Knowledge tracking method and model based on learning ability

Abstract

The invention provides a knowledge tracking method and model based on learning ability, and the method comprises the following steps: introducing a difficulty vector of a question, coding the question and a question and answer pair, and calculating the learning ability of a student for a knowledge concept corresponding to the question at the current moment; the method comprises the following steps of: modeling knowledge states of students at different time steps by using a learning ability enhanced sequence neural network, fusing the learning ability of the students into an acquisition and updating process of the knowledge states of the students, and then extracting knowledge states related to a to-be-predicted problem at the current moment by using an encoder-decoder structure based on a self-attention mechanism; combining the obtained learning ability of the student at the current moment with the knowledge state, and predicting the answering performance of the student at the next moment. According to the invention, the learning process of students can be described more comprehensively and accurately, and the knowledge tracking performance is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge tracing, and particularly relates to a knowledge tracing method and model based on learning ability. Background Art

[0002] In recent years, the rise of online education platforms represented by massive open online courses (MOOCs) and intelligent tutoring systems (ITS) has promoted a revolution in the way of education. Especially under the influence of the global COVID-19 pandemic, the traditional education model has been challenged, accelerating the digital transformation of education. This has led to a surge in learning-related data, providing opportunities for the development of educational data mining (EDM). Among them, knowledge tracing (KT) technology has become a research hotspot, which can model the real-time learning state based on students' historical learning records, track the changes in knowledge levels, and predict future performance.

[0003] However, existing knowledge tracing models mainly focus on students' knowledge mastery level, while ignoring the important factor of learning ability. In fact, students' answering performance is not only related to their current knowledge state, but also closely related to their learning ability. Under the same knowledge state, different students may show different learning speeds and effects. In addition, learning ability also affects the knowledge acquisition process. Early models assumed that students obtained the same knowledge growth after completing questions, while recent research has begun to improve the knowledge update process. For example, some studies have considered factors such as the current knowledge state, knowledge acquisition sensitivity, learning gain, and knowledge absorption. However, these models still do not fully consider the impact of learning ability on knowledge acquisition.

[0004] In short, in personalized education, accurately evaluating and tracing students' academic progress is the key to improving teaching quality. Existing methods often lack in-depth analysis of the learning ability of trainees, resulting in the inability to provide precise personalized teaching support. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a knowledge tracing method and model based on learning ability aiming at the deficiencies of the prior art, which can more comprehensively and accurately depict students' learning processes and improve the performance of knowledge tracing.

[0006] To solve the above technical problem, the technical solution of the present invention is as follows:

[0007] A knowledge tracing method based on learning ability, comprising the following steps:

[0008] S1, introducing the difficulty vector of the question, encoding the question and the question-answer pair, and calculating the learning ability of the student for the knowledge concept corresponding to the question;

[0009] S2. Use a sequential neural network with enhanced learning ability to model the student's knowledge state at different time steps, integrate the student's learning ability into the process of obtaining and updating the student's knowledge state, and then use an encoder-decoder structure based on the self-attention mechanism to extract the knowledge state related to the problem to be predicted at the current moment.

[0010] S3. Combine the obtained learning ability and knowledge state of the student at the current moment to predict the student's answering performance at the next moment.

[0011] Preferably, the step S1 specifically includes the following operations: 1) Problem encoding: For the embedding of the knowledge concept corresponding to the question, introduce the difficulty vector of the question and perform embedding calculation on the question that the student is doing at the current moment. 2) Problem-answer pair encoding: For the embedding of the knowledge concept-answer pair corresponding to the question, also introduce the difficulty vector of the question and encode the problem-answer pair. 3) Learning ability encoding: Calculate the student's learning ability for the corresponding knowledge concept at the current moment according to the student's historical performance on the knowledge concept corresponding to the question.

[0012] Preferably, in the learning ability encoding, k latest question-and-answer records with the same knowledge concept are selected from the student's question-solving history, and the GRU is used to calculate the student's learning ability for this knowledge concept.

[0013] Preferably, in the learning ability encoding, the learning behavior data of the student is obtained from the online education platform, and the machine learning algorithm is used to analyze the student's data to evaluate their learning ability; the learning behavior data includes learning time, learning frequency, and the speed and quality of completing homework; the learning ability includes comprehension ability, memory ability, and analysis ability.

