Knowledge tracking depth depolarization method, electronic equipment and storage medium

Abstract

The invention discloses a knowledge tracking depth depolarization method, electronic equipment and a storage medium. The method comprises the steps of 1, data acquisition; 2, building a knowledge tracking network, an adversarial embedding generation network and an adversarial knowledge tracking network; and 3, alternately training the knowledge tracking network, the adversarial embedding generation network and the adversarial knowledge tracking network to obtain an optimal knowledge tracking model. The method combines adversarial data enhancement and loss reweighting, generates adversarial deviation data, reduces the learning rate of deviation samples in the model training process, effectively prevents the model from using the correlation between the deviation attribute and the label as a prediction shortcut, alleviates the deviation problem of the deep knowledge tracking model, and improves the prediction accuracy of the deep knowledge tracking model. Therefore, the knowledge state of the student can be evaluated more accurately.

Description

Technical Field

[0001] The present invention belongs to the field of data mining, and relates to knowledge tracking and model debiasing technology, and specifically to a knowledge tracking deep debiasing method, an electronic device and a storage medium. Background Art

[0002] Knowledge tracking technology is a basic task in intelligent education and personalized learning. Its goal is to predict students' future performance given their historical exercise records. Thanks to the rapid development of deep learning technology and the large amount of data accumulated by online learning platforms, deep knowledge tracking models have achieved great success. However, existing advanced deep knowledge tracking models have ignored the problem of bias. The biased samples in the data set associate the biased attributes (the accuracy of students' historical answers and the accuracy of target exercises) with the labels. The model trained on the biased data set will learn and amplify the false association between the biased attributes and the labels, resulting in the bias problem of the model. Since the biased attributes of biased samples are highly correlated with the labels, the model can use the biased association to make accurate predictions for the biased samples. Therefore, using biased data sets to evaluate the knowledge tracking model cannot truly reflect the model's ability to extract students' knowledge status. Therefore, it is very necessary to construct an unbiased set to evaluate the existing model unbiasedly and design a debiased training method to improve the model's ability to resist bias under the condition of training only with biased data sets. Summary of the invention

[0003] In order to address the deficiencies of the above-mentioned prior art, the present invention proposes a knowledge tracking deep debiasing method, an electronic device and a storage medium, in order to alleviate the bias problem of the deep knowledge tracking model, thereby achieving a more accurate assessment of students' knowledge status.

[0004] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0005] A knowledge tracking deep debiasing method of the present invention is characterized in that it comprises the following steps:

[0006] Step 1: Data collection;

[0007] Get the student collection , where s n represents the nth student, N is the number of students;

[0008] Get Exercise Set , where q i represents the i-th exercise, and I is the number of exercises;

[0009] Let the nth student s n For the i-th exercise q i The answer is rn,i , if r n,i =1, indicating the answer is correct, if r n,i =0, indicating an incorrect answer;

[0010] Define the nth student s n The previous T historical answer records are H n , T = [ ( q n , 1 , r n , 1 ) , ( q n , 2 , r n , 2 ) , … , ( q n , t , r n , t ) , … , ( q n , T , r n , T ) ] ,in, represents the nth student s n The tth historical answer record, q n,t represents the nth student s n The tth history exercise, r n,t represents the nth student s n The tth historical answer result of , T is the nth student s n Number of historical questions answered;

[0011] Define the nth student s n The tth answer sample is , where H n,t-1 represents the nth student s n The previous t-1 historical answer records, For the nth student s n When answering the history question for the tth time, the correct rate of the most recent w answers is season ; For the nth student s n The tth history exercise q n,t The correct rate of all students in the previous T historical answer records, when q n,t In the previous T historical answer records of all students, the number of times it appears is less than L times, let ;

[0012] like Satisfy the conditions , then Defined as the historical answer performance bias guide sample, denoted as ;

[0013] like Satisfy the conditions , then Defined as the historical answer performance deviation conflict sample, denoted as ;

[0014] like Satisfy the conditions , then Defined as the sample guided by the deviation of the exercise answer, denoted as ;

[0015] like Satisfy the conditions , then Defined as a conflicting sample of exercise answer deviation, denoted as ; Among them, a th is the threshold for judging sample deviation;

[0016] The unbiased set of historical answer performance is constructed from the conflicting samples of historical answer performance of all students, denoted as ;

[0017] The unbiased set of exercise answers is constructed from the conflicting samples of all students' exercise answers, denoted as ;

[0018] Step 2: Build a knowledge tracking network, adversarial embedding generation network, and adversarial knowledge tracking network;

