A method for learner cognitive diagnosis based on multi-layer attention mechanism

Abstract

A learner cognitive diagnosis method based on a multi-layer attention mechanism, which comprises the following steps: selecting and preprocessing a data set, dividing the data set, constructing a multi-layer attention neural network, constructing a question knowledge point matrix, constructing a learner cognitive level matrix, constructing a learner exercise score matrix, evaluating the final score of the learner, training the multi-layer attention neural network, testing the multi-layer attention neural network, evaluating the multi-layer attention neural network, and evaluating the cognitive level diagnosis result. The multi-layer attention neural network adopts a data enhancement module, which increases the amount of training data and reduces the risk of network overfitting. Compared with existing cognitive level diagnosis methods, the network construction cancels the generation of learner embedding vectors, improves the prediction ability of the network, solves the cold start problem of learners, and can be used for learner cognitive diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a method for diagnosing learner cognition. Background Technology

[0002] In recent years, with the increasing prevalence of big data technology in various fields, the development of intelligent education and online education has accelerated, and education systems have developed intelligent tutoring systems that are not limited by time or location. Cognitive diagnostics, as a core technology of personalized learning, aims to predict a student's mastery of relevant concepts based on their learning activity records over time. In a cognitive diagnostic system, learners practice, using their scores as input for cognitive diagnosis, and the system outputs the student's cognitive state. Based on these cognitive states, it can predict learners' test scores, recommend learning paths, or provide learning alerts to promote personalized and targeted learning. Therefore, an efficient cognitive diagnostic system is extremely important.

[0003] Early diagnostic methods primarily utilized probability statistics and educational psychology. However, the limitations of manually designed cognitive diagnostic functions led to technical issues, such as the difficulty of extracting information from the textual content of items. Later, some scholars proposed a neurocognitive diagnostic framework combining cognitive diagnosis with deep learning methods. This framework constructs neural networks to learn complex nonlinear item interactions, mapping learners and exercises to factor vectors and utilizing multi-layer neural modeling of the interactions between learners and items to enhance cognitive diagnostic results. However, when new learners join the intelligent tutor system, the trained cognitive diagnostic model fails to grasp the new learners' learning characteristics. Consequently, the prediction accuracy of existing cognitive diagnostic models decreases when predicting new learners' practice scores, resulting in the cold-start problem for new learners. Summary of the Invention

[0004] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art, namely the cold start problem of new learners, and to propose a learner cognitive diagnosis method based on a multi-layer attention mechanism that is interpretable, has good network performance, and has a wide range of applications.

[0005] The technical solution adopted to solve the above technical problems consists of the following steps:

[0006] (1) Selecting and preprocessing the dataset

[0007] Select the response records of learners with more than 15 responses from the MathEC dataset from Edi, including learners, exercises, and knowledge points. The learner set L is as follows:

[0008] L = {l1, l2, ..., l u ,...,l I}

[0009] Among them, l u Let u represent the u-th learner, 1≤u≤I, where I represents the total number of learners and is a finite positive integer.

[0010] The set of exercises E is as follows:

[0011] E = {e1, e2, ..., e m}

[0012] Among them, e m Let m represent the m-th exercise, 1 ≤ m ≤ M, where M represents the total number of exercises and is a finite positive integer.

[0013] The knowledge point set C is as follows:

[0014] C = {c1, c2, ..., c} k}

[0015] Among them, c k Let K represent the k-th knowledge point, where 1 ≤ k ≤ K, and K represents the total number of knowledge points, which is a finite positive integer.

[0016] learner l u Answer record set R u as follows:

[0017] R u ={r u1 ,r u2 ,...,r um}

[0018] Where, r um Represents the u-th learner l u The score on the m-th exercise.

[0019] (2) Split the dataset

[0020] The dataset was divided into training and test sets in an 8:2 ratio based on the learners.

[0021] (3) Constructing a multi-layer attention neural network

[0022] The multilayer attention neural network is composed of a data augmentation module, attention mechanism module 1, attention mechanism module 2, attention mechanism module 3, and a guessing error module connected in series.

[0023] The data enhancement module is composed of a data partitioning module and a data aggregation module connected in series.

[0024] The aforementioned guessing error module is composed of a guessing processing module and an error processing module connected in parallel and then connected in series with an information aggregation module.

[0025] (4) Construct a matrix of knowledge points for answering questions

[0026] 1) Constructing the learner's historical answer matrix Z

[0027] Construct the learner's historical answer matrix Z using the following formula:

[0028] Z = (z im ) (βI)×M

[0029]

[0030] Among them, X i This indicates that the learner's data has been processed by the data augmentation module. i A collection of historical answers to questions.

[0031] 2) Construct the learner prediction response matrix W

[0032] Construct the learner prediction matrix W using the following formula:

[0033] W = (w im ) (βI)×M

[0034]

[0035] Among them, Y i This indicates that after processing by the data augmentation module, in learner l i A collection of verified answer records.

[0036] 3) Construct a matrix of historical answer knowledge points D

[0037] Construct the learner's historical answer knowledge point matrix D using the following formula:

[0038] D=(d ik ) (βI)×K

[0039]

[0040] (5) Constructing a learner cognitive level matrix

[0041] 1) Construct a matrix Q representing the relationship between exercises and knowledge points.

[0042] Construct the relationship matrix Q between exercises and knowledge points using the following formula:

[0043] Q = (q mk ) M×K

[0044]

[0045] 2) Constructing the linear interaction matrix of the exercises

[0046] Construct the M-row, H-column linear interaction matrix P″ of the exercise according to the following formula:

[0047] P″=(p″ mh ) M×H

[0048] Where, p″ mh Let p' be the vector in the m-th row of the linear interaction matrix P″. m The h-th element in the array, where H is a hyperparameter with a value of 2. 5 ~2 8 .

[0049] 3) Construct a linear interaction matrix of knowledge points

[0050] Construct a linear interaction matrix T″ of knowledge points with K rows and H columns using the following formula:

[0051] T″=(t″ kh ) K×H

[0052] Where, t″ kh Let t be the vector in the k-th row of the linear interaction matrix T″. k The h-th element in.

[0053] 4) Construct a cognitive level matrix of learners for the knowledge points contained in the answered exercises.

[0054] Build learners using the following formula i The cognitive level matrix b of the knowledge points included in the answered exercises i :

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] Among them, b ik Represents the i-th learner l i In knowledge point c k The level of cognition above, 0 < b ik <1, Indicates learner l i In Exercise e m The weighting of questions in historical answer records, |C i | represents the i-th learner. i The set of answer records R iThe total number of knowledge points contained in the exercises is 0 < |C i |<K;p mk This represents knowledge point c. k For exercise e m Normalized weights, 0≤p mk ≤1, sim(p″ m ,t″ k ) represents the calculation exercise e m With knowledge point c k similarity, p″ m Let t' be the H-dimensional vector in the m-th row of matrix P''. k This represents the H-dimensional vector in the k-th row of matrix T″. It's a hyperparameter. The value ranges from 2 to 9.

[0061]

[0062] 5) Assess learners' cognitive levels a i

[0063] a i =(a ik ) 1×K

[0064]

[0065]

[0066]

[0067] Among them, a ik Indicates learner l i Regarding knowledge point c k cognitive level Indicates learner l i Regarding knowledge point c k In the already answered knowledge point set C i The normalization result on t vk This represents knowledge point c. v Regarding knowledge point c k Normalized weights, t uk This represents knowledge point c. u Regarding knowledge point c k Normalized weights.

[0068] (6) Construct a learner exercise score matrix

[0069] Construct learner exercise score matrix using the following formula

[0070]

[0071]

[0072]

[0073] in, Indicates learner l i In Exercise e m The upper score prediction, Indicates the meaning of exercise e m The normalized result of the included knowledge points

[0074]

[0075] (7) Assess learners’ final scores

[0076] Assess learners' final scores The method is as follows:

[0077] 1) Construct the probability matrix S for guessing:

[0078] S=(s1,s2,...,s m )

[0079] Among them, s m The weight value represents the probability that the learner guesses correctly on the m-th exercise, and is a trainable parameter with an initial value of [value missing].

[0080] 2) Construct the error probability matrix G:

[0081] G = (g1, g2, ..., g m )

[0082] Where g m The weights represent the probability of a learner making a mistake on the m-th question, and are trainable parameters with an initial value of [value missing].

[0083] 3) Assess learners' final scores

[0084]

[0085]

[0086] in, Indicates learner l i In Exercise e m Score predictions on the screen.

[0087] (8) Training a multi-layer attention neural network

[0088] 1) Constructing the loss function

[0089] The loss function Loss is constructed as follows:

[0090]

[0091] Among them, y im Represents the i-th learner l i In the m-th exercise e m The actual score on the screen, Represents the i-th learner l i In the m-th exercise e m The predicted score is given by βI, which represents the number of questions I answered by each learner increased by β times after data augmentation.

[0092] 2) Training a multi-layer attention neural network

[0093] Learner i The set of answer records R i The input data augmentation module enhances the data into a set R of β answer records. i .

[0094] The linear transformation matrix P″ of the exercises, the linear transformation matrix T″ of the knowledge points, the relationship matrix Q between the exercises and the knowledge points, and the learner l are constructed. i The set of answer records R i Input attention mechanism module 1, extract learner's cognitive level matrix b of the knowledge points contained in the answered questions. i :

[0095]

[0096] The linear transformation matrix T″ of the knowledge points and the learner's cognitive level matrix b of the knowledge points covered in the answered exercises are used to represent the knowledge points. i The input is fed into attention mechanism module 2 to extract the learner's cognitive level matrix a across all knowledge points. i :

[0097] a i =(a ik ) 1×K

[0098] The linear transformation matrix P″ of the exercises, the linear transformation matrix T″ of the knowledge points, and the learner's cognitive level matrix a across all knowledge points are used to... i The input is fed into attention mechanism module 3 to obtain a score matrix based on the learner's cognitive level.

[0099]

[0100] The guessing probability matrix S, the error probability matrix G, and the learner's score prediction based on cognitive level are combined. The module for inputting incorrect guesses assesses the learner's final score.

[0101]

[0102] in, This indicates that the guessing probability s was taken into account. m And the probability of error g m Afterwards, learners l i In the m-th exercise e m The final score prediction.

[0103] During training, batch sizes of 2 were used respectively. 3 ~2 8 Training was performed with a learning rate set to 0.003–0.03 and a hyperparameter H value of 2. 5 ~2 8 hyperparameters The value of is 2 to 9. During training, the loss function Loss is minimized to update the network parameters until the loss function Loss converges.

[0104] (9) Testing multi-layer attention neural networks

[0105] The test set is input into a multi-layer attention neural network for testing, and the results of the learner's cognitive level on the test set are output.

[0106] (10) Evaluate multi-layer attention neural networks

[0107] The performance of multilayer attention neural networks on the test set was evaluated using accuracy (ACC), root mean square error (RMSE), and area under the curve (AUC) metrics.

[0108] The accuracy ACC is determined by the following formula:

[0109]

[0110] in, Indicates learner l u As the total number of verification exercises, y um Indicates learner l u In Exercise e m The actual score on the screen, Indicates learner l u In Exercise e m Score prediction on x um Indicates whether the learner answered exercise e. m .

[0111] The root mean square error (RMSE) is determined by the following formula:

[0112]

[0113] Among them, the closer the accuracy (ACC) value is to 1 and the closer the root mean square error (RMSE) value is to 0, the better the prediction performance of the network.

[0114] (11) Assessing cognitive level and diagnostic results

[0115] The degree of consistency (DOA) is determined by the following formula:

[0116]

[0117]

[0118]

[0119] Among them, a uk Indicates learner l u In knowledge point c k The cognitive state above, a vk Indicates learner l v In knowledge point c k The cognitive state above.

[0120] A higher DOA value indicates that the network has better diagnostic performance in assessing learners' cognitive levels.

[0121] In step (2) of this invention, which involves constructing a multi-layer attention neural network, the data augmentation module is constructed as follows:

[0122] The data partitioning module ξ in the data augmentation module is constructed as follows:

[0123] ξ=KF(β,R u )

[0124] Wherein, KF represents the learner l u The set of answer records R u The data is divided into β-folds, where β takes the value of an integer from 2 to 6;

[0125] {R u1 ,R u2 ,...,R uβ} for R u The result of partitioning the data into β parts.

[0126] The data aggregation module ε in the data augmentation module is constructed as follows:

[0127] ε=CB({R u1 ,R u2 ,...,R uβ})

[0128] Where CB represents a data aggregation function, defined by CB({Ru1 ,R u2 ,...,R uβ}) to get {({R u1 ,R u2 ,...,R uβ}-{R u1},{R u1}),({R u1 ,R u2 ,...,R uβ}-{R u2},{R u2}),…,({R u1 ,R u2 ,...,R uβ}-{R uβ},{R u β})}, where β takes the value of an integer from 2 to 6, {R u1 ,R u2 ,...,R uβ}-{R uj} is recorded as The learner before data augmentation u The set of historical answer records obtained after processing by the data augmentation module, 1≤j≤β, {R} uj} is recorded as The learner before data augmentation u The j-th set of verification answer records obtained after data augmentation is obtained. The number of learners is increased from I to β.I, X after further processing by the data augmentation module. i Indicates the learner after data augmentation l i A set of historical answer records, 1≤i≤β.I,X i Corresponding to The relationship between i, u, and j can be expressed as i = (u-1) × β + j, Y i Indicates the learner after data augmentation l i The set of verification answer records, Y i Corresponding to

[0129] In step (2) of this invention, which involves constructing a multi-layer attention neural network, the data augmentation module is constructed as follows:

[0130] The data partitioning module ξ in the data augmentation module is constructed as follows:

[0131] ξ=KF(β,R u )

[0132] Wherein, KF represents the learner l u The set of answer records R uFor the β-fold data partitioning, the optimal value of β is 4; {R u1 ,R u2 ,...,R uβ} for R u The result of partitioning the data into β parts.

[0133] The data aggregation module ε in the data augmentation module is constructed as follows:

[0134] ε=CB({R u1 ,R u2 ,...,R uβ})

[0135] Where CB represents a data aggregation function, defined by CB({R u1 ,R u2 ,...,R uβ}) to get {({R u1 ,R u2 ,...,R uβ}-{R u1},{R u1}),({R u1 ,R u2 ,...,R uβ}-{R u2},{R u2}),…,({R u1 ,R u2 ,...,R uβ}-{R uβ},{R uβ})}, the optimal value of β is 4, {R u1 ,R u2 ,...,R uβ}-{R uj} is recorded as The learner before data augmentation u The set of historical answer records obtained after processing by the data augmentation module, 1≤j≤β, {R} uj} is recorded as The learner before data augmentation u The j-th set of verification answer records obtained after data augmentation is obtained. The number of learners is increased from I to β.I, X after further processing by the data augmentation module. i Indicates the learner after data augmentation l i A set of historical answer records, 1≤i≤β.I,X i Corresponding to The relationship between i, u, and j can be expressed as i = (u-1) × β + j, Y i Indicates the learner after data augmentation l i The set of verification answer records, Yi Corresponding to

[0136] In step (2) of this invention, which involves constructing a multi-layer attention neural network, the method for constructing the guessing error module is as follows:

[0137] The construction method of the guessing error handling module is as follows:

[0138]

[0139] in, This indicates that after considering the guess probability weights, the learner l i In Exercise e m Score predictions on the screen.

[0140] The method for constructing the error handling module of the guessing error module is as follows:

[0141]

[0142] in, This indicates that after considering the error probability weights, the learner l i In Exercise e m Score predictions on the screen.

[0143] The construction method of the information fusion module for the guessing error module is as follows:

[0144]

[0145] In step (8) of the present invention, training a multi-layer attention neural network, the optimal batch size used during training is 2. 6 During training, the optimal learning rate is 0.016, and the optimal value for the hyperparameter H is 2. 6 hyperparameters The optimal value is 6. During training, the loss function Loss is minimized to update the network parameters until the loss function Loss converges.

[0146] Compared with the prior art, the present invention has the following advantages:

[0147] Because the multi-layer attention neural network of this invention employs a data augmentation module construction method, it reduces the risk of network overfitting while increasing the amount of training data. Compared to existing cognitive level diagnostic methods, it eliminates the need to generate embedding vectors from learner IDs during network construction, improving the network's predictive ability and effectively solving the learner's cold start problem. In constructing the multi-layer attention neural network, it fully utilizes the interactions between knowledge points and the quantitative relationship between exercises and knowledge points, while considering both error probability and guessing probability, and the output of each module is between 0 and 1. This invention has advantages such as good network predictive performance, reduced risk of network overfitting, and solving the learner's cold start problem, and can be used for learner cognitive diagnosis. Attached Figure Description

[0148] Figure 1 This is a flowchart of Embodiment 1 of the present invention.

[0149] Figure 2 This is a schematic diagram of a multi-layer attention neural network.

[0150] Figure 3 yes Figure 2 A schematic diagram of the data enhancement module.

[0151] Figure 4 yes Figure 2 A schematic diagram of the error guessing module.

[0152] Figure 5 The figure shows the simulation results of the comparison between the method in Example 1 and the comparative experimental method. Detailed Implementation

[0153] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.

[0154] Example 1

[0155] The learner cognitive diagnosis method based on multi-layer attention mechanism in this embodiment consists of the following steps (see...). Figure 1 ):

[0156] (1) Selecting and preprocessing the dataset

[0157] Select the response records of learners with more than 15 responses from the MathEC dataset from Edi, including learners, exercises, and knowledge points. The learner set L is as follows:

[0158] L = {l1, l2, ..., l u ,...,l I}

[0159] Among them, lu Let u represent the u-th learner, 1≤u≤I, where I represents the total number of learners and is a finite positive integer.

[0160] The set of exercises E is as follows:

[0161] E = {e1, e2, ..., e m}

[0162] Among them, e m Let m represent the m-th exercise, 1 ≤ m ≤ M, where M represents the total number of exercises and is a finite positive integer.

[0163] The knowledge point set C is as follows:

[0164] C = {c1, c2, ..., c} k}

[0165] Among them, c k Let K represent the k-th knowledge point, where 1 ≤ k ≤ K, and K represents the total number of knowledge points, which is a finite positive integer.

[0166] learner l u Answer record set R u as follows:

[0167] R u ={r u1 ,r u2 ,...,r um}

[0168] Where, r um Represents the u-th learner l u The score on the m-th exercise.