[0014] Preferably, in the step S2, using a sequential neural network with enhanced learning ability to model the student's knowledge state at different time steps specifically includes the following operations: 1) Obtain the subjective question difficulty: Calculate the student's subjective perception of the question difficulty at the current moment by subtracting the student's knowledge state at the previous moment from the projected question embedding corresponding to the question. 2) Knowledge acquisition: Subtract their subjective question difficulty from the student's learning ability, utilize the difference between the two, and combine the question-and-answer pair embedding to obtain the personalized knowledge of the student at the current moment. 3) Knowledge update: Concatenate the student's knowledge state at the previous moment, the knowledge obtained at the current moment, and the learning ability at the current moment and input them into a fully connected network to adaptively update the knowledge state and obtain the student's knowledge state at the current moment.

[0015] Preferably, in step S2, after modeling the knowledge states of students at different time steps using a sequential neural network with enhanced learning ability, the knowledge states of students are further updated in the encoder; the encoder includes a multi-head attention layer, and the inputs of this multi-head attention layer all come from the output of the sequential neural network.

[0016] Preferably, in step S2, in the decoder, first, the sequential information of the question sequence is modeled by GRU, and then, for the question to be predicted at the current moment, the knowledge state related to the question to be predicted is extracted from the encoder according to the relevance between the encoder and the historical questions.

[0017] Preferably, in step S2, the decoder includes two multi-head attention layers. The inputs of the first multi-head attention layer all come from the output of GRU, and the inputs of the second multi-head attention layer come from the output of the first multi-head attention layer and the knowledge states of students at different time steps obtained by the encoder.

[0018] Preferably, in step S1, according to the calculated learning ability of students for the knowledge concepts corresponding to the question at the current moment, the difficulty and progress of the teaching content are dynamically adjusted, and feedback on the learning effect is provided.

[0019] A knowledge tracing model based on learning ability, which is constructed and generated by the above knowledge tracing method.

[0020] The beneficial effects of the present invention are:

[0021] The present invention proposes a knowledge tracing method and model based on students' learning ability. By analyzing the Q&A records of students on the same knowledge points, their current learning ability on specific knowledge points is calculated; through a sequential neural network with enhanced learning ability, the acquisition and update of knowledge states are calculated by combining students' learning ability; then an encoder-decoder structure based on the self-attention mechanism is used to extract the knowledge state related to the question to be predicted; finally, the knowledge state and learning ability of students are combined to jointly predict students' answering performance. The model proposed by the present invention performs well on three real-world public datasets. Ablation experiments verify the effectiveness of each module, providing valuable insights for knowledge tracing models. Description of the Drawings

[0022] Figure 1 It is a structural diagram of the knowledge tracing model LAKT based on students' learning ability;

[0023] Figure 2 It is an architecture diagram of the knowledge tracing model LAKT based on students' learning ability;

[0024] Figure 3It is a structural diagram of the sequential neural network LASNN. Specific implementation manner

[0025] To facilitate the understanding of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners. Those skilled in the art should understand that the described embodiments are only for helping to understand the present invention and should not be regarded as specific limitations on the present invention.

[0026] Aiming at the problem that the current knowledge tracking model lacks the modeling of the learning ability characteristics of individual students, the present invention proposes a knowledge tracking method and model based on students' learning ability, as Figure 2 shown.

[0027] The knowledge tracking method based on learning ability includes the following steps:

[0028] S1. Introduce the difficulty vector of the question, encode the question and the question-answer pair, and calculate the learning ability of the student for the knowledge concept corresponding to the question;

[0029] S2. Use the learning ability enhanced sequential neural network to model the knowledge state of the student at different time steps, integrate the learning ability of the student into the process of obtaining and updating the student's knowledge state, and then use the encoder-decoder structure based on the self-attention mechanism to extract the knowledge state related to the question to be predicted at the current moment;

[0030] S3. Combine the obtained learning ability and knowledge state of the student at the current moment to predict the student's answering performance at the next moment.

[0031] Generally, students who master knowledge points through fewer question exercises show better learning ability. Based on this understanding, the present invention designs a learning ability enhanced sequential neural network (LASNN), which takes learning ability as a key factor in the process of obtaining and updating the student's knowledge state. The present invention uses the encoder-decoder structure based on the self-attention mechanism to extract the knowledge state related to the question to be predicted, and uses the gated recurrent unit GRU to capture the sequential information in the student's question sequence. By combining these two types of information - the student's knowledge state and learning ability - the student's subsequent performance on the question is predicted.

