Knowledge tracking method and system based on multi-view dynamic evolution

Abstract

The invention discloses a knowledge tracking method and system based on multi-view dynamic evolution, which are used for modeling knowledge of students from three perspectives of problems, concepts and entirety and comprehensively capturing the evolution of knowledge states of the students. According to the invention, by using and combining the LSTM network, the self-attention mechanism and the IRT, complex interaction between different knowledge dimensions can be captured, so that the learning process of students can be known more comprehensively. LSTM is used to process an interaction sequence and capture changes of knowledge states, an IRT model is used to predict performance of students on different knowledge points, learner ability and problem features are quantified, and model interpretability is enhanced. In addition, the knowledge state is dynamically updated through a self-attention mechanism, so that the model can capture long-term dependency and relation between knowledge points and learning situations. According to the invention, the learning condition of the student can be scientifically and comprehensively predicted, and the purpose of assisting the teacher in precise teaching is achieved.

Description

Technical Field

[0001] The present invention relates to the fields of educational big data mining, neural networks and student behavior modeling, and in particular to a knowledge tracking method and system based on multi-perspective dynamic evolution. Background Art

[0002] With the development of information technology, knowledge tracking (KT) has attracted more and more attention in the field of education. As a core technology for monitoring and predicting learners' knowledge status, KT tracks and predicts students' knowledge status in real time by analyzing learning behavior data. This technology has important value in personalized teaching and adