Knowledge tracking method and system based on cognitive fluctuation enhancement

Abstract

The invention relates to the technical field of knowledge tracking, in particular to a knowledge tracking method and system based on cognitive fluctuation enhancement, and the method comprises the steps: obtaining an answer interaction record of a student; the question answering interaction record is input into a preset knowledge tracking model, the next question answering situation of the student is predicted, the knowledge tracking model is used for extracting cognitive features in the question answering interaction record, cognitive fluctuation enhancement is conducted after the cognitive features are decomposed, and then the next question answering situation is predicted. According to the method, the comprehensive modeling ability and the model length generalization ability of long-term and short-term cognitive features can be enhanced, and the accurate prediction effect on the answering condition of the student is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge tracing, and in particular to a knowledge tracing method and system based on cognitive fluctuation enhancement. Background Art

[0002] Knowledge tracing (KT) is a key sequence prediction task that aims to use students' historical learning interaction data to predict their performance on future problems. By modeling human cognitive behavior, KT can deeply understand the cognitive process and provide technical support for the advancement of personalized learning and intelligent education. With the significant progress made in the development of deep learning and innovative technologies, a large number of efficient KT technologies have emerged. DKT (Deep knowledge tracing) first introduced recurrent neural networks (RNN) to model students' learning process. DKVMN (Dynamic key-value memory networks) improves the accuracy of prediction by adding a dual memory structure. GKT (Graph-based Knowledge Tracing) uses graph neural networks (GNN) to structure knowledge into graphs. In recent years, attention-based models have become the mainstream in the field of KT. SAKT (Self Attentive Knowledge Tracing) uses a self-attention mechanism to deal with data sparsity and improve prediction accuracy. SAINT (Separated Self-Attentive Neural Knowledge Tracing) uses an encoder-decoder architecture to effectively model the relationship between practice and answering questions. The AKT (Attentiveknowledge tracing) model uses three attention modules to extract the latent relations of questions and interactions respectively to model student behaviors. And introduces relative time interval bias in the attention matrix to integrate latent learning factors. SimpleKT handles the KT task by modeling question-specific changes based on the Rasch model, and uses a common dot product attention function to extract time-aware information embedded in student learning interactions.

[0003] Defects and shortcomings of the existing technology:

[0004] In KT tasks, human cognitive behavior has the following two core characteristics:

[0005] 1. Long-term cognitive trend: reflects the stable process of gradual accumulation of knowledge status over time.

[0006] 2. Short-term cognitive fluctuations: brief cognitive changes caused by momentary factors such as forgetfulness, fatigue, inattention or speculation.

[0007] These two types of cognitive characteristics jointly affect students' knowledge status, so it is crucial to capture both types of characteristics in KT modeling. Although the existing KT model based on the attention mechanism can effectively capture long-term cognitive trends, it still has significant defects in modeling short-term cognitive fluctuations. These deficiencies are specifically manifested as follows:

[0008] 1. Short-term cognitive features are ignored. The current attention mechanism focuses too much on long-term cognitive trends, resulting in insufficient capture of the features of short-term cognitive fluctuations. This deficiency may cause the cognitive behavior features extracted by the model to be too smooth and unable to accurately reflect the students' instantaneous cognitive changes, thereby reducing the model's ability to express dynamic cognitive behaviors.

[0009] 2. The model’s generalization ability to sequence length is insufficient. The attention mechanism’s neglect of short-term cognitive fluctuation characteristics also affects its length generalization ability. When the length of test data exceeds the length of training data, the attention distribution of the existing model may be too dispersed, resulting in its inability to effectively predict the short-term cognitive fluctuation characteristics in unknown sequences, ultimately leading to performance degradation. In actual online learning systems, the length of students’ interaction data is often variable, and the existing model cannot generalize well to long sequence prediction scenarios after short sequence training, limiting its application effect in real scenarios.

[0010] 3. Insufficient dynamic adaptability of cognitive state modeling. The existing attention mechanism lacks the ability to dynamically decompose and adjust the weights of long-term cognitive trends and short-term cognitive fluctuations. This static modeling method cannot flexibly cope with the complexity of cognitive states changing over time, limiting the prediction accuracy of the model.