[0032] Step S1 specifically includes the following operations: 1) Question encoding: For the embedding of the knowledge concept corresponding to the question, introduce the difficulty vector of the question and perform embedding calculation on the question that the student is doing at the current moment; 2) Question-answer pair encoding: For the embedding of the knowledge concept-answer pair corresponding to the question, also introduce the difficulty vector of the question and encode the question-answer pair; 3) Learning ability encoding: Calculate the learning ability of the student for the corresponding knowledge concept at the current moment t according to the student's historical performance on the knowledge concept corresponding to the question.

[0033] In the learning ability encoding, k latest Q&A records with the same knowledge concepts are selected from the student's question-solving history, and the GRU is used to calculate the student's learning ability for this knowledge concept.

[0034] In the learning ability encoding, the learning behavior data of the student is obtained from the online education platform, and machine learning algorithms are used to analyze the student's data to evaluate their learning ability; the learning behavior data includes learning time, learning frequency, and the speed and quality of completing homework; the learning ability includes comprehension ability, memory ability, and analysis ability.

[0035] In step S2, using the learning ability enhanced sequential neural network to model the student's knowledge state at different time steps specifically includes the following operations: 1) Obtain the subjective question difficulty: Calculate the student's subjective perception of the question difficulty at the current moment by subtracting the student's knowledge state at the previous moment from the projected question embedding corresponding to the question; 2) Knowledge acquisition: Subtract their subjective question difficulty from the student's learning ability, utilize the difference between the two, and combine with the Q&A pair embedding to obtain the personalized knowledge of the student at the current moment; 3) Knowledge update: Concatenate the student's knowledge state at the previous moment, the knowledge obtained at the current moment, and the learning ability at the current moment and input them into the fully connected network to adaptively update the knowledge state and obtain the student's knowledge state at the current moment.

[0036] In step S2, after using the learning ability enhanced sequential neural network to model the student's knowledge state at different time steps, the student's knowledge state is further updated in the encoder; the encoder includes a multi-head attention layer, and the input of this multi-head attention layer all comes from the output of the sequential neural network.

[0037] In step S2, in the decoder, first, the sequential information of the question sequence is modeled by the GRU, and then, for the question to be predicted at the current moment, the knowledge state related to the question to be predicted is extracted from the encoder according to the relevance between the encoder and the historical questions.

[0038] In step S2, the decoder includes two multi-head attention layers. The input of the first multi-head attention layer all comes from the output of the GRU, and the input of the second multi-head attention layer comes from the output of the first multi-head attention layer and the knowledge state of the student at different time steps obtained by the encoder.

[0039] In step S1, according to the calculated learning ability of the student for the knowledge concept corresponding to the question at the current moment, the difficulty and progress of the teaching content are dynamically adjusted, and feedback on the learning effect is provided.

[0040] The present invention also provides a knowledge tracing model LAKT based on learning ability, and the knowledge tracing model is constructed and generated by the above-mentioned knowledge tracing method. Figure 1 As shown in the figure, the knowledge tracing model LAKT based on students' learning ability includes an encoding layer, a knowledge state layer, and a prediction layer.

[0041] The task of knowledge tracing (KT) is to predict a student's performance on future questions based on the student's historical response data. This involves constructing a model to process and analyze sequential data and extract latent information to predict the next response. Specifically, a student's answer record can be represented as:

[0042] X = ((q 1 , r 1 ), (q 2 , r 2 ), (q 3 , r 3 ), …, (q T , r T ))

[0043] Among them, q i represents a question, and r i is a binary variable (0 or 1), indicating whether the student correctly answered the question q i (0 indicates incorrect, 1 indicates correct).

[0044] For χ, knowledge tracing aims to predict whether the student will correctly answer q T+1 , that is, calculate the probability:

[0045] P(r T+1 = 1|q T+1 , χ)

[0046] Encoding layer: Encode questions and question-answer pairs, and screen out the recent k exercise records with the same knowledge points from the history of doing questions to calculate the student's learning ability on this knowledge point.