[0169] (2) Split the dataset

[0170] The dataset was divided into training and test sets in an 8:2 ratio based on the learners.

[0171] (3) Constructing a multi-layer attention neural network

[0172] Figure 2 A schematic diagram of the multilayer attention neural network structure in this embodiment is provided. Figure 2 In this embodiment, the multilayer attention neural network is composed of a data augmentation module, attention mechanism module 1, attention mechanism module 2, attention mechanism module 3, and a guessing error module connected in series.

[0173] Figure 3 Given Figure 2 A schematic diagram of the data augmentation module. Figure 3 In this embodiment, the data enhancement module is composed of a data partitioning module and a data aggregation module connected in series.

[0174] Figure 4 Given Figure 2 A schematic diagram of the error-guessing module. Figure 4 In this embodiment, the guessing error module is composed of a guessing processing module and an error processing module connected in parallel and then connected in series with the information aggregation module.

[0175] The data augmentation module is constructed as follows:

[0176] The data partitioning module ξ in the data augmentation module is constructed as follows:

[0177] ξ=KF(β,R u )

[0178] Wherein, KF represents the learner l u The set of answer records R u The data is divided into β-folds, with β taking the value of an integer from 2 to 6; in this embodiment, β is 4. {R u1 ,R u2 ,...,R uβ} for R u The result of partitioning the data into β parts.

[0179] The data aggregation module ε in the data augmentation module is constructed as follows:

[0180] ε=CB({R u1 ,R u2 ,...,R uβ})

[0181] Where CB represents a data aggregation function, defined by CB({R u1 ,R u2 ,...,R uβ}) to get {({R u1 ,R u2 ,...,R uβ}-{R u1},{R u1}),({R u1 ,R u2 ,...,R uβ}-{R u2},{R u2}),…,({R u1 ,R u2 ,...,R uβ}-{R uβ},{R uβ})}, β takes the value of an integer from 2 to 6, and in this embodiment, β is 4. {R u1 ,R u2 ,...,R uβ}-{Ruj} is recorded as The learner before data augmentation u The set of historical answer records obtained after processing by the data augmentation module, 1≤j≤β, {R} uj} is recorded as The learner before data augmentation u The j-th set of verification answer records obtained after data augmentation is obtained. The number of learners is increased from I to β.I, X after further processing by the data augmentation module. i Indicates the learner after data augmentation l i A set of historical answer records, 1≤i≤β.I,X i Corresponding to The relationship between i, u, and j can be expressed as i = (u-1) × β + j, Y i Indicates the learner after data augmentation l i The set of verification answer records, Y i Corresponding to

[0182] The method for constructing the error guessing module is as follows:

[0183] The construction method of the guessing error handling module is as follows:

[0184]

[0185] in, This indicates that after considering the guess probability weights, the learner l i In Exercise e m Score predictions on the screen.

[0186] The method for constructing the error handling module of the guessing error module is as follows:

[0187]

[0188] in, This indicates that after considering the error probability weights, the learner l i In Exercise e m Score predictions on the screen.

[0189] The construction method of the information fusion module for the guessing error module is as follows:

[0190]

[0191] This embodiment employs a data augmentation module construction method to construct a data augmentation module, which reduces the risk of network overfitting while increasing the amount of training data. Compared with existing cognitive level diagnostic methods, the generation of embedding vectors from learner IDs is eliminated during network construction, improving the network's predictive ability and effectively solving the learner's cold start problem.

[0192] When constructing a multi-layer attention neural network, it can make full use of the interaction between knowledge points and the quantitative relationship between exercises and knowledge points, while taking into account the probability of error and the probability of guessing, and the output of each module is between 0 and 1.

[0193] (4) Construct a matrix of knowledge points for answering questions

[0194] 1) Constructing the learner's historical answer matrix Z

[0195] Construct the learner's historical answer matrix Z using the following formula:

[0196] Z = (z im ) (βI)×M

[0197]

[0198] Among them, X i This indicates that the learner's data has been processed by the data augmentation module. i A collection of historical answers to questions.

[0199] 2) Construct the learner prediction response matrix W

[0200] Construct the learner prediction matrix W using the following formula:

[0201] W = (w im ) (βI)×M

[0202]

[0203] Among them, Y i This indicates that after processing by the data augmentation module, in learner l i A collection of verified answer records.

[0204] 3) Construct a matrix of historical answer knowledge points D

[0205] Construct the learner's historical answer knowledge point matrix D using the following formula:

[0206] D=(d ik ) (βI)×K

[0207]

[0208] (5) Constructing a learner cognitive level matrix

[0209] 1) Construct a matrix Q representing the relationship between exercises and knowledge points.

[0210] Construct the relationship matrix Q between exercises and knowledge points using the following formula:

[0211] Q = (q mk ) M×K

[0212]

[0213] 2) Constructing the linear interaction matrix of the exercises

[0214] Construct the M-row, H-column linear interaction matrix P″ of the exercise according to the following formula:

[0215] P″=(p″ mh ) M×H

[0216] Where, p″ mh Let p' be the vector in the m-th row of the linear interaction matrix P″. m The h-th element in the array, where H is a hyperparameter with a value of 2. 5 ~2 8 In this embodiment, the value of H is 2. 6 .

[0217] 3) Construct a linear interaction matrix of knowledge points

[0218] Construct a linear interaction matrix T″ of knowledge points with K rows and H columns using the following formula:

[0219] T″=(t″ kh ) K×H

[0220] Among them, t k " h Let t represent the k-th row vector of the linear interaction matrix T″. k The h-th element in ″.

[0221] 4) Construct a cognitive level matrix of learners for the knowledge points contained in the answered exercises.

[0222] Build learners using the following formula i The cognitive level matrix b of the knowledge points included in the answered exercises i :

[0223]

[0224]

[0225]

[0226]

[0227]

[0228] Among them, b ik Represents the i-th learner l i In knowledge point c k The level of cognition above, 0 < b ik <1, Indicates learner l i In Exercise e m The weighting of questions in historical answer records, |C i | represents the i-th learner. i The set of answer records R i The total number of knowledge points contained in the exercises is 0 < |C i |<K;p mk This represents knowledge point c. k For exercise e m Normalized weights, 0≤p mk ≤1, sim(p″ m ,t″ k ) represents the calculation exercise e m With knowledge point c k similarity, p″ m Let t' be the H-dimensional vector in the m-th row of matrix P''. k This represents the H-dimensional vector in the k-th row of matrix T″. It's a hyperparameter. The value of is 2 to 9. This is the hyperparameter in this embodiment. The value is 6.