[0019] Step 2.1, the knowledge tracking network includes: embedding layer E, knowledge state extraction layer and prediction layer , and used for the nth student s n Exercises for Sequence Data H n,T Process and get the nth student s n The probability of answering the next exercise correctly, thereby constructing the loss of the knowledge tracking network , used to train the embedding layer E and the knowledge state extraction layer and prediction layer , and get the first trained embedding layer E 1 , the first trained knowledge state extraction layer And the prediction layer trained initially ;

[0020] Step 2.2, the embedding layer E trained initially 1 , the first trained knowledge state extraction layer And the prediction layer trained initially Construct an adversarial embedding generation network and use it to train the nth student s n Exercises for Sequence Data H n,T Process and get the nth student s n The probability of correct adversarial generation in the next answer exercise is used to construct the loss of the adversarial embedding generation network , used to train the initially trained embedding layer E 1 , and get the retrained embedding layer E 2 ;

[0021] Step 2.3, the embedded layer E trained again 2 , the first trained knowledge state extraction layer And the prediction layer trained initially Construct an adversarial knowledge tracking network and use it to track the nth student s n Exercises for Sequence Data H n,T Process and get the nth student s n The correct adversarial prediction probability of the next answer exercise is constructed to construct the loss of the adversarial knowledge tracking network , used to train the initially trained knowledge state extraction layer And the prediction layer trained initially , and obtain the retrained knowledge state extraction layer And the retrained prediction layer ;

[0022] Step 3: E 1 , and Assign corresponding values ​​to E, and Finally, the knowledge tracking network, adversarial embedding generation network and adversarial knowledge tracking network are trained alternately according to the process from step 2.1 to step 2.3 to continuously adjust the network parameters until the prediction accuracy of the knowledge tracking network no longer improves, thereby obtaining the optimal knowledge tracking model and using it to process the unbiased set of historical answer performance and the unbiased set of exercise answers to obtain the correct prediction results of students' exercise answers after debiasing.

[0023] The knowledge tracing deep debiasing method of the present invention is also characterized in that step 2.1 comprises:

[0024] Step 2.1.1, the embedding layer E is used for the nth student s n Exercises for Sequence Data H n,T Process and output the nth student s n The embedded representation sequence of the second to Tth answer questions E M B n q = [ e m b n , 2 q , e m b n , 3 q , … , e m b n , t q , … , e m b n , T q ] and the nth student s n The embedded representation sequence of the interactive behaviors of answering exercises before T-1 times E M B n q r = [ e m b n , 1 q r , e m b n , 2 q r , … , e m b n , t q r , … , e m b n , T − 1 q r ] ;in, represents the nth student s n Answer exercise q for the tth time n,t The embedded representation of represents the nth student s n Answer exercise q for the tth time n,t And the answer result n,t Embedded representation of

[0025] The knowledge state extraction layer right Process and output the nth student s n The knowledge state representation sequence when answering the exercises from the 2nd to the Tth time E M B n h = [ e m b n , 2 h , e m b n , 3 h , … , e m b n , t h , … , e m b n , T h ] ;in, represents the nth student s n The knowledge status representation when answering the exercise for the tth time;

[0026] The prediction layer right and Process and output the nth student s n The probability of answering the next exercise correctly [ r ^ n , 2 , r ^ n , 3 , … , r ^ n , t , … , r ^ n , T ] ;in, represents the nth student s n The probability of answering the exercise correctly the tth time;

[0027] Step 2.1.2, based on the degree of sample deviation, use formula (1) to calculate the sample The loss weight ;

[0028] (1)

[0029] In formula (1), if To guide the sample based on historical performance bias, we use formula (2) to calculate Unbiased loss weights for historical answer performance , otherwise, let ;like To guide the sample for the exercise answer bias, we use (3) to calculate Answers to exercises on unbiased loss weights , otherwise, let ;

[0030] (2)

[0031] (3)

[0032] In formula (2) and formula (3), w m It is the maximum reduction in the loss weight of the sample guided by the bias;

[0033] Step 2.1.3, use formula (4) to construct the loss function of the knowledge tracking network ;

[0034] (4)

[0035] Step 2.1.4, using the loss function of the knowledge tracking network Training embedding layer E, knowledge state extraction layer and prediction layer , get the initially trained embedding layer E 1 , the first trained knowledge state extraction layer And the prediction layer trained initially .