[0011] The existence of the above defects indicates that the current KT model still has room for improvement in comprehensively modeling students' cognitive behaviors, especially in optimizing the short-term fluctuation modeling ability and improving the generalization ability of the model sequence length. Therefore, the present invention proposes a knowledge tracking method and system based on cognitive fluctuation enhancement. Summary of the invention

[0012] The purpose of the present invention is to provide a knowledge tracking method and system based on cognitive fluctuation enhancement, to enhance the comprehensive modeling ability of long-term and short-term cognitive characteristics and the model length generalization ability, and to improve the accurate prediction effect of students' answers.

[0013] To achieve the above object, the present invention provides the following solutions:

[0014] A knowledge tracking method based on cognitive fluctuation enhancement, comprising:

[0015] Get students' interactive records of answering questions;

[0016] The question-answering interaction record is input into a preset knowledge tracking model to predict the student's next question-answering situation, wherein the knowledge tracking model is used to extract cognitive features in the question-answering interaction record, decompose the cognitive features and then perform cognitive fluctuation enhancement to predict the next question-answering situation.

[0017] Optionally, the knowledge tracking model includes: an information embedding unit, a feature decomposition unit, an attention processing unit, and a question answering prediction unit, wherein the information embedding unit is used to encode the question answering interaction record into a multi-dimensional feature vector; the feature decomposition unit is used to decompose the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhance them; the attention processing unit is used to extract the enhanced long-term cognitive trend features and short-term cognitive fluctuation features layer by layer based on a multi-layer attention mechanism to obtain cognitive state features; the question answering prediction unit is used to predict the student's next answer based on the cognitive state features.

[0018] Optionally, the method by which the information embedding unit encodes the question-answering interaction record into a multi-dimensional feature vector is:

[0019]

[0020]

[0021] Among them, x t Represents the problem concept encoding, which is the problem q at time step t t and the concepts involved t The encoding, Encode the difficulty of the question, Indicates the problem variation encoding covering knowledge point KC, is the one-hot encoding of the knowledge point KC, ⊙ and Represents element-wise multiplication and element-wise addition operations, y t Represents the student interaction code, which is the student answer status r t After x t Enhanced coding, Indicates c t and r t The joint encoding of Represents the question variation coding covering the knowledge points KC and the student's answer results r.

[0022] Optionally, the feature decomposition unit decomposes the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhances them, including:

[0023] Performing convolution operation on the problem concept coding and the student interaction coding to extract the long-term cognitive trend features;

[0024] Subtracting the long-term cognitive trend feature from the problem concept code and the student interaction code to obtain the short-term cognitive fluctuation feature;

[0025] The long-term cognitive trend characteristics and short-term cognitive fluctuation characteristics are combined to complete cognitive fluctuation enhancement processing.

[0026] Optionally, a method of performing a convolution operation on the problem concept code and the student interaction code to extract the long-term cognitive trend feature is:

[0027] x ′ t =P(x t ,[0,2,1]),

[0028] y t ′ =P(y t ,[0,2,1]),

[0029] Among them, x t and t Represents the problem concept coding and student interaction coding, respectively, ′ t and t ′ They represent the problem concept encoding and student interaction encoding after the permutation operation, P(·) represents the permutation operation, C(·) represents the causal convolution operation, Represents long-term cognitive trend characteristics.

[0030] Optionally, the method of subtracting the long-term cognitive trend feature from the problem concept code and the student interaction code to obtain the short-term cognitive fluctuation feature is:

[0031]

[0032] Among them, S[·] and They represent short-term cognitive fluctuation characteristics and long-term cognitive trend characteristics respectively.

[0033] Optionally, the method of combining the long-term cognitive trend feature and the short-term cognitive fluctuation feature is:

[0034]

[0035] Here, μ and v represent learnable aggregation parameters.

[0036] Optionally, the multi-layer attention mechanism adopts a kernel function based on a conditional positive kernel to model the algebraic differences between different positions in the feature sequence.