[0229]

[0230] 5) Assess learners' cognitive levels a i

[0231] a i =(a ik ) 1×K

[0232]

[0233]

[0234]

[0235] Among them, a ik Indicates learner l i Regarding knowledge point c k cognitive level Indicates learner l i Regarding knowledge point c kIn the already answered knowledge point set C i The normalization result on t vk This represents knowledge point c. v Regarding knowledge point c k Normalized weights, t uk This represents knowledge point c. u Regarding knowledge point c k Normalized weights.

[0236] (6) Construct a learner exercise score matrix

[0237] Construct learner exercise score matrix using the following formula

[0238]

[0239]

[0240]

[0241] in, Indicates learner l i In Exercise e m The upper score prediction, Indicates the meaning of exercise e m The normalized result of the included knowledge points

[0242]

[0243] (7) Assess learners’ final scores

[0244] Assess learners' final scores The method is as follows:

[0245] 1) Construct the probability matrix S for guessing:

[0246] S=(s1,s2,...,s m )

[0247] Among them, s m The weight value represents the probability that the learner guesses correctly on the m-th exercise, and is a trainable parameter with an initial value of [value missing].

[0248] 2) Construct the error probability matrix G:

[0249] G = (g1, g2, ..., g m )

[0250] Where g m The weights represent the probability of a learner making a mistake on the m-th question, and are trainable parameters with an initial value of [value missing].

[0251] 3) Assess learners' final scores

[0252]

[0253]

[0254] in, Indicates learner l i In Exercise e m Score predictions on the screen.

[0255] (8) Training a multi-layer attention neural network

[0256] 1) Constructing the loss function

[0257] The loss function Loss is constructed as follows:

[0258]

[0259] Among them, y im Represents the i-th learner l i In the m-th exercise e m The actual score on the screen, Represents the i-th learner l i In the m-th exercise e m The predicted score is given by βI, which represents the number of questions I answered by each learner increased by β times after data augmentation.

[0260] 2) Training a multi-layer attention neural network

[0261] Learner i The set of answer records R i The input data augmentation module enhances the data into a set R of β answer records. i .

[0262] The linear transformation matrix P″ of the exercises, the linear transformation matrix T″ of the knowledge points, the relationship matrix Q between the exercises and the knowledge points, and the learner l are constructed. i The set of answer records R i Input attention mechanism module 1, extract learner's cognitive level matrix b of the knowledge points contained in the answered questions. i :

[0263]

[0264] The linear transformation matrix T″ of the knowledge points and the learner's cognitive level matrix b of the knowledge points covered in the answered exercises are used to represent the knowledge points. i The input is fed into attention mechanism module 2 to extract the learner's cognitive level matrix a across all knowledge points. i .

[0265] a i =(a ik ) 1×K

[0266] The linear transformation matrix P″ of the exercises, the linear transformation matrix T″ of the knowledge points, and the learner's cognitive level matrix a across all knowledge points are used to... i The input is fed into attention mechanism module 3 to obtain a score matrix based on the learner's cognitive level.

[0267]

[0268] The guessing probability matrix S, the error probability matrix G, and the learner's score prediction based on cognitive level are combined. The module for inputting incorrect guesses assesses the learner's final score.

[0269]

[0270] in, This indicates that the guessing probability s was taken into account. m And the probability of error g m Afterwards, learners l i In the m-th exercise e m The final score prediction.

[0271] During training, batch sizes of 2 were used respectively. 3 ~2 8 During training, in this embodiment, a batch size of 2 is used respectively. 6 The training learning rate ranges from 0.003 to 0.03; in this embodiment, the learning rate is 0.016, and the hyperparameter H is 2. 5 ~2 8 In this embodiment, the hyperparameter H is set to 2. 6 ', hyperparameters The value of is 2 to 9. This is the hyperparameter in this embodiment. The value of is 6. During training, the loss function Loss is minimized to update the network parameters until the loss function Loss converges.

[0272] (9) Testing multi-layer attention neural networks

[0273] The test set is input into a multi-layer attention neural network for testing, and the results of the learner's cognitive level on the test set are output.

[0274] (10) Evaluate multi-layer attention neural networks

[0275] The performance of multilayer attention neural networks on the test set was evaluated using accuracy (ACC), root mean square error (RMSE), and area under the curve (AUC) metrics.

[0276] The accuracy ACC is determined by the following formula:

[0277]

[0278] in, Indicates learner l u As the total number of verification exercises, y um Indicates learner l u In Exercise e m The actual score on the screen, Indicates learner l u In Exercise e m Score prediction on x um Indicates whether the learner answered exercise e. m ;

[0279] The root mean square error (RMSE) is determined by the following formula:

[0280]

[0281] Among them, the closer the accuracy (ACC) value is to 1 and the closer the root mean square error (RMSE) value is to 0, the better the prediction performance of the network.

[0282] (11) Assessing cognitive level and diagnostic results

[0283] The degree of consistency (DOA) is determined by the following formula:

[0284]

[0285]

[0286]

[0287] Among them, a uk Indicates learner l u In knowledge point c k The cognitive state above, a vk Indicates learner l v In knowledge point c k The cognitive state above;

[0288] A higher DOA value indicates that the network has better diagnostic performance in assessing learners' cognitive levels.

[0289] Develop a cognitive diagnostic method for learners based on a multi-layered attention mechanism.

[0290] Example 2

[0291] The learner cognitive diagnosis method based on a multi-layer attention mechanism in this embodiment consists of the following steps:

[0292] (1) Selecting and preprocessing the dataset

[0293] The steps are the same as in Example 1.

[0294] (2) Split the dataset

[0295] The steps are the same as in Example 1.

[0296] (3) Constructing a multi-layer attention neural network

[0297] The structure of the multi-layer attention neural network is the same as that in Example 1.

[0298] The data partitioning module ξ in the data augmentation module of this embodiment is constructed as follows:

[0299] ξ=KF(β,R u )

[0300] Wherein, KF represents the learner l u The set of answer records R u The data is divided into β-folds, with β taking the value of an integer from 2 to 6. In this embodiment, β is 2. {R u1 ,R u2 ,...,R uβ} for R u The result of partitioning the data into β parts.