[0036] Furthermore, the step 2.2 includes:

[0037] Step 2.2.1, the initially trained embedding layer E 1 For the nth student s n Exercises for Sequence Data H n,T Process and output the nth student s n The adversarial generation of embedding representation sequences for the 2nd to Tth answer exercises E M B n q , a d v = [ e m b n , 2 q , a d v , e m b n , 3 q , a d v , … , e m b n , t q , a d v , … , e m b n , T q , a d v ] and the nth student s n The adversarial output embedding representation sequence of the interactive behavior of the first T-1 times of answering exercises E M B n q r , a d v = [ e m b n , 1 q r , a d v , e m b n , 2 q r , a d v , … , e m b n , t q r , a d v , … , e m b n , T − 1 q r , a d v ] ;in, represents the nth student s n Answer exercise q for the tth time n,t The adversarial generation of embedding representations, represents the nth student s n Answer exercise q for the tth time n,t And the answer result n,t adversarial generation of embedding representations;

[0038] The initially trained knowledge state extraction layer right Process and output the nth student s n The adversarial generation of knowledge state representation sequence when answering exercises from the 2nd to the Tth time E M B n h , a d v = [ e m b n , 2 h , a d v , e m b n , 3 h , a d v , … , e m b n , t h , a d v , … , e m b n , T h , a d v ] ;in, represents the nth student s n The adversarial knowledge state representation when answering the exercise for the tth time;

[0039] The initially trained prediction layer right and Process and output the nth student s n The probability of correct adversarial generation in the next answer exercise [ r ^ a d v n , 2 , r ^ a d v n , 3 , … , r ^ a d v n , t , … , r ^ a d v n , T ] ;in, represents the nth student s n The correct adversarial generation probability of answering the exercise for the tth time;

[0040] Step 2.2.2: Use formula (5) to construct the loss function of the adversarial embedding generation network. :

[0041] (5)

[0042] In formula (5), is the mean square error loss The weight of is the prediction loss of the adversarial embedding generation network and is obtained by formula (6); is the mean square error loss between the student's knowledge state representation and the adversarial knowledge state representation, and is obtained by formula (7);

[0043] (6)

[0044] (7)

[0045] In formula (7), Indicates L 2 norm; Indicates the first training, the knowledge state extraction layer trained initially Output the nth student s n The adversarial generation of knowledge state representation sequence when answering the exercise for the tth time; Indicates that except for the first training, the knowledge state extraction layer trained initially is Output the nth student s n The adversarial generation of knowledge state representation sequence when answering the exercise for the tth time;

[0046] Step 2.2.3, Fixation and , and use the loss function of the adversarial embedding generation network For E 1 Train and get the retrained embedding layer E 2 .

[0047] Further, the step 2.3 includes:

[0048] Step 2.3.1, the retrained embedding layer E 2 For the nth student s n Exercises for Sequence Data H n,T Process and output the nth student s n The adversarial prediction embedding representation sequence of the 2nd to Tth answer questions E M B n q , ∗ = [ e m b n , 2 q , ∗ , e m b n , 3 q , ∗ , … , e m b n , t q , ∗ , … , e m b n , T q , ∗ ] and the nth student s n The adversarial prediction embedding representation sequence of the interactive behavior of the first T-1 answering exercises E M B n q r , ∗ = [ e m b n , 1 q r , ∗ , e m b n , 2 q r , ∗ , … , e m b n , t q r , ∗ , … , e m b n , T − 1 q r , ∗ ] ;in, represents the nth student s n Answer exercise q for the tth time n,t The adversarial prediction embedding representation of represents the nth student s n Answer exercise q for the tth time n,t And the answer result n,t adversarial prediction embedding representation;

[0049] The initially trained knowledge state extraction layer right Process and output the nth student s n The sequence of adversarial prediction knowledge states when answering exercises from the 2nd to the Tth time E M B n h , * = [ e m b n , 2 h , * , e m b n , 3 h , * , … , e m b n , t h , * , … , e m b n , T h , * ] ;in, represents the nth student s n The adversarial prediction knowledge state representation when answering the exercise for the tth time;

[0050] The initially trained prediction layer right and Process and output the nth student s n The correct adversarial prediction probability for the next answer exercise [ r ^ * n , 2 , r ^ * n , 3 , … , r ^ * n , t , … , r ^ * n , T ] ;in, represents the nth student s n The correct adversarial prediction probability of answering the exercise for the tth time;

[0051] Step 2.3.2: Use formula (8) to construct the loss function of the adversarial knowledge tracking network. ;

[0052] (8)

[0053] Step 2.3.3, fix E 2 Parameters of the network, using adversarial knowledge to track the loss function of the network Extract the knowledge state layer from the initial training And the prediction layer trained initially Train to obtain a retrained knowledge state extraction layer And the retrained prediction layer .

[0054] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the knowledge tracking deep debiasing method, and the processor is configured to execute the program stored in the memory.

