i-th student, the m-th exercise, and the p-th knowledge point in the source domain be denoted as , ... and 1≤i≤N s , 1≤m≤M s 1≤p≤K s ;

[0078] Let the i-th student in the source domain For the m-th exercise The answer log is recorded as in, It is the i-th student in the source domain. For the m-th exercise The answer score, if This indicates that the answer is correct. This indicates an incorrect answer;

[0079] definition This represents the m-th exercise in the source domain. Includes the p-th knowledge point The Q matrix;

[0080] definition This represents the nth exercise in the target domain. Includes the q-th knowledge point The Q matrix;

[0081] In this embodiment, as Figure 2As shown, this method introduces domain adaptation technology into cognitive diagnosis tasks for the first time, proposing an adaptive cognitive diagnosis (ACD) learning problem. This involves using a cognitive diagnosis model based on existing problem-solving records to achieve cognitive diagnosis in different scenarios where the problem-solving records are unknown. Furthermore, addressing the problem of the coexistence and mutual influence of various distributional differences among students, problems, knowledge points, and cognitive states, a two-layer adversarial adaptive cognitive diagnosis method is proposed. First, at the representation layer, adaptive learning is performed on knowledge point proficiency, problem difficulty, and problem discrimination. Based on this, further adaptive learning is performed at the diagnosis layer. Specifically:

[0082] Step 2: Construct a two-layer adversarial adaptive cognitive diagnostic model, including: source domain representation layer, source domain diagnostic layer, target domain representation layer, target domain diagnostic layer, two adaptive learning layers, and a prediction layer;

[0083] Step 2.1: The source domain representation layer uses equations (1), (2), (3), and (4) to obtain the i-th student respectively. Knowledge point proficiency vector The m-th exercise Difficulty vector The m-th exercise Discrimination vector The m-th exercise Includes the p-th knowledge point Q matrix vector

[0084]

[0085]

[0086]

[0087]

[0088] In equations (1)-(4), They represent the i-th student. and the m-th exercise One-hot encoded vector; These represent three matrices to be trained;

[0089] Step 2.2: The target domain representation layer uses equations (5), (6), (7), and (8) to obtain the j-th student respectively. Knowledge point proficiency vector The nth exercise Difficulty vector The nth exercise Discrimination vector The nth exercise Includes the q-th knowledge point Q matrix vector

[0090]

[0091]

[0092]

[0093]

[0094] In equations (5)-(8), Let each represent the j-th student. and the nth exercise One-hot encoded vector; These represent three matrices to be trained;

[0095] In this embodiment, as Figure 3 As shown, the adaptive learning process proceeds as follows:

[0096] Step 2.3: The first adaptive learning layer includes a first generator and a first discriminator;

[0097] Step 2.3.1: The first generator uses equation (9) to generate the knowledge point proficiency vectors in the two domains. and Encode:

[0098]

[0099] In equation (9), and These are the feature representations of the knowledge proficiency vectors in the source and target domains, respectively, where u1, u2, d1, and d2 represent the training parameters of the hidden layer. These are the transposes of u1 and u2, respectively;

[0100] Step 2.3.2: The first generator uses equation (10) to calculate the difficulty vectors of the exercises in the two domains. and Encode:

[0101]

[0102] In equation (10), and , , and , respectively, represent the feature representations of the problem difficulty vectors in the source and target domains, and u3, u4, d3, and d4 represent the training parameters of the hidden layer. These are the transposes of u3 and u4, respectively;

[0103] Step 2.3.3: The first generator uses equation (11) to evaluate the discriminant vectors of exercises in the two domains. and Encode:

[0104]

[0105] In equation (11), and , which are the feature representations of the exercise discrimination vectors in the source and target domains, respectively, and u5, u6, d5, and d6 represent the training parameters of the hidden layer. These are the transposes of u5 and u6, respectively;

[0106] Step 2.3.4: Construct the objective function of the knowledge point proficiency of the first discriminator using equations (12)-(14). Objective function of problem difficulty Objective function of discrimination between exercises