[0301] The data aggregation module ε in the data augmentation module is constructed as follows:

[0302] ε=CB({R u1 ,R u2 ,...,R uβ})

[0303] Where CB represents a data aggregation function, defined by CB({R u1 ,R u2 ,...,R uβ}) to get {({R u1 ,R u2 ,...,R uβ}-{R u1},{R u1}),({R u1 ,R u2 ,...,R uβ}-{R u2},{R u2}),…,({R u1 ,R u2 ,...,Ruβ}-{R uβ},{R uβ})}, where β takes the value of an integer from 2 to 6, and in this embodiment, β takes the value of 2. The meanings and value ranges of the other parameters in this step are the same as in Embodiment 1.

[0304] The method for constructing the guessing error module is the same as in Example 1.

[0305] (4) Construct a matrix of knowledge points for answering questions

[0306] The steps are the same as in Example 1.

[0307] (5) Constructing a learner cognitive level matrix

[0308] 1) Construct a matrix Q representing the relationship between exercises and knowledge points.

[0309] The steps are the same as in Example 1.

[0310] 2) Constructing the linear interaction matrix of the exercises

[0311] Construct the M-row, H-column linear interaction matrix P″ of the exercise according to the following formula:

[0312] P″=(p″ mh ) M×H

[0313] Where, p″ mh Let p' be the vector in the m-th row of the linear interaction matrix P″. m The h-th element in the array, where H is a hyperparameter with a value of 2. 5 ~2 8 In this embodiment, the value of H is 2. 5 .

[0314] This step is the same as in Example 1.

[0315] (6) Construct a learner exercise score matrix

[0316] The steps are the same as in Example 1.

[0317] (7) Assess learners’ final scores

[0318] The steps are the same as in Example 1.

[0319] (8) Training a multi-layer attention neural network

[0320] 1) Constructing the loss function

[0321] The steps are the same as in Example 1.

[0322] 2) Training a multi-layer attention neural network

[0323] Learneri The set of answer records R i The input data augmentation module enhances the data into a set R of α answer records. i .

[0324] The linear transformation matrix P″ of the exercises, the linear transformation matrix T″ of the knowledge points, the relationship matrix Q between the exercises and the knowledge points, and the learner l are constructed. i The set of answer records R i Input attention mechanism module 1, extract learner's cognitive level matrix b of the knowledge points contained in the answered questions. i :

[0325]

[0326] The linear transformation matrix T″ of the knowledge points and the learner's cognitive level matrix b of the knowledge points covered in the answered exercises are used to represent the knowledge points. i The input is fed into attention mechanism module 2 to extract the learner's cognitive level matrix a across all knowledge points. i .

[0327] a i =(a ik ) 1×K

[0328] The linear transformation matrix P″ of the exercises, the linear transformation matrix T″ of the knowledge points, and the learner's cognitive level matrix a across all knowledge points are used to... i The input is fed into attention mechanism module 3 to obtain a score matrix based on the learner's cognitive level.

[0329]

[0330] The guessing probability matrix S, the error probability matrix G, and the learner's score prediction based on cognitive level are combined. The module for inputting incorrect guesses assesses the learner's final score.

[0331]

[0332] in, This indicates that the guessing probability s was taken into account. m And the probability of error g m Afterwards, learners l i In the m-th exercise e m The final score prediction.

[0333] During training, batch sizes of 2 were used respectively. 3 ~2 8 During training, in this embodiment, a batch size of 2 is used respectively. 3The training learning rate ranges from 0.003 to 0.03; in this embodiment, the learning rate is 0.003, and the hyperparameter H is 2. 5 ~2 8 In this embodiment, the hyperparameter H is set to 2. 5 hyperparameters The value of is 2 to 9. This is the hyperparameter in this embodiment. The value of is 2. During training, the loss function Loss is minimized to update the network parameters until the loss function Loss converges.

[0334] The other steps are the same as in Example 1.

[0335] Develop a cognitive diagnostic method for learners based on a multi-layered attention mechanism.

[0336] Example 3

[0337] The learner cognitive diagnosis method based on a multi-layer attention mechanism in this embodiment consists of the following steps:

[0338] (1) Selecting and preprocessing the dataset

[0339] The steps are the same as in Example 1.

[0340] (2) Split the dataset

[0341] The steps are the same as in Example 1.

[0342] (3) Constructing a multi-layer attention neural network

[0343] The structure of the multi-layer attention neural network is the same as that in Example 1.

[0344] The data partitioning module ξ in the data augmentation module of this embodiment is constructed as follows:

[0345] ξ=KF(β,R u )

[0346] Wherein, KF represents the learner l u The set of answer records R u The data is divided into β-folds, with β taking the value of an integer from 2 to 6. In this embodiment, β is 6. {R u1 ,R u2 ,...,R uβ} for R u The result of partitioning the data into β parts.

[0347] The data aggregation module ε in the data augmentation module is constructed as follows:

[0348] ε=CB({R u1 ,R u2 ,...,Ruβ})

[0349] Where CB represents a data aggregation function, defined by CB({R u1 ,R u2 ,...,R uβ}) to get {({R u1 ,R u2 ,...,R uβ}-{R u1},{R u1}),({R u1 ,R u2 ,...,R uβ}-{R u2},{R u2}),…,({R u1 ,R u2 ,...,R uβ}-{R uβ},{R uβ})}, where β takes the value of an integer from 2 to 6, and in this embodiment, β takes the value of 6. The meanings and value ranges of the other parameters in this step are the same as in Embodiment 1.

[0350] The method for constructing the guessing error module is the same as in Example 1.

[0351] (4) Construct a matrix of knowledge points for answering questions

[0352] The steps are the same as in Example 1.

[0353] (5) Constructing a learner cognitive level matrix

[0354] 1) Construct a matrix Q representing the relationship between exercises and knowledge points.

[0355] The steps are the same as in Example 1.

[0356] 2) Constructing the linear interaction matrix of the exercises

[0357] Construct the M-row, H-column linear interaction matrix P″ of the exercise according to the following formula:

[0358] P″=(p″ mh ) M×H

[0359] Where, p″ mh Let p' be the vector in the m-th row of the linear interaction matrix P″. m The h-th element in the array, where H is a hyperparameter with a value of 2. 5 ~2 8 In this embodiment, the value of H is 2. 8 .