[0055] The present invention provides a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. The characteristic is that when the computer program is executed by a processor, the steps of the knowledge tracking deep debiasing method are executed.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. The present invention proposes a method for dividing unbiased samples in a knowledge tracking dataset, namely, judging whether a sample is an unbiased sample by the association between its bias attributes and labels, thereby realizing the construction of an unbiased set for more accurately measuring the ability of the knowledge tracking model to evaluate students' knowledge status.

[0058] 2. The present invention proposes an adversarial data enhancement method and a loss reweighting method suitable for knowledge tracking. First, the adversarial prediction loss is used to generate enhanced data that is adversarial to the bias. Then, the loss weight is calculated based on the degree of bias of the sample. The weighted prediction loss is used on the original data and enhanced data to train the knowledge tracking model to reduce the learning rate of the model for biased samples, thereby achieving deep debiasing of the model and more accurately evaluating the students' knowledge status. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a block diagram of the knowledge tracking deep debiasing method of the present invention. DETAILED DESCRIPTION

[0060] In this embodiment, Figure 1 As shown in FIG, a deep debiasing method for knowledge tracking is to generate data against bias by combining adversarial data enhancement and loss reweighting, and reduce the learning rate of biased samples during model training, effectively avoiding the model from using the association between biased attributes and labels as a prediction shortcut, so as to alleviate the bias problem of the deep knowledge tracking model; specifically, the method includes the following steps:

[0061] Step 1: Data collection;

[0062] Get the student collection , where s n represents the nth student, N is the number of students;

[0063] Get Exercise Set , where q i represents the i-th exercise, and I is the number of exercises;

[0064] Let the nth student s n For the i-th exercise q i The answer is r n,i , if r n,i=1, indicating the answer is correct, if r n,i =0, indicating an incorrect answer;

[0065] Define the nth student s n The previous T historical answer records are H n , T = [ ( q n , 1 , r n , 1 ) , ( q n , 2 , r n , 2 ) , … , ( q n , t , r n , t ) , … , ( q n , T , r n , T ) ] ,in, represents the nth student s n The tth historical answer record, q n,t represents the nth student s n The tth history exercise, r n,t represents the nth student s n The tth historical answer result of , T is the nth student s n The number of historical answers.

[0066] Define the nth student s n The tth answer sample is , where H n,t-1 represents the nth student s n The previous t-1 historical answer records, For the nth student s n The correctness rate of the most recent w answers in the tth historical answering. In this embodiment, w=10. season ; For the nth student s n The tth history exercise q n,t The correct rate of all students in the previous T historical answer records, when q n,t In the previous T historical answer records of all students, the number of times it appears is less than L times, let ; In this embodiment, L=3.

[0067] like Satisfy the conditions , then Defined as the historical answer performance bias guide sample, denoted as ;

[0068] like Satisfy the conditions , then Defined as the historical answer performance deviation conflict sample, denoted as ;

[0069] like Satisfy the conditions , then Defined as the sample guided by the deviation of the exercise answer, denoted as ;

[0070] like Satisfy the conditions , then Defined as a conflicting sample of exercise answer deviation, denoted as ; Among them, a th is the threshold for determining sample deviation. In this embodiment, a th =0.4.

[0071] The unbiased set of historical answer performance is constructed from the conflicting samples of historical answer performance of all students, denoted as ;

[0072] The unbiased set of exercise answers is constructed from the conflicting samples of all students' exercise answers, denoted as .

[0073] Step 2: Build a knowledge tracking network, adversarial embedding generation network, and adversarial knowledge tracking network;

[0074] Step 2.1, the knowledge tracking network includes: embedding layer E, knowledge state extraction layer and prediction layer , and used for the nth student s n Exercises for Sequence Data H n,T Process and get the nth student s n The probability of answering the next exercise correctly, thereby constructing the loss of the knowledge tracking network , used to train the embedding layer E and the knowledge state extraction layer and prediction layer , and get the first trained embedding layer E 1 , the first trained knowledge state extraction layer And the prediction layer trained initially .