[0037] To further achieve the above-mentioned purpose, the present invention also provides a knowledge tracking system based on cognitive fluctuation enhancement, comprising: an information acquisition module and a knowledge tracking module, wherein the information acquisition module is used to obtain students' answering interaction records; the knowledge tracking module is used to input the answering interaction records into a preset knowledge tracking model to predict the student's next answering situation, wherein the knowledge tracking model comprises: an information embedding unit, a feature decomposition unit, an attention processing unit, and an answer prediction unit, wherein the information embedding unit is used to encode the answering interaction records into a multi-dimensional feature vector; the feature decomposition unit is used to decompose the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhance them; the attention processing unit is used to extract the enhanced long-term cognitive trend features and short-term cognitive fluctuation features layer by layer based on a multi-layer attention mechanism to obtain cognitive state features; the answer prediction unit is used to predict the student's next answering situation based on the cognitive state features.

[0038] Optionally, the feature decomposition unit decomposes the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhances them, including:

[0039] Performing convolution operation on the problem concept coding and the student interaction coding to extract the long-term cognitive trend features;

[0040] Subtracting the long-term cognitive trend feature from the problem concept code and the student interaction code to obtain the short-term cognitive fluctuation feature;

[0041] The long-term cognitive trend characteristics and short-term cognitive fluctuation characteristics are combined to complete cognitive fluctuation enhancement processing.

[0042] The beneficial effects of the present invention are:

[0043] (1) Comprehensive modeling capabilities of long-term and short-term cognitive characteristics: The present invention effectively distinguishes and dynamically weights long-term cognitive trends and short-term cognitive fluctuation characteristics through cognitive feature decomposition, thereby solving the problem that traditional knowledge tracking models are unable to capture short-term fluctuations and significantly improving the modeling accuracy of students' cognitive behaviors.

[0044] (2) The model’s length generalization capability is significantly enhanced: The present invention enhances the model’s sensitivity to local short-term information through a kernelized bias attention score penalty mechanism, so that it can still maintain high performance when the length of the test data exceeds the length of the training data. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 The figure is a schematic diagram of the structure and workflow of the knowledge tracking model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] This embodiment provides a knowledge tracking method based on cognitive fluctuation enhancement, including:

[0050] Get students' interactive records of answering questions;

[0051] The question-answering interaction record is input into a preset knowledge tracking model to predict the student's next question-answering situation, wherein the knowledge tracking model is used to extract cognitive features in the question-answering interaction record, decompose the cognitive features and then perform cognitive fluctuation enhancement to predict the next question-answering situation.

[0052] Specifically, this embodiment effectively distinguishes and dynamically weights long-term cognitive trends and short-term cognitive fluctuations through cognitive feature decomposition, solves the problem that traditional knowledge tracking models fail to capture short-term fluctuations, and significantly improves the modeling accuracy of students' cognitive behaviors. This embodiment enhances the model's sensitivity to local short-term information through a kernelized bias attention score penalty mechanism, allowing it to maintain high performance even when the length of the test data exceeds the length of the training data.

[0053] Furthermore, the knowledge tracking model includes: an information embedding unit, a feature decomposition unit, an attention processing unit, and a question answering prediction unit, wherein the information embedding unit is used to encode the question answering interaction record into a multi-dimensional feature vector; the feature decomposition unit is used to decompose the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhance them; the attention processing unit is used to extract the enhanced long-term cognitive trend features and short-term cognitive fluctuation features layer by layer based on a multi-layer attention mechanism to obtain cognitive state features; the question answering prediction unit is used to predict the student's next answer based on the cognitive state features.

[0054] The feature decomposition unit decomposes the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhances them, including:

[0055] Performing convolution operation on the problem concept coding and the student interaction coding to extract the long-term cognitive trend features;

[0056] Subtracting the long-term cognitive trend feature from the problem concept code and the student interaction code to obtain the short-term cognitive fluctuation feature;

[0057] The long-term cognitive trend characteristics and short-term cognitive fluctuation characteristics are combined to complete cognitive fluctuation enhancement processing.

[0058] The multi-layer attention mechanism adopts a kernel function based on a conditional positive kernel to model the algebraic differences between different positions in the feature sequence.