[0360] This step is the same as in Example 1.

[0361] (6) Construct a learner exercise score matrix

[0362] The steps are the same as in Example 1.

[0363] (7) Assess learners’ final scores

[0364] The steps are the same as in Example 1.

[0365] (8) Training a multi-layer attention neural network

[0366] 1) Constructing the loss function

[0367] The steps are the same as in Example 1.

[0368] 2) Training a multi-layer attention neural network

[0369] Learner i The set of answer records R i The input data augmentation module enhances the data into a set R of β answer records. i .

[0370] The linear transformation matrix P″ of the exercises, the linear transformation matrix T″ of the knowledge points, the relationship matrix Q between the exercises and the knowledge points, and the learner l are constructed. i The set of answer records R i Input attention mechanism module 1, extract learner's cognitive level matrix b of the knowledge points contained in the answered questions. i :

[0371]

[0372] The linear transformation matrix T″ of the knowledge points and the learner's cognitive level matrix b of the knowledge points covered in the answered exercises are used to represent the knowledge points. i The input is fed into attention mechanism module 2 to extract the learner's cognitive level matrix a across all knowledge points. i .

[0373] a i =(a ik ) 1×K

[0374] The linear transformation matrix P″ of the exercises, the linear transformation matrix T″ of the knowledge points, and the learner's cognitive level matrix a across all knowledge points are used to... i The input is fed into attention mechanism module 3 to obtain a score matrix based on the learner's cognitive level.

[0375]

[0376] The guessing probability matrix S, the error probability matrix G, and the learner's score prediction based on cognitive level are combined. The module for inputting incorrect guesses assesses the learner's final score.

[0377]

[0378] in, This indicates that the guessing probability s was taken into account. m And the probability of error g m Afterwards, learners l i In the m-th exercise e m The final score prediction.

[0379] During training, batch sizes of 2 were used respectively. 3 ~2 8 During training, in this embodiment, a batch size of 2 is used respectively. 8 The training learning rate ranges from 0.003 to 0.03; in this embodiment, the learning rate is 0.03, and the hyperparameter H is 2. 5 ~2 8 In this embodiment, the hyperparameter H is set to 2. 8 hyperparameters The value of is 2 to 9. This is the hyperparameter in this embodiment. The value of is 9. During training, the loss function Loss is minimized to update the network parameters until the loss function Loss converges.

[0380] The other steps are the same as in Example 1.

[0381] Develop a cognitive diagnostic method for learners based on a multi-layered attention mechanism.

[0382] To verify the beneficial effects of this method, learners with at least 15 answer records were selected from the ASSIST0910, ASSIST2017, FrcSub, JunYi, and MathEC datasets. A comparative simulation experiment was conducted using the method of Embodiment 1 of this invention (hereinafter referred to as SCD) against Deterministic Inputs (hereinafter referred to as DINA), Noisy “and” Gate (hereinafter referred to as NCDM), Neural Cognitive Diagnosis Model, and Using Knowledge Concept Aggregation towards Accurate Cognitive Diagnosis (hereinafter referred to as CDGK). The performance of the attention neural network on the test set was evaluated using accuracy (ACC), area under the curve (AUC), and root mean square error (RMSE). The experimental results are shown in Table 1.

[0383] Table 1. Comparison of simulation experiment results between the method of the present invention and the comparative method.

[0384]

[0385] As shown in Table 1, the ACC, AUC, and RMSE indices of the method of this invention are superior to those of the comparative methods.

[0386] The method of this invention was used to conduct a cognitive diagnostic experiment on new learners. The experimental results are shown in [the table below]. Figure 5 . Figure 5 The experimental results shown on the above dataset indicate that a higher DOA (Different Opinion on Analytical Ability) value indicates better diagnostic performance of the network in assessing learners' cognitive levels. Figure 5 In the graph, the horizontal axis represents the five datasets, and the vertical axis represents the degree of consistency (DOA) value. Figure 5 It is evident that the DOA value of the method in Embodiment 1 of this invention is significantly better than that of the comparative method.