[0047] Question encoding: Due to data sparsity, some existing models have used knowledge concepts to replace questions as the input of the model. However, the existing models ignore the differences between questions covering the same knowledge concept, such as different difficulty levels between questions. And difficulty is an important feature of questions. To better link questions with knowledge concepts, the present invention introduces the difficulty vector of questions, and calculates the embedding of the question q t done by the student at the current time t as:

[0048] q t = c t + μ t ⊙ d t

[0049] wherein, ⊙ represents the Hadamard product, and c t is the embedding of the knowledge concept corresponding to the question q t , μ t is the difficulty vector of the question q t , and d t is the change vector of the knowledge concept corresponding to the question q t .

[0050] Question - answer pair encoding: Similar to the question encoding, the embedding of the question - answer pair at time t is defined as:

[0051] a t = e t + μ t ⊙ f t

[0052] wherein, e t is the embedding of the knowledge concept - answer pair (c t , r t ) corresponding to the question q t , and f t is the change vector of the knowledge concept - answer pair (c t , r t ) corresponding to the question q t .

[0053] Student ability encoding: Considering that students have different learning abilities for different knowledge concepts, when predicting the performance of students on the question q t at the current time t, the present invention calculates their learning ability for the corresponding knowledge concept according to the historical performance of the students on the knowledge concept c t corresponding to this question q t .

[0054] To avoid interference from practice records that are too far apart, the present invention only selects k latest question - answer records with the same knowledge concept c t , denoted as wherein, Intuitively, if students can solve increasingly difficult problems with fewer mistakes during the learning process, then they may have stronger learning abilities.

[0055] To capture the learning ability of students, we input into the GRU:

[0056]

[0057] wherein, represents the concatenation operation, μ tt' is the difficulty vector of the question, and r tt'is a vector of all 1s or 0s, with the same dimension as μ tt' Same.

[0058] Finally, we obtain k output vectors [p 1 , p 2 ,..., p k , and use p k as the learning ability m t of the student at time t:

[0059] m t = tanh(p k )

[0060] where tanh is the tanh activation function.

[0061] The knowledge state layer includes the Learning Ability-aware Sequential Neural Network (LASNN) and an encoder-decoder. The encoder-decoder uses the self-attention mechanism to calculate the correlation between the question and the question-answer pair.

[0062] First, the learning ability is incorporated into the process of students' knowledge acquisition and knowledge update through the LASNN; then, the encoder is used to further model the knowledge state of the student at different time steps; finally, in order to predict the question in the decoder, the knowledge state related to the question is extracted from the encoder.

[0063] The LASNN integrates the learning ability of the student into the process of knowledge state change. Its structure is as Figure 2 shown, consisting of three parts: (a) subjective question difficulty, (b) knowledge acquisition, and (c) knowledge update.

[0064] (a) Subjective question difficulty

[0065] Previously, an objective difficulty was defined for each question. However, each student's perception of the question difficulty is different. Specifically, if the student's knowledge mastery level meets the requirements of the question, they will find the question relatively easy; for students who do not meet the requirements of the question, they will correspondingly find the question more difficult. The present invention calculates the subjective question difficulty perception at the current time t by subtracting the student's previous knowledge state h t from the projected question embedding x t corresponding to the question q. The difference between the two represents the student's subjective perception of the question difficulty: t-1

[0066]

[0067] where σ is the Sigmoid activation function, ⊙ represents the Hadamard product (element-wise multiplication), and W 1 , W 2 , W 3 are weight matrices, and b 1 , b 2 , b 3 are bias vectors. is the direct difference between x t and h t-1 . The gating vector is used to retain or remove the information in , and sd t is the final student subjective question difficulty vector obtained by the present invention.

[0068] (b) Knowledge acquisition

[0069] For each question-answer pair (q t , r t ), the actual amount of knowledge obtained by each student is different. Comparing the learning ability m t of the student with the subjective question difficulty sd t , if the subjective question difficulty sd t is within the ability range of the student, the student can understand the question content and learn more from this problem-solving experience. However, if the subjective question difficulty sd t exceeds the current ability of the student, the student will obtain corresponding benefits after completing the question.

[0070] To solve this problem, the present invention subtracts their subjective question difficulty sd t from the learning ability m t of the student, and uses the difference between the two, combined with the question-answer pair embedding a t , to define the personalized knowledge acquisition of the student:

[0071]

[0072] h t ’ = ka t ⊙ a t

[0073] where W 4 , W 5 are weight matrices, b 4 , b 5 are bias vectors, is the direct difference between m t and sd t . The gating vector is used to retain or remove the information in , and h t' represents the knowledge that the student can acquire after this exercise, i.e., at the current time t.