[0075] Step 2.1.1, embedding layer E for the nth student s n Exercises for Sequence Data H n,T Process and output the nth student s n The embedded representation sequence of the second to Tth answer questions E M B n q = [ e m b n , 2 q , e m b n , 3 q , … , e m b n , t q , … , e m b n , T q ] and the nth student s n The embedded representation sequence of the interactive behaviors of answering exercises before T-1 times E M B n q r = [ e m b n , 1 q r , e m b n , 2 q r , … , e m b n , t q r , … , e m b n , T − 1 q r ] ;in, represents the nth student s n Answer exercise q for the tth time n,t The embedded representation of represents the nth student s n Answer exercise q for the tth time n,t And the answer result n,t Embedded representation of

[0076] Knowledge State Extraction Layer right Process and output the nth student s n The knowledge state representation sequence when answering the exercises from the 2nd to the Tth time E M B n h = [ e m b n , 2 h , e m b n , 3 h , … , e m b n , t h , … , e m b n , T h ] ;in, represents the nth student s n The knowledge status representation when answering the exercise for the tth time;

[0077] Prediction Layer right and Process and output the nth student s n The probability of answering the next exercise correctly [ r ^ n , 2 , r ^ n , 3 , … , r ^ n , t , … , r ^ n , T ] ;in, represents the nth student s n The probability of answering the exercise correctly the tth time.

[0078] Step 2.1.2, based on the degree of sample deviation, use formula (1) to calculate the sample The loss weight ;

[0079] (1)

[0080] In formula (1), if To guide the sample based on historical performance bias, we use formula (2) to calculate Unbiased loss weights for historical answer performance , otherwise, let ;like To guide the sample for the exercise answer bias, we use (3) to calculate Answers to exercises on unbiased loss weights , otherwise, let ;

[0081] (2)

[0082] (3)

[0083] In formula (2) and formula (3), w m is the maximum reduction in the sample loss weight guided by the deviation. In this embodiment, w m =0.7.

[0084] Step 2.1.3, use formula (4) to construct the loss function of the knowledge tracking network ;

[0085] (4)

[0086] Step 2.1.4, using the loss function of the knowledge tracking network Training embedding layer E, knowledge state extraction layer and prediction layer , get the initially trained embedding layer E 1 , the first trained knowledge state extraction layer And the prediction layer trained initially .

[0087] Step 2.2, the embedding layer E trained initially 1 , the first trained knowledge state extraction layer And the prediction layer trained initially Construct an adversarial embedding generation network and use it to train the nth student s n Exercises for Sequence Data H n,T Process and get the nth student s n The probability of correct adversarial generation in the next answer exercise is constructed to construct the loss of the adversarial embedding generation network , used to train the initially trained embedding layer E 1 , and get the retrained embedding layer E 2 ;

[0088] Step 2.2.1, the first trained embedding layer E 1 For the nth student s n Exercises for Sequence Data H n,T Process and output the nth student s n The adversarial generation of embedding representation sequences for the 2nd to Tth answer exercises E M B n q , a d v = [ e m b n , 2 q , a d v , e m b n , 3 q , a d v , … , e m b n , t q , a d v , … , e m b n , T q , a d v ] and the nth student s n The adversarial output embedding representation sequence of the interactive behavior of the first T-1 times of answering exercises E M B n q r , a d v = [ e m b n , 1 q r , a d v , e m b n , 2 q r , a d v , … , e m b n , t q r , a d v , … , e m b n , T − 1 q r , a d v ] ;in, represents the nth student s n Answer exercise q for the tth time n,t The adversarial generation of embedding representations, represents the nth student s n Answer exercise q for the tth time n,t And the answer result n,t adversarial generation of embedding representations;

[0089] Initially trained knowledge state extraction layer right Process and output the nth student s nThe adversarial generation of knowledge state representation sequence when answering exercises from the 2nd to the Tth time E M B n h , a d v = [ e m b n , 2 h , a d v , e m b n , 3 h , a d v , … , e m b n , t h , a d v , … , e m b n , T h , a d v ] ;in, represents the nth student s n The adversarial knowledge state representation when answering the exercise for the tth time;

[0090] Initially trained prediction layer right and Process and output the nth student s n The probability of correct adversarial generation in the next answer exercise [ r ^ a d v n , 2 , r ^ a d v n , 3 , … , r ^ a d v n , t , … , r ^ a d v n , T ] ;in, represents the nth student s n The correct adversarial generation probability for answering the exercise for the tth time.

[0091] Step 2.2.2: Use formula (5) to construct the loss function of the adversarial embedding generation network. :

[0092] (5)

[0093] In formula (5), is the mean square error loss The weight of, in this embodiment, =1, is the prediction loss of the adversarial embedding generation network and is obtained by formula (6); is the mean square error loss between the student's knowledge state representation and the adversarial knowledge state representation, and is obtained by formula (7);

[0094] (6)

[0095] (7)

[0096] In formula (7), Indicates L 2 norm; Indicates the first training, the knowledge state extraction layer trained initially Output the nth student s n The adversarial generation of knowledge state representation sequence when answering the exercise for the tth time; Indicates that except for the first training, the knowledge state extraction layer trained initially is Output the nth student s n The adversarial generation of knowledge state representation sequence when answering the exercise for the tth time;

[0097] Step 2.2.3, Fixation and , and use the loss function of the adversarial embedding generation network For E 1 Train and get the retrained embedding layer E 2 .