[0059] Combine the following Figure 1 The knowledge tracking model proposed in this embodiment is introduced in detail. The model mainly includes the following parts:

[0060] (1) Information embedding unit: The embedding layer encodes the student's interactive sequence into a multi-dimensional feature vector containing questions, knowledge points, and answer results. The learnable question difficulty and knowledge point variation characteristics are combined to enhance the expression ability of the interactive features. The student interaction sequence is represented as {(q1,c1,r1),...,(q T ,c T ,r T )}. Where T represents the time step, q represents the question number, c represents the concept number involved in the question, and the student's answer r∈{0,1} represents whether the student answered correctly, 0 represents a wrong answer, and 1 represents a correct answer.

[0061] The student interaction coding can be expressed as:

[0062]

[0063] Among them, x tRepresents the problem concept encoding, which is the problem q at time step t t and the concepts involved t The encoding. Code the difficulty of the question. Indicates the variation encoding of questions covering knowledge points (KC). is the one-hot encoding of KC. ⊙ and Represents element-wise multiplication and element-wise addition operations respectively. t Represents the student interaction code, which is to add the student's answer situation r t After x t Enhanced encoding. Indicates c t and r t The joint encoding of Represents the problem variation encoding covering KC and r.

[0064] (2) Feature decomposition unit: With the help of causal convolution and wavelet transform methods, the interactive features are decomposed into long-term cognitive trend features and short-term cognitive fluctuation features, and these features are reconstructed through adaptive weighted aggregation to strengthen the expression of short-term cognitive fluctuation features and avoid the excessive smoothing effect of long-term features on short-term features. First, the problem concept encoding x t Interactive coding with students t Perform convolution operations to filter out the long-term cognitive trend features. Specifically:

[0065] x ′ t =P(x t ,[0,2,1]),

[0066] y t ′ =P(y t ,[0,2,1]),

[0067] Among them, x t and t Represents the problem concept coding and student interaction coding, respectively, ′ t and t ′ They represent the problem concept encoding and student interaction encoding after the permutation operation, P(·) represents the permutation operation, C(·) represents the causal convolution operation, Represents long-term cognitive trend characteristics.

[0068] Then, from the original feature x t ,y t Subtract the long-term cognitive trend characteristics from the short-term cognitive fluctuation characteristics to obtain the short-term cognitive fluctuation characteristics:

[0069]

[0070] Among them, S[·] and They represent short-term cognitive fluctuation characteristics and long-term cognitive trend characteristics respectively.

[0071] Finally, the long-term cognitive trend features and short-term cognitive fluctuation features are adaptively combined as follows:

[0072]

[0073] Where μ and v represent learnable aggregation parameters. Through this explicit decomposition operation, the present embodiment forces the two entangled cognitive features to be separated and reassembles them in a weighted aggregation manner to enhance the short-term cognitive fluctuation features in the original input. In addition, by using learnable weights, the attention mechanism can adjust its focus on the two types of cognitive features according to the downstream tasks.

[0074] (3) Attention processing unit: A multi-layer attention mechanism is used to extract students’ cognitive state characteristics layer by layer, analyze the interaction between long-term trends and short-term fluctuation characteristics, and solve the problem that multi-layer attention calculations lead to single features, ensuring that the model can effectively capture complex cognitive change patterns. In the attention layer, a kernelized bias penalty mechanism is introduced to suppress long-term attention scores and focus on short-term high-frequency information, thereby improving the model’s ability to capture local features and enhancing its ability to extrapolate to unseen data sequences.

[0075] First, we use scaled click attention to compute the input question sequence x t These attention scores are then used as weights with the student’s response encoding y t Multiply and weight them together to get the student's knowledge status.

[0076]

[0077] Where Q = x t W Q , K = x t W K , V = y t W V , W Q , W K , W V is a learnable parameter, is the scaling factor. For the lth attention layer, the output of its interaction sequence is represented as H (l) , the attention operation is summarized as follows:

[0078]

[0079] A kernel function based on conditionally positive definite kernel (CPD) is introduced in the attention score to model the algebraic differences between different positions in the sequence. Specifically, for the element a in the attention score matrix A, ij :

[0080]

[0081] in, is the scaling factor, q i is the i-th query vector, k j is the j-th key vector, n is the total number of key vectors, τ1, τ2 are two learnable parameters that satisfy 0<τ1≤1, 0<τ2≤1.