[0074] (c) Knowledge update

[0075] After calculating the personalized knowledge acquisition, it is necessary to calculate the student's knowledge state at the current time t. In the present invention, through a gating mechanism, the knowledge state h of the student at the previous time step t-1 , the knowledge h' obtained at the current time t t and the learning ability m at the current time t t are concatenated and input into a fully connected network to adaptively update the knowledge state:

[0076]

[0077] h t = Γ t ⊙ h’ t +(1 - Γ t ) ⊙ h t-1

[0078] In the formula, represents the vector concatenation operation, W 6 is the weight matrix, b 6 is the bias vector, and h t is the knowledge state of the student at the current time t.

[0079] Encoder: As Figure 3 shown, the correlation calculation between the question and the question - answer pair is completed through an encoder - decoder structure, where both the encoder and the decoder use the multi - head attention mechanism. The formula for multi - head attention is:

[0080] MHA(Q, K, V) = [head 1 ,..., head h W O

[0081]

[0082] In the formula, W O , is the weight matrix, and d k is the dimension of the key - value vector.

[0083] The Learning Ability Enhanced Sequential Neural Network (LASNN) is used to model the knowledge state of the student at different time steps. Next, in the encoder, the knowledge state of the student is further updated:

[0084] M = LN(Q + MHA(Q, K, V))

[0085] H E= LN(M + FFN(M)

[0086] where LN is the normalization operation, FFN is the feed-forward neural network, MHA is the multi-head attention layer, and Q, K, and V all come from the output of LASNN. H E represents the knowledge state of the student at different time steps after passing through the multi-head attention layer.

[0087] Decoder: In the decoder, for the question q to be predicted at the current time t t , relevant knowledge states are extracted from the encoder according to the relevance between the encoder and the historical questions [q 1 , q 2 ,..., q t-1 . Before extracting the relevant knowledge states from the encoder, the sequential information of the question sequence is first modeled by GRU:

[0088] q' t = GRU(q t , q' t-1 )

[0089]

[0090] The decoder includes two multi-head attention layers, and the process is as follows:

[0091] M 1 = LN(Q + MHA(Q, K, V))

[0092] M 2 = LN(M 1 + MHA(M 1 , M 1 , H E ))

[0093] H D = LN(M 2 + FFN(M 2 ))

[0094] where Q, K, and V all come from Finally, H D is the extracted knowledge state of the student related to the question to be predicted.

[0095] (3) Prediction layer:

[0096] Combine the knowledge state and learning ability of the student to predict the performance of the student when answering questions at the next moment.

[0097] Through the above steps, we obtain the learning ability m of the student at time t t and the extracted knowledge state related to the question to be predicted To predict a student's performance on questions, we combine the learning ability m t and the knowledge state as follows:

[0098]

[0099] where W out is the weight matrix, b out is the bias vector, is the predicted probability that the student answers the question correctly. Finally, we use the cross-entropy loss function to train the model:

[0100]

[0101] Experimental analysis:

[0102] Experimental data: The present invention was experimented on three public datasets: ASSIST2017, Junyi, and EdNet. Their statistical information is shown in Table 1.

[0103] Table 1 Statistical information of public datasets

[0104]

[0105] The ASSIST2017 dataset comes from the 2017 ASSISTments data mining competition and records the learning data of students on the ASSISTments platform from 2004 to 2007; the Junyi dataset comes from Junyi Academy and records the records of students solving problems since 2015; for the Educational Network (EdNet) dataset, due to its large scale, 5,000 students were randomly selected from EdNet-KT1 for model testing.

[0106] Evaluation metrics: In this experiment, the area under the ROC curve (AUC) is used to evaluate the prediction performance of the model. The ROC curve or Receiver Operating Characteristic curve has the true positive rate on the y-axis and the false positive rate on the x-axis. The area value is a specific evaluation metric, and the larger the area value, the better the model performance. In addition, AUC is applicable to scenarios with imbalanced positive and negative samples, so the present invention uses AUC to measure the model performance.

[0107] Experimental parameters: This experiment was implemented using the PyTorch deep learning framework and run on an RTX3090 GPU. For model training, 5-fold cross-validation was used to find the optimal parameters. The embedding dimensions of questions, concepts, question-answer pairs, and learning ability were all set to 256, and the attention layer had 8 attention heads. All learnable parameters were optimized using Adam, with an initial learning rate of 0.001, which decreased as the number of iterations increased.