[0107]

[0108]

[0109]

[0110] Step 2.4: The source domain diagnostic layer includes a source domain interaction layer and two source domain fully connected layers;

[0111] Step 2.4.1: The source domain interaction layer uses equation (15) to obtain the i-th student. Do the m-th exercise that contains the p-th knowledge point. Interaction vectors

[0112]

[0113] Step 2.4.2: The first source domain fully connected layer and the second source domain fully connected layer use equation (16) to obtain the i-th student. Do the m-th exercise that contains the p-th knowledge point. Feature representation and dimensionality reduction feature representation

[0114]

[0115] In equation (16), φ is the activation function Sigmoid, W1 and W2 are the weight matrices of the source domain fully connected layer, and b1 and b2 are the bias vectors of the source domain fully connected layer.

[0116] Step 2.5: The target domain diagnostic layer includes a target domain interaction layer and two fully connected target domain layers;

[0117] Step 2.5.1: The target domain interaction layer uses equation (17) to establish the j-th student. Do the nth exercise that includes the qth knowledge point. Interaction vectors

[0118]

[0119] Step 2.5.2: The first target domain fully connected layer and the second target domain fully connected layer use equation (18) to obtain the j-th student. Do the nth exercise that includes the qth knowledge point. Feature representation and dimensionality reduction feature representation

[0120]

[0121] In equation (18), W3 and W4 are the weight matrices of the target domain fully connected layer, and b3 and b4 are the bias vectors of the target domain fully connected layer.

[0122] Step 2.6, the second adaptive learning layer includes: a second generator and a second discriminator:

[0123] Step 2.6.1: The second generator uses equation (19) to represent the dimensionality reduction features in the two domains. and Encode:

[0124]

[0125] In equation (19), and , , and d7 represent the feature representations of the dimensionality-reduced feature representations in the source and target domains, respectively, and u7, u8, d7, and d8 represent the training parameters of the hidden layer. These are the transposes of u7 and u8, respectively;

[0126] Step 2.6.2: Use equation (20) to establish the objective function of the second discriminator.

[0127]

[0128] Step 2.7: The prediction layer uses equation (21) to predict the i-th student. Do the m-th exercise that contains the p-th knowledge point. Score y imp :

[0129]

[0130] In equation (21), W5 is the weight matrix of the source domain fully connected layer, and b5 is the bias vector of the source domain fully connected layer.

[0131] Step 3: Training the two-layer adversarial adaptive cognitive diagnostic network:

[0132] Step 3.1: Initialize all network parameters using the Xavier initialization method;

[0133] Step 3.2: Use formula (22) to calculate the i-th student. Total loss for answering all exercises i,j :

[0134]

[0135] In equation (22), λ1, λ2, λ3 and λ4 are four weight parameters;

[0136] Step 4.3: Train the two-layer adversarial adaptive cognitive diagnostic network using the Adam optimizer and minimize the loss. i,j Until Loss i,j The optimal cognitive diagnostic network is obtained by converging until the network reaches convergence. This network is used to diagnose students' mastery of knowledge points using a two-layer adversarial adaptive cognitive diagnostic network.

[0137] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the adaptive cognitive diagnostic method described above. The processor is configured to execute the program stored in the memory.

[0138] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the adaptive cognitive diagnosis method described above.

[0139] Example

[0140] To verify the effectiveness of the method of this invention, experiments were conducted on four datasets: Assistment09, Mooper, Math1, and NIPS2020. Assistment09 only provides student response logs and knowledge points, containing 4163 students, 17751 problems, and 123 knowledge points. Mooper is derived from the concept of Massive Open Online Practice, containing 26398 students, 7217 problems, and 3277 knowledge points. Math1 is data from a specific high school's final mathematics test, with objective and subjective problems, containing 4209 students, 20 problems, and 11 knowledge points. The NIPS2020 dataset comes from a diagnostic problem competition: the NeurIPS2020 Educational Challenge, which provides logs of students' responses to mathematics problems over two academic years (2018-2020), containing 6147 students, 948 problems, and 86 knowledge points.