Claims

1.A learner cognitive diagnosis method based on a multi-layer attention mechanism, characterized in that It consists of the following steps: (1) Selecting and preprocessing the dataset Select the response records of learners with more than 15 responses from the MathEC dataset from Edi, including learners, exercises, and knowledge points. The learner set L is as follows: wherein, denotes the u-th learner, I denotes the total number of learners, is a finite positive integer; The set of exercises E is as follows: wherein, denotes the m-th exercise, M denotes the total number of exercises, a finite positive integer; The knowledge point set C is as follows: wherein, represents the kth knowledge point, K represents the total number of knowledge points, which is a finite positive integer; learners answer record set as follows: wherein, denotes the u-th learner score on the m-th exercise; (2) Split the dataset The dataset was divided into training and test sets in an 8:2 ratio based on the learners. (3) Construct a multi-layer attention neural network The multilayer attention neural network is composed of a data augmentation module, attention mechanism module 1, attention mechanism module 2, attention mechanism module 3, and a guessing error module connected in series. The data enhancement module is composed of a data partitioning module and a data aggregation module connected in series; The aforementioned guessing error module is composed of a guessing processing module and an error processing module connected in parallel, and then connected in series with an information aggregation module; (4) Construct a matrix of knowledge points for answering questions 1) Constructing the learner's historical answer matrix Z Construct the learner's historical answer matrix Z using the following formula: ， in, This indicates the learner's performance after processing by the data augmentation module. A collection of historical answer records; 2) Construct the learner prediction response matrix W Construct the learner prediction matrix W using the following formula: ， wherein, represents that after data augmentation, the number of answers of each learner I is expanded times, represents the validation answer record set of the learner after processing by the data augmentation module; 3) Construct a historical answer knowledge point matrix D Construct the learner's historical answer knowledge point matrix D using the following formula: ； (5) Constructing a learner cognitive level matrix 1) Construct a matrix Q representing the relationship between exercises and knowledge points. Construct the relationship matrix Q between exercises and knowledge points using the following formula: ； 2) Constructing the linear interaction matrix of the exercises The M by H problem linear interaction matrix is constructed as follows : wherein, represents a linear interaction matrix the mth row vector the hth element in the vector, H is a hyperparameter, H takes a value of 2 5 ~ 2 8 ; 3) Construct a linear interaction matrix of knowledge points The knowledge point linear interaction matrix of K rows and H columns is constructed as follows : wherein, represents a linear interaction matrix the kth row vector the hth element in 4) Construct a cognitive level matrix of learners for the knowledge points contained in the answered exercises. The learner is constructed as follows The cognitive level matrix of the knowledge points contained in the answered questions : in, Represents the i-th learner In the knowledge points cognitive level , Indicate learner In the exercises The weighting of historical answer records in practice questions. Represents the i-th learner Collection of answer records The total number of knowledge points covered in the exercises. ; Representing knowledge points For exercises Normalized weights, , Representing calculation exercises With knowledge points similarity, Representation matrix The H-dimensional vector in the m-th row, Representation matrix The k-th row of the H-dimensional vector, It's a hyperparameter. The value ranges from 2 to 9; ； 5) Assessing the cognitive level of a learner in, Indicate learner Knowledge points cognitive level Indicate learner Knowledge points In the answerable knowledge point set The normalization result on, Representing knowledge points Knowledge points Normalized weights, Representing knowledge points Knowledge points Normalized weights; (6) Construct a learner exercise score matrix The learner exercise score matrix is constructed as follows : wherein, represents the learner in the exercise score prediction, represents the number of exercises I each learner answers after data augmentation is performed times; represents the normalized result of the knowledge point contained in the exercise ​ (7) Assess learners’ final scores Assessing learner final score The method is as follows: 1) Construct the probability matrix S for guessing: wherein, represents the weight value of the probability of the learner guessing correctly on the mth exercise, is a trainable parameter, and the initial value is ; 2) Construct the error probability matrix G: wherein, represents the weight value of the probability of the learner failing on the mth exercise, is a trainable parameter, and the initial value is ; 3) Assessing the final score of the learner : wherein, representing learners score prediction on exercises on exercises. (8) Training a multi-layer attention neural network 1) Constructing the loss function The loss function Loss is constructed as follows: in, Represents the i-th learner In the m-th exercise The actual score on the screen, Represents the i-th learner In the m-th exercise The predicted score on the screen This indicates that after data augmentation, the number of questions I answered by each learner was increased. times; 2) Training a multi-layer attention neural network a learner a set of answer records an input data augmentation module, to augment a set of answer records ; The linear transformation matrix of the constructed exercises Linear transformation matrix of knowledge points Matrix of the relationship between exercises and knowledge points and learners Collection of answer records Input attention mechanism module 1 extracts the learner's cognitive level matrix of the knowledge points contained in the answered questions. : a linear transformation matrix of knowledge points and a cognitive level matrix of the learner on the knowledge points contained in the answered exercise input to the attention mechanism module 2, to extract the cognitive level matrix of the learner on all knowledge points : in, Indicate learner Knowledge points The cognitive level will determine the linear transformation matrix of the exercises. Linear transformation matrix of knowledge points and learners' cognitive level matrix across all knowledge points The input is fed into attention mechanism module 3 to obtain a score matrix based on the learner's cognitive level. : The guessing probability matrix S, the error probability matrix G and the score prediction of the learner based on the cognitive level are combined An input guessing error module assesses the final score of the learner : wherein, represents the probability of guessing and the probability of error after the learner scores the final prediction on the mth problem . The batch size is 2 3 ~ 2 8 The learning rate of the training is set to 0.003-0.03, the value of the hyperparameter is 2 5 ~ 2 8 , the value of the hyperparameter is 2-9 The loss function Loss is minimized during training to update the network parameters until the loss function Loss converges. (9) Testing multi-layer attention neural networks The test set is input into a multi-layer attention neural network for testing, and the learner's cognitive level results on the test set are output. (10) Evaluate multi-layer attention neural networks Use accuracy Root mean square error And area under curve AUC evaluation index to evaluate the effect of multi-layer attention neural network on the test set; The accuracy is determined by the following formula : wherein, represents a learner as a total number of validation exercises, represents a learner a true score on an exercise , represents a learner a score prediction on an exercise , represents whether a learner answered an exercise ; The root mean square error is determined as follows : Wherein, the more the value of accuracy rate tends to 1, the more the value of root mean square error tends to 0, indicating that the prediction performance of the network is better; (11) Assess the diagnostic results of cognitive level The degree of consistency (DOA) is determined by the following formula: ， wherein, representing a learner cognitive state, on a knowledge point, representing a learner cognitive state, on a knowledge point; A higher DOA value indicates that the network has better diagnostic performance in assessing learners' cognitive levels. 2.The learning cognitive diagnosis method based on multi-layer attention mechanism according to claim 1, characterized in that In step (3), the data augmentation module is constructed using the following method: Data partitioning module in data augmentation module The construction method is as follows: Wherein, KF represents the answer record set R of the learner β-fold data division of R u , β is an integer between 2 and 6; β data division results of R u ; Data aggregation module in data augmentation module The construction method is as follows: Where CB represents the data aggregation function, which is derived from CB( )get{( - ,{R u1 }), ( - , {R u2 }), … , ( - , {R u β })}, where β takes the value of an integer from 2 to 6. - Recorded as Represents learners before data augmentation The set of historical answer records obtained after processing by the data augmentation module, 1 ≤ j ≤ β, Recorded as Represents learners before data augmentation The j-th set of verification answer records obtained after data augmentation has its number of learners increased from I to β.I after processing by the data augmentation module. Indicates learners after data augmentation A set of historical answer records, 1 ≤ i ≤ β.I, Corresponding to The relationship between i, u, and j can be expressed as i = (u-1) × β + j. Indicates learners after data augmentation The set of verification answer records Corresponding to . 3.The learning cognitive diagnosis method based on multi-layer attention mechanism according to claim 1, characterized in that In step (3) of constructing the multi-layer attention neural network, the method for constructing the guessing error module is as follows: The construction method of the guessing error handling module is as follows: wherein, represents the score prediction of the learner on the exercise after considering the guess probability weight. The method for constructing the error handling module of the guessing error module is as follows: wherein, represents the score prediction of a learner on the practice problems after considering the probability of error weights; The construction method of the information fusion module for the guessing error module is as follows: 。

Images