[0098] Step 2.3, the embedded layer E trained again 2 , the first trained knowledge state extraction layer And the prediction layer trained initially Construct an adversarial knowledge tracking network and use it to track the nth student s n Exercises for Sequence Data H n,T Process and get the nth student s n The correct adversarial prediction probability of the next answer exercise is constructed to construct the loss of the adversarial knowledge tracking network , used to train the initially trained knowledge state extraction layer And the prediction layer trained initially , and obtain the retrained knowledge state extraction layer And the retrained prediction layer ;

[0099] Step 2.3.1, retrain the embedding layer E 2 For the nth student s n Exercises for Sequence Data H n,T Process and output the nth student s n The adversarial prediction embedding representation sequence of the 2nd to Tth answer questions E M B n q , ∗ = [ e m b n , 2 q , ∗ , e m b n , 3 q , ∗ , … , e m b n , t q , ∗ , … , e m b n , T q , ∗ ] and the nth student s n The adversarial prediction embedding representation sequence of the interactive behavior of the first T-1 answering exercises E M B n q r , ∗ = [ e m b n , 1 q r , ∗ , e m b n , 2 q r , ∗ , … , e m b n , t q r , ∗ , … , e m b n , T − 1 q r , ∗ ] ;in, represents the nth student s n Answer exercise q for the tth time n,t The adversarial prediction embedding representation of represents the nth student s n Answer exercise q for the tth time n,t And the answer result n,t Adversarial prediction embedding representation.

[0100] Initially trained knowledge state extraction layer right Process and output the nth student s n The sequence of adversarial prediction knowledge states when answering exercises from the 2nd to the Tth time E M B n h , * = [ e m b n , 2 h , * , e m b n , 3 h , * , … , e m b n , t h , * , … , e m b n , T h , * ] ;in, represents the nth student s n The adversarial prediction knowledge state representation when answering the exercise for the tth time;

[0101] Initially trained prediction layer right and Process and output the nth student s n The correct adversarial prediction probability for the next answer exercise [ r ^ * n , 2 , r ^ * n , 3 , … , r ^ * n , t , … , r ^ * n , T ] ;in, represents the nth student s n The adversarial prediction probability of answering the exercise correctly for the tth time.

[0102] Step 2.3.2: Use formula (8) to construct the loss function of the adversarial knowledge tracking network. ;

[0103] (8)

[0104] Step 2.3.3, fix E 2 Parameters of the network, using adversarial knowledge to track the loss function of the network Extract the knowledge state layer from the initial training And the prediction layer trained initially Train to obtain a retrained knowledge state extraction layer And the retrained prediction layer .

[0105] Step 3: E 1 , and Assign corresponding values ​​to E, and Finally, the knowledge tracking network, adversarial embedding generation network and adversarial knowledge tracking network are trained alternately according to the process from step 2.1 to step 2.3 to continuously adjust the network parameters until the prediction accuracy of the knowledge tracking network no longer improves, thereby obtaining the optimal knowledge tracking model and using it to process the unbiased set of historical answer performance and the unbiased set of exercise answers to obtain the biased prediction results of the student exercise answer accuracy.

[0106] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0107] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0108] The test results of the present invention are further described in conjunction with the following charts:

[0109] In order to verify the debiasing effect of the present invention on the existing knowledge tracing model, two commonly used benchmark models qDKT (deep knowledge tracing using exercise information) and AKT (context-aware knowledge tracing) were selected to conduct experiments on three commonly used knowledge tracing datasets: Assist2009, Ednet-kt1, and Statics2011. Ours represents the application of the method proposed in the present invention on the corresponding benchmark deep knowledge tracing model, and for fair comparison, the settings of hyperparameters, including model hyperparameters and training hyperparameters, such as learning rate, L2 regularization coefficient, mini-batch size, etc., are consistent between Ours and the benchmark model. The historical answer performance unbiased set D and the exercise answer unbiased set D of the three datasets are used respectively. q An unbiased evaluation is performed on the optimal knowledge tracking model obtained after training, using AUC (Area Under the Curve) and ACC (Accuracy) as evaluation indicators. The experimental results are shown in Tables 1 and 2.

[0110] Table 1 Performance comparison between the baseline model and the proposed method on the unbiased set of historical answer performance

[0111]

[0112] Table 2 Performance comparison between the baseline model and the proposed method on the unbiased set of exercise answers

[0113]

[0114] It can be seen from Table 1 and Table 2 that the method proposed in the present invention significantly improves the AUC and ACC indicators of the baseline model on the unbiased set of historical answer performance and the unbiased set of exercise answers, verifying the effectiveness of the present invention in debiasing the knowledge tracking model.