[0141] 1. First, the dataset was partitioned. To construct different scenarios in the adaptive cognitive diagnosis task, different scenarios were designed for different students, different exercises, and different knowledge points. Different students mean that the students in the source scenario and the students in the target scenario are different. Obviously, when the knowledge points are different, the exercises will be different, while when the exercises are different, the knowledge points may be the same.

[0142] 2. Secondly, we compare the algorithms used in the selection of topics, including classic cognitive diagnostic models DINA and MIRT, as well as neural network models NCD, RCD, and KSCD.

[0143] 3. Evaluation metrics, such as the commonly used AUC metric.

[0144] Table 1 Experimental Results

[0145] (a)

[0146]

[0147] (b)

[0148]

[0149] (c)

[0150]

[0151] Table 1 shows the experimental results of this method. In Table 1, (a), (b), and (c) represent different students, exercises, and knowledge points in the source and target scenarios, respectively. NCD-A, RCD-A, and KSCD-A are the proposed methods added to NCD, RCD, and KSCD, respectively. The proposed adaptive diagnostic model (ACD) significantly outperforms the baseline across all datasets. The proposed two-layer adversarial adaptive cognitive diagnostic method can achieve adaptive cognitive diagnosis by adding adaptive modules to the NCD, RCD, and KSCD models, demonstrating the general applicability of this method.