[0082] (4) Answer prediction unit: Based on the extracted behavioral sequence features, a two-layer fully connected neural network is used to predict the students' answer results, and the binary cross-entropy loss function is used to optimize the model's prediction ability, thereby improving the accuracy of the prediction of students' answers.

[0083] After L layers of attention layers, the model hierarchically extracts the knowledge state information in the sequence of interaction between students and knowledge points. The final generated sequence is represented by h t+1 It represents the features extracted by the model at the t+1 time step. The prediction layer uses a two-layer fully connected neural network for prediction. And by minimizing the true answer r t+1 With predicted probability The binary cross entropy loss function between them is used to optimize the model:

[0084]

[0085] Among them, γ and δ represent the Sigmoid function and Relu function respectively, W1 and W2 are the learnable parameter matrices in the fully connected neural network, b1 and b2 are bias terms, and h t+1 is the final feature sequence extracted, x t+1 Encode the problem to be predicted.

[0086] To further optimize the above technical solution, this embodiment also provides a knowledge tracking system based on cognitive fluctuation enhancement, including: an information acquisition module and a knowledge tracking module, the information acquisition module is used to obtain students' answering interaction records; the knowledge tracking module is used to input the answering interaction records into a preset knowledge tracking model to predict the student's next answering situation, wherein the knowledge tracking model includes: an information embedding unit, a feature decomposition unit, an attention processing unit, and an answer prediction unit, wherein the information embedding unit is used to encode the answering interaction records into a multi-dimensional feature vector; the feature decomposition unit is used to decompose the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhance them; the attention processing unit is used to extract the enhanced long-term cognitive trend features and short-term cognitive fluctuation features layer by layer based on a multi-layer attention mechanism to obtain cognitive state features; the answer prediction unit is used to predict the student's next answering situation based on the cognitive state features.

[0087] The feature decomposition unit decomposes the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhances them, including:

[0088] Performing convolution operation on the problem concept coding and the student interaction coding to extract the long-term cognitive trend features;

[0089] Subtracting the long-term cognitive trend feature from the problem concept code and the student interaction code to obtain the short-term cognitive fluctuation feature;

[0090] The long-term cognitive trend characteristics and short-term cognitive fluctuation characteristics are combined to complete cognitive fluctuation enhancement processing.

[0091] The multi-layer attention mechanism adopts a kernel function based on a conditional positive kernel to model the algebraic differences between different positions in the feature sequence.

[0092] Specifically, the knowledge tracking model in this embodiment mainly includes the following parts:

[0093] (1) Information embedding unit: The embedding layer encodes the student's interactive sequence into a multi-dimensional feature vector containing questions, knowledge points, and answer results. The learnable question difficulty and knowledge point variation characteristics are combined to enhance the expression ability of the interactive features. The student interaction sequence is represented as {(q1,c1,r1),...,(q T ,c T ,r T )}. Where T represents the time step, q represents the question number, c represents the concept number involved in the question, and the student's answer r∈{0,1} represents whether the student answered correctly, 0 represents a wrong answer, and 1 represents a correct answer.

[0094] The student interaction coding can be expressed as:

[0095]

[0096] Among them, x t Represents the problem concept encoding, at time step t, problem q t and the concepts involved t The encoding. Code the difficulty of the question. Indicates the variation encoding of questions covering knowledge points (KC). is the one-hot encoding of KC. ⊙ and Represents element-wise multiplication and element-wise addition operations respectively. t Represents the student interaction code, which is to add the student's answer situation r t After x t Enhanced encoding. Indicates c t and r t The joint encoding of Represents the problem variation encoding covering KC and r.

[0097] (2) Feature decomposition unit: With the help of causal convolution and wavelet transform methods, the interactive features are decomposed into long-term cognitive trend features and short-term cognitive fluctuation features, and these features are reconstructed through adaptive weighted aggregation to strengthen the expression of short-term cognitive fluctuation features and avoid the excessive smoothing effect of long-term features on short-term features. First, the problem concept encoding x t Interactive coding with students t Perform convolution operations to filter out the long-term cognitive trend features. Specifically:

[0098] x ′ t =P(x t ,[0,2,1]),

[0099] y t ′ =P(y t ,[0,2,1]),

[0100] Among them, x t and t Represents the problem concept coding and student interaction coding, respectively, ′ t and t ′ They represent the problem concept encoding and student interaction encoding after the permutation operation, P(·) represents the permutation operation, C(·) represents the causal convolution operation, Represents long-term cognitive trend characteristics.