[0108] Comparative experiments: To evaluate the performance of the model proposed in this invention, the LAKT model was compared with the following baseline models: DKT is the first deep learning-based knowledge tracing model, which represents the knowledge state of students through the hidden vector in the RNN; DKT-Q uses questions instead of concepts as the input of the DKT model; DKVMN introduced a dynamic key-value memory network, which uses static and dynamic matrices to store the static embeddings of concepts and the mastery level of each concept by students respectively; DKVMN-Q uses questions instead of concepts as the input of the DKVMN model; SAKT uses a self-attention mechanism to identify the interactions related to a given concept from the past problem-solving interactions of students and predicts their knowledge mastery based on their performance in the relevant interactions; AKT uses a new monotonic attention mechanism to link the performance of students on the problem to be predicted with their historical problem-solving performance.

[0109] The experimental results are shown in Table 2. The model of this invention outperforms other baseline models on all three datasets, indicating that the LAKT model can effectively improve the prediction of students' answer performance. Specifically, the AUC values of the LAKT model on the ASSIST17, Junyi, and EdNet datasets increased by 1.55%, 0.98%, and 1.43% respectively.

[0110] Table 2. Comparative experimental results of LAKT student answer prediction

[0111]

[0112] It can be seen from the experimental results that the sequence-based DKT and DKVMN models have similar overall results, while the self-attention-based SAKT and AKT models show a greater performance gap. This is because SAKT only applies a self-attention network, which may not have strong feature extraction capabilities. The AKT model adopts an encoder-decoder structure, effectively improving the accuracy of the prediction results, indicating that this structure can extract relevant useful information more effectively. The model of this invention also uses an encoder-decoder structure, and its performance is better than that of AKT.

[0113] Ablation study: To prove the effectiveness of each module in the model, the present invention designed the following ablation experiments: LAKT-RLA removed the learning ability of the students in the prediction module and only used the knowledge state and question information to predict the students' responses; LAKT-RKS deleted the knowledge state of the students in the prediction module and only used the learning ability and question information to predict the students' answers; LAKT-RSN removed the LASNN module and directly embedded the question-answer pairs into the encoder; LAKT-RLG replaced the LASNN module with GRU and directly embedded the question-answer pairs into GRU; LAKT-RGRU removed GRU before the decoder, that is, did not integrate the order information of the student question sequence into the question embedding. The results of the ablation experiments are shown in Table 3.

[0114] Table 3 Results of the LAKT Ablation Experiment

[0115]

[0116] It can be observed from Table 3 that: 1) The AUC values of LAKT-RLA, which removed the learning ability from the prediction module, decreased by 5.73%, 1.16%, and 0.31% respectively on the three datasets, indicating that incorporating the learning ability of students in the model of the present invention can effectively improve the model performance. At the same time, the AUC values of LAKT-RKS, which removed the knowledge state from the prediction module, decreased by 0.23%, 2.23%, and 2.45% respectively on the 3 datasets, indicating the necessity of using the knowledge state of students in the prediction module. 2) The AUC values of LAKT-RSN, which removed the LASNN module, decreased by 0.57%, 1.77%, and 0.59% respectively on the 3 datasets, proving the effectiveness of the LASNN module designed by the present invention. 3) The AUC values of LAKT-RLG, which replaced LASNN with GRU, decreased by 0.49%, 1.11%, and 0.29% respectively on the 3 datasets, indicating that the LASNN module designed by the present invention is superior to directly using GRU. 4) The AUC values of LAKT-RGRU, which removed GRU before the decoder, decreased by 0.47%, 0.72%, and 0.12% respectively on the three datasets, indicating that using GRU before inputting the data into the self-attention layer can capture the order information of the student question sequence and improve the model performance.

[0117] Analysis of the hyperparameter k: In the student ability modeling part of the LAKT model, k latest question-answer records with the same knowledge points were selected for each student. Here, k is a key hyperparameter, and corresponding experiments were designed to explore the appropriate value of k. Specifically, k was set to take values from {4, 6, 8, 10, 12, 14}, and the performance of the model on the three datasets ASSIST17, Junyi, and EdNet was observed. The AUC results for different values of k are shown in Table 4.