Claims

1. An adaptive cognitive diagnosis method based on two-layer adversarial approaches, characterized in that, The procedure is as follows: Step 1: Define the number of students in the target domain as N. t The number of exercises is M t The number of knowledge points is K t Let the j-th student, the n-th exercise, and the q-th knowledge point in the target domain be denoted as _j_, _n_, _q ... and 1≤j≤N t , 1≤n≤M t , 1≤q≤K t ; Define the number of students in the source domain as N. s The number of exercises is M s The number of knowledge points is K s Let the i-th student, the m-th exercise, and the p-th knowledge point in the source domain be denoted as , ... and 1≤i≤N s , 1≤m≤M s 1≤p≤K s ; Let the i-th student in the source domain For the m-th exercise The answer log is recorded as in, It is the i-th student in the source domain. For the m-th exercise The answer score, if This indicates that the answer is correct. This indicates an incorrect answer; definition This represents the m-th exercise in the source domain. Includes the p-th knowledge point The Q matrix; definition This represents the nth exercise in the target domain. Includes the q-th knowledge point The Q matrix; Step 2: Construct a two-layer adversarial adaptive cognitive diagnostic model, including: source domain representation layer, source domain diagnostic layer, target domain representation layer, target domain diagnostic layer, two adaptive learning layers, and a prediction layer; Step 2.1: The source domain representation layer obtains the i-th student using equations (1), (2), (3), and (4) respectively. Knowledge point proficiency vector The m-th exercise Difficulty vector The m-th exercise Discrimination vector The m-th exercise Includes the p-th knowledge point Q matrix vector In equations (1)-(4), They represent the i-th student. and the m-th exercise One-hot encoded vector; These represent three matrices to be trained; Step 2.2: The target domain representation layer uses equations (5), (6), (7), and (8) to obtain the j-th student respectively. Knowledge point proficiency vector The nth exercise Difficulty vector The nth exercise Discrimination vector The nth exercise Includes the q-th knowledge point Q matrix vector In equations (5)-(8), Let each represent the j-th student. and the nth exercise One-hot encoded vector; These represent three matrices to be trained; Step 2.3: The first adaptive learning layer includes a first generator and a first discriminator; Step 2.3.1: The first generator uses equation (9) to generate the knowledge point proficiency vectors in the two domains. and Encode: In equation (9), and These are the feature representations of the knowledge proficiency vectors in the source and target domains, respectively, where u1, u2, d1, and d2 represent the training parameters of the hidden layer. These are the transposes of u1 and u2, respectively; Step 2.3.2: The first generator uses equation (10) to calculate the difficulty vectors of the exercises in the two domains. and Encode: In equation (10), and , , and , respectively, represent the feature representations of the problem difficulty vectors in the source and target domains, and u3, u4, d3, and d4 represent the training parameters of the hidden layer. These are the transposes of u3 and u4, respectively; Step 2.3.3: The first generator uses equation (11) to evaluate the discriminant vectors of exercises in the two domains. and Encode: In equation (11), and , which are the feature representations of the exercise discrimination vectors in the source and target domains, respectively, and u5, u6, d5, and d6 represent the training parameters of the hidden layer. These are the transposes of u5 and u6, respectively; Step 2.3.4: Construct the objective function of the knowledge point proficiency of the first discriminator using equations (12)-(14). Objective function of problem difficulty Objective function of discrimination between exercises Step 2.4: The source domain diagnostic layer includes a source domain interaction layer and two source domain fully connected layers; Step 2.4.1: The source domain interaction layer obtains the i-th student using equation (15). Do the m-th exercise that contains the p-th knowledge point. Interaction vectors Step 2.4.2: The first source domain fully connected layer and the second source domain fully connected layer use equation (16) to obtain the i-th student. Do the m-th exercise that contains the p-th knowledge point. Feature representation and dimensionality reduction feature representation In equation (16), φ is the activation function Sigmoid, W1 and W2 are the weight matrices of the source domain fully connected layer, and b1 and b2 are the bias vectors of the source domain fully connected layer. Step 2.5: The target domain diagnostic layer includes a target domain interaction layer and two target domain fully connected layers; Step 2.5.1: The target domain interaction layer uses equation (17) to establish the j-th student. Do the nth exercise that includes the qth knowledge point. Interaction vectors Step 2.5.2: The first target domain fully connected layer and the second target domain fully connected layer use equation (18) to obtain the j-th student. Do the nth exercise that includes the qth knowledge point. Feature vector representation of the degree of mastery of knowledge points Dimensionally reduced feature vector representation of the degree of mastery of knowledge points In equation (18), W3 and W4 are the weight matrices of the target domain fully connected layer, and b3 and b4 are the bias vectors of the target domain fully connected layer. Step 2.6, the second adaptive learning layer includes: a second generator and a second discriminator: Step 2.6.1: The second generator uses equation (19) to represent the dimensionality-reduced feature vector of the mastery level of knowledge points in the two domains. and Encode: In equation (19), and , , and d7 are the feature representations of the degree of knowledge mastery in the source and target domains, respectively, and u7, u8, d7, and d8 represent the training parameters of the hidden layer. These are the transposes of u7 and u8, respectively; Step 2.6.2: Use equation (20) to establish the objective function of the second discriminator. Step 2.7: The prediction layer uses equation (21) to predict the i-th student. Do the m-th exercise that contains the p-th knowledge point. Score y imp : In equation (21), W5 is the weight matrix of the source domain fully connected layer, and b5 is the bias vector of the source domain fully connected layer. Step 3: Training the two-layer adversarial adaptive cognitive diagnostic network: Step 3.1: Initialize all network parameters using the Xavier initialization method; Step 3.2: Use formula (22) to calculate the i-th student. Total loss for answering all exercises i,j : In equation (22), λ1, λ2, λ3 and λ4 are four weight parameters; Step 4.3: Train the two-layer adversarial adaptive cognitive diagnostic network using the Adam optimizer and minimize the loss. i,j Until Loss i,j The optimal cognitive diagnostic network is obtained by converging until the network reaches convergence. This network is used to diagnose students' mastery of knowledge points using a two-layer adversarial adaptive cognitive diagnostic network.

2. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the adaptive cognitive diagnostic method of the two-layer adversarial approach of claim 1, wherein the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the adaptive cognitive diagnostic method of claim 1, which is a two-layer adversarial approach.

Images