[0101] Then, from the original feature x t ,y t Subtract the long-term cognitive trend characteristics from the short-term cognitive fluctuation characteristics to obtain the short-term cognitive fluctuation characteristics:

[0102]

[0103] where S[·] and They represent short-term cognitive fluctuation characteristics and long-term cognitive trend characteristics respectively.

[0104] Finally, the long-term cognitive trend features and short-term cognitive fluctuation features are adaptively combined as follows:

[0105]

[0106] Where μ and v represent learnable aggregation parameters. Through this explicit decomposition operation, the present embodiment forces the two entangled cognitive features to be separated and reassembles them in a weighted aggregation manner to enhance the short-term cognitive fluctuation features in the original input. In addition, by using learnable weights, the attention mechanism can adjust its focus on the two types of cognitive features according to the downstream tasks.

[0107] (3) Attention processing unit: A multi-layer attention mechanism is used to extract students’ cognitive state characteristics layer by layer, analyze the interaction between long-term trends and short-term fluctuation characteristics, and solve the problem that multi-layer attention calculations lead to single features, ensuring that the model can effectively capture complex cognitive change patterns. In the attention layer, a kernelized bias penalty mechanism is introduced to suppress long-term attention scores and focus on short-term high-frequency information, thereby improving the model’s ability to capture local features and enhancing its ability to extrapolate to unseen data sequences.

[0108] First, we use scaled click attention to compute the input question sequence x t These attention scores are then used as weights with the student’s response encoding y t Multiply and weight them together to get the student's knowledge status.

[0109]

[0110] Where Q = x t W Q , K = x t W K , V = y t W V , W Q , W K , W V is a learnable parameter, is the scaling factor. For the lth attention layer, the output of its interaction sequence is represented as H (l) , the attention operation is summarized as follows:

[0111]

[0112] A kernel function based on conditionally positive definite kernel (CPD) is introduced in the attention score to model the algebraic differences between different positions in the sequence. Specifically, for the element a in the attention score matrix A, ij :

[0113]

[0114] in, is the scaling factor, q i is the i-th query vector, k j is the j-th key vector, n is the total number of key vectors, τ1, τ2 are two learnable parameters that satisfy 0<τ1≤1, 0<τ2≤1.

[0115] (4) Answer prediction unit: Based on the extracted behavioral sequence features, a two-layer fully connected neural network is used to predict the students' answer results, and the binary cross-entropy loss function is used to optimize the model's prediction ability, thereby improving the accuracy of the prediction of students' answers.

[0116] After L layers of attention layers, the model hierarchically extracts the knowledge state information in the sequence of interaction between students and knowledge points. The final generated sequence is represented by h t+1 It represents the features extracted by the model at the t+1 time step. The prediction layer uses a two-layer fully connected neural network for prediction. And by minimizing the true answer r t+1 With predicted probability The binary cross entropy loss function between them is used to optimize the model:

[0117]

[0118] Among them, γ and δ represent the Sigmoid function and Relu function respectively, W1 and W2 are the learnable parameter matrices in the fully connected neural network, b1 and b2 are bias terms, and h t+1 is the final feature sequence extracted, x t+1 Encode the problem to be predicted.

[0119] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A knowledge tracking method based on cognitive fluctuation enhancement, characterized in that: include: Get students' interactive records of answering questions; The question-answering interaction record is input into a preset knowledge tracking model to predict the student's next question-answering situation, wherein the knowledge tracking model is used to extract cognitive features in the question-answering interaction record, decompose the cognitive features and then perform cognitive fluctuation enhancement to predict the next question-answering situation.

2. The knowledge tracking method based on cognitive fluctuation enhancement according to claim 1 is characterized in that: The knowledge tracking model includes: an information embedding unit, a feature decomposition unit, an attention processing unit, and a question answering prediction unit, wherein the information embedding unit is used to encode the question answering interaction record into a multi-dimensional feature vector; the feature decomposition unit is used to decompose the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhance them; the attention processing unit is used to extract the enhanced long-term cognitive trend features and short-term cognitive fluctuation features layer by layer based on a multi-layer attention mechanism to obtain cognitive state features; the question answering prediction unit is used to predict the student's next answer based on the cognitive state features.