Claims

1. A knowledge tracking deep debiasing method, characterized in that: The following steps are involved: Step 1: Data collection; Get the student collection , where s n represents the nth student, N is the number of students; Get Exercise Set , where q i represents the i-th exercise, and I is the number of exercises; Let the nth student s n For the i-th exercise q i The answer is r n,i , if r n,i =1, indicating the answer is correct, if r n,i =0, indicating an incorrect answer; Define the nth student s n The previous T historical answer records are ,in, represents the nth student s n The tth historical answer record, q n,t represents the nth student s n The tth history exercise, r n,t represents the nth student s n The tth historical answer result of , T is the nth student s n Number of historical questions answered; Define the nth student s n The tth answer sample is , where H n,t-1 represents the nth student s n The previous t-1 historical answer records, For the nth student s n When answering the history question for the tth time, the correct rate of the most recent w answers is season ; For the nth student s n The tth history exercise q n,t The correct rate of all students in the previous T historical answer records, when q n,t In the previous T historical answer records of all students, the number of times it appears is less than L times, let ; like Satisfy the conditions , then Defined as the historical answer performance bias guide sample, denoted as ; like Satisfy the conditions , then Defined as the historical answer performance deviation conflict sample, denoted as ; like Satisfy the conditions , then Defined as the sample guided by the deviation of the exercise answer, denoted as ; like Satisfy the conditions , then Defined as a conflicting sample of exercise answer deviation, denoted as ; Among them, a th is the threshold for judging sample deviation; The unbiased set of historical answer performance is constructed from the conflicting samples of historical answer performance of all students, denoted as ; The unbiased set of exercise answers is constructed from the conflicting samples of all students' exercise answers, denoted as ; Step 2: Build a knowledge tracking network, adversarial embedding generation network, and adversarial knowledge tracking network; Step 2.1, the knowledge tracking network includes: embedding layer E, knowledge state extraction layer and prediction layer , and used for the nth student s n Exercises for Sequence Data H n,T Process and get the nth student s n The probability of answering the next exercise correctly, thereby constructing the loss of the knowledge tracking network , used to train the embedding layer E and the knowledge state extraction layer and prediction layer , and get the first trained embedding layer E 1 , the first trained knowledge state extraction layer And the prediction layer trained initially ; Step 2.2, the embedding layer E trained initially 1 , the first trained knowledge state extraction layer And the prediction layer trained initially Construct an adversarial embedding generation network and use it to train the nth student s n Exercises for Sequence Data H n,T Process and get the nth student s n The probability of correct adversarial generation in the next answer exercise is constructed to construct the loss of the adversarial embedding generation network , used to train the initially trained embedding layer E 1 , and get the retrained embedding layer E 2 ; Step 2.3, the embedded layer E trained again 2 , the first trained knowledge state extraction layer And the prediction layer trained initially Construct an adversarial knowledge tracking network and use it to track the nth student s n Exercises for Sequence Data H n,T Process and get the nth student s n The correct adversarial prediction probability of the next answer exercise is constructed to construct the loss of the adversarial knowledge tracking network , used to train the initially trained knowledge state extraction layer And the prediction layer trained initially , and obtain the retrained knowledge state extraction layer And the retrained prediction layer ; Step 3: E 1 , and Assign corresponding values ​​to E, and Finally, the knowledge tracking network, adversarial embedding generation network and adversarial knowledge tracking network are trained alternately according to the process from step 2.1 to step 2.3 to continuously adjust the network parameters until the prediction accuracy of the knowledge tracking network no longer improves, thereby obtaining the optimal knowledge tracking model and using it to process the unbiased set of historical answer performance and the unbiased set of exercise answers to obtain the correct prediction results of students' exercise answers after debiasing.