[0118] The results show that the AUC of the model is basically close in the 6 experiments, and the optimal k values for ASSIST17, Junyi, and EdNet are 10, 8, and 8 respectively. The experimental results show that when the k value is relatively small, the model may not be able to learn useful features; while when the k value is relatively large, it will introduce too much noise because learning records too long ago may not have a significant impact on the present.

[0119] Table 4 AUC results under different k values

[0120]

[0121] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A knowledge tracking method based on learning ability, characterized in that: The following steps are involved: S1, introduces the difficulty vector of the question, encodes the question and the question-answer pair, and calculates the student's learning ability of the knowledge concept corresponding to the question; S2, using learning ability to enhance the sequential neural network to model the students' knowledge status at different time steps, integrating the students' learning ability into the process of acquiring and updating the students' knowledge status, and then using the encoder-decoder structure based on the self-attention mechanism to extract the knowledge status related to the problem to be predicted at the current moment; S3, combines the student’s learning ability and knowledge status at the current moment to predict the student’s answering performance at the next moment.

2. The knowledge tracking method based on learning ability according to claim 1 is characterized in that: The step S1 specifically includes the following operations: 1) Question encoding: for the embedding of the knowledge concept corresponding to the question, the difficulty vector of the question is introduced, and the embedding calculation is performed on the question currently being done by the student; 2) Question-answer pair encoding: for the embedding of the knowledge concept-answer pair corresponding to the question, the difficulty vector of the question is also introduced, and the question-answer pair is encoded; 3) Learning ability encoding: based on the student's historical performance on the knowledge concept corresponding to the question, the student's learning ability for the corresponding knowledge concept at the current moment is calculated.

3. The knowledge tracking method based on learning ability according to claim 2 is characterized in that: In the learning ability coding, k latest question and answer records with the same knowledge concept are selected from the student's question-solving history, and the GRU is used to calculate the student's learning ability for the knowledge point concept.

4. The knowledge tracking method based on learning ability according to claim 2 is characterized in that: In the learning ability coding, the learning behavior data of students is obtained from the online education platform, and the data of students is analyzed using a machine learning algorithm to evaluate their learning ability; The learning behavior data includes learning time, learning frequency, and the speed and quality of completing homework; the learning ability includes comprehension, memory, and analysis.

5. The knowledge tracking method based on learning ability according to claim 1 is characterized in that: In the step S2, the learning ability enhanced sequential neural network is used to model the students' knowledge states at different time steps, which specifically includes the following operations: 1) Obtaining subjective problem difficulty: calculating the students' subjective perception of problem difficulty at the current moment by subtracting the students' knowledge state at the previous moment from the projected problem embedding corresponding to the question; 2) Knowledge acquisition: subtracting their subjective problem difficulty from their learning ability, using the difference between the two, and combining with the question-answer pair embedding to obtain the students' personalized knowledge at the current moment; 3) Knowledge update: splicing the students' knowledge state at the previous moment, the knowledge acquired at the current moment, and the learning ability at the current moment together and inputting them into the fully connected network to adaptively update the knowledge state and obtain the students' knowledge state at the current moment.

6. The knowledge tracking method based on learning ability according to claim 1, characterized in that: In the step S2, after the student's knowledge state at different time steps is modeled using the learning ability enhanced sequential neural network, the student's knowledge state is further updated in the encoder; the encoder includes a multi-head attention layer, and the input of the multi-head attention layer all comes from the output of the sequential neural network.

7. The knowledge tracking method based on learning ability according to claim 1 is characterized in that: In step S2, in the decoder, the sequential information of the question sequence is first modeled by GRU, and then, for the question to be predicted at the current moment, the knowledge state related to the question to be predicted is extracted from the encoder according to the correlation between the encoder and the historical questions.

8. The knowledge tracking method based on learning ability according to claim 7 is characterized in that: In step S2, the decoder includes two multi-head attention layers, the input of the first multi-head attention layer comes from the output of GRU, and the input of the second multi-head attention layer comes from the output of the first multi-head attention layer and the knowledge state of the student obtained by the encoder at different time steps.

9. The knowledge tracking method based on learning ability according to claim 1, characterized in that: In step S1, the difficulty and progress of the teaching content are dynamically adjusted based on the calculated student's learning ability of the knowledge concept corresponding to the question at the current moment, and feedback on the learning effect is provided.

10. A knowledge tracking model based on learning ability, characterized in that: The knowledge tracking model is constructed and generated by the knowledge tracking method as described in any one of claims 1-9.

Images