3. The knowledge tracking method based on cognitive fluctuation enhancement according to claim 2 is characterized in that: The method by which the information embedding unit encodes the question-answering interaction record into a multi-dimensional feature vector is: Among them, x t Represents the problem concept encoding, which is the problem q at time step t t and the concepts involved t The encoding, Encode the difficulty of the question, Indicates the problem variation encoding covering knowledge point KC, is the one-hot encoding of the knowledge point KC, ⊙ and Represents element-wise multiplication and element-wise addition operations, y t Represents the student interaction code, which is the student answer status r t After x t Enhanced coding, Indicates c t and r t The joint encoding of Represents the question variation coding covering the knowledge points KC and the student's answer results r.

4. The knowledge tracking method based on cognitive fluctuation enhancement according to claim 2 is characterized in that: The feature decomposition unit decomposes the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhances them, including: Performing convolution operation on the problem concept coding and the student interaction coding to extract the long-term cognitive trend features; Subtracting the long-term cognitive trend feature from the problem concept code and the student interaction code to obtain the short-term cognitive fluctuation feature; The long-term cognitive trend characteristics and short-term cognitive fluctuation characteristics are combined to complete cognitive fluctuation enhancement processing.

5. The knowledge tracking method based on cognitive fluctuation enhancement according to claim 4 is characterized in that: The method of performing convolution operation on the problem concept coding and the student interaction coding to extract the long-term cognitive trend characteristics is as follows: Among them, x t and t Represent the problem concept coding and student interaction coding, respectively, ′ t and t ′ They represent the problem concept encoding and student interaction encoding after the permutation operation, P(·) represents the permutation operation, C(·) represents the causal convolution operation, Represents long-term cognitive trend characteristics.

6. The knowledge tracking method based on cognitive fluctuation enhancement according to claim 5 is characterized in that: The method of subtracting the long-term cognitive trend feature from the problem concept code and the student interaction code to obtain the short-term cognitive fluctuation feature is: Among them, S[·] and They represent short-term cognitive fluctuation characteristics and long-term cognitive trend characteristics respectively.

7. The knowledge tracking method based on cognitive fluctuation enhancement according to claim 6 is characterized in that: The method of combining the long-term cognitive trend characteristics and the short-term cognitive fluctuation characteristics is: Here, μ and v represent learnable aggregation parameters.

8. The knowledge tracking method based on cognitive fluctuation enhancement according to claim 2 is characterized in that: The multi-layer attention mechanism adopts a kernel function based on a conditional positive kernel to model the algebraic differences between different positions in the feature sequence.

9. A knowledge tracking system based on cognitive fluctuation enhancement, characterized in that: include: An information acquisition module and a knowledge tracking module, wherein the information acquisition module is used to obtain students' interactive records of answering questions; The knowledge tracking module is used to input the question-answering interaction record into a preset knowledge tracking model to predict the student's next answering situation, wherein the knowledge tracking model includes: an information embedding unit, a feature decomposition unit, an attention processing unit, and a question-answering prediction unit, wherein the information embedding unit is used to encode the question-answering interaction record into a multi-dimensional feature vector; the feature decomposition unit is used to decompose the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhance them; the attention processing unit is used to extract the enhanced long-term cognitive trend features and short-term cognitive fluctuation features layer by layer based on a multi-layer attention mechanism to obtain cognitive state features; the question-answering prediction unit is used to predict the student's next answering situation based on the cognitive state features.

10. The knowledge tracking system based on cognitive fluctuation enhancement according to claim 9 is characterized in that: The feature decomposition unit decomposes the multi-dimensional feature vector into long-term cognitive trend features and short-term cognitive fluctuation features and enhances them, including: Performing convolution operation on the problem concept coding and the student interaction coding to extract the long-term cognitive trend features; Subtracting the long-term cognitive trend feature from the problem concept code and the student interaction code to obtain the short-term cognitive fluctuation feature; The long-term cognitive trend characteristics and short-term cognitive fluctuation characteristics are combined to complete cognitive fluctuation enhancement processing.

Images