2. The knowledge tracing deep debiasing method according to claim 1, characterized in that: The step 2.1 comprises: Step 2.1.1, the embedding layer E is used for the nth student s n Exercises for Sequence Data H n,T Process and output the nth student s n The embedded representation sequence of the second to Tth answer questions and the nth student s n The embedded representation sequence of the interactive behaviors of answering exercises before T-1 times ;in, represents the nth student s n Answer exercise q for the tth time n,t The embedded representation of represents the nth student s n Answer exercise q for the tth time n,t And the answer result n,t Embedded representation of The knowledge state extraction layer right Process and output the nth student s n The knowledge state representation sequence when answering the exercises from the 2nd to the Tth time ;in, represents the nth student s n The knowledge status representation when answering the exercise for the tth time; The prediction layer right and Process and output the nth student s n The probability of answering the next exercise correctly ;in, represents the nth student s n The probability of answering the exercise correctly the tth time; Step 2.1.2, based on the degree of sample deviation, use formula (1) to calculate the sample The loss weight ; (1) In formula (1), if To guide the sample based on historical performance bias, we use formula (2) to calculate Unbiased loss weights for historical answer performance , otherwise, let ;like To guide the sample for the exercise answer bias, use (3) to calculate Answers to exercises on unbiased loss weights , otherwise, let ; (2) (3) In formula (2) and formula (3), w m It is the maximum reduction in the loss weight of the sample guided by the bias; Step 2.1.3, use formula (4) to construct the loss function of the knowledge tracking network ; (4) Step 2.1.4, using the loss function of the knowledge tracking network Training embedding layer E, knowledge state extraction layer and prediction layer , get the initially trained embedding layer E 1 , the first trained knowledge state extraction layer And the prediction layer trained initially .

3. The knowledge tracking deep debiasing method according to claim 2, characterized in that: The step 2.2 comprises: Step 2.2.1, the initially trained embedding layer E 1 For the nth student s n Exercises for Sequence Data H n,T Process and output the nth student s n The adversarial generation of embedding representation sequences for the 2nd to Tth answer exercises and the nth student s n The adversarial output embedding representation sequence of the interactive behavior of the first T-1 times of answering exercises ;in, represents the nth student s n Answer exercise q for the tth time n,t The adversarial generation of embedding representations, represents the nth student s n Answer exercise q for the tth time n,t And the answer result n,t adversarial generation of embedding representations; The initially trained knowledge state extraction layer right Process and output the nth student s n The adversarial generation of knowledge state representation sequence when answering exercises from the 2nd to the Tth time ;in, represents the nth student s n The adversarial knowledge state representation when answering the exercise for the tth time; The initially trained prediction layer right and Process and output the nth student s n The probability of correct adversarial generation in the next answer exercise ;in, represents the nth student s n The correct adversarial generation probability of answering the exercise for the tth time; Step 2.2.2: Use formula (5) to construct the loss function of the adversarial embedding generation network. : (5) In formula (5), is the mean square error loss The weight of is the prediction loss of the adversarial embedding generation network and is obtained by formula (6); is the mean square error loss between the student's knowledge state representation and the adversarial knowledge state representation, and is obtained by formula (7); (6) (7) In formula (7), represents the L2 norm; Indicates the first training, the knowledge state extraction layer trained initially Output the nth student s n The adversarial generation of knowledge state representation sequence when answering the exercise for the tth time; Indicates that except for the first training, the knowledge state extraction layer trained initially is Output the nth student s n The adversarial generation of knowledge state representation sequence when answering the exercise for the tth time; Step 2.2.3, Fixation and , and use the loss function of the adversarial embedding generation network For E 1 Train and get the retrained embedding layer E 2 .

4. The knowledge tracking deep debiasing method according to claim 3, characterized in that: The step 2.3 comprises: Step 2.3.1, the retrained embedding layer E 2 For the nth student s n Exercises for Sequence Data H n,T Process and output the nth student s n The adversarial prediction embedding representation sequence of the 2nd to Tth answer questions and the nth student s n The adversarial prediction embedding representation sequence of the interactive behavior of the first T-1 answering exercises ;in, represents the nth student s n Answer exercise q for the tth time n,t The adversarial prediction embedding representation of represents the nth student s n Answer exercise q for the tth time n,t And the answer result n,t adversarial prediction embedding representation; The initially trained knowledge state extraction layer right Process and output the nth student s n The sequence of adversarial prediction knowledge states when answering exercises from the 2nd to the Tth time ;in, represents the nth student s n The adversarial prediction knowledge state representation when answering the exercise for the tth time; The initially trained prediction layer right and Process and output the nth student s n The correct adversarial prediction probability for the next answer exercise ;in, represents the nth student s n The correct adversarial prediction probability of answering the exercise for the tth time; Step 2.3.2: Use formula (8) to construct the loss function of the adversarial knowledge tracking network. ; (8) Step 2.3.3, fix E 2 Parameters of the network, using adversarial knowledge to track the loss function of the network Extract the knowledge state layer from the initial training And the prediction layer trained initially Train to obtain a retrained knowledge state extraction layer And the retrained prediction layer .

5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the knowledge tracking deep debiasing method described in any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the knowledge tracing deep debiasing method according to any one of claims 1 to 4 are executed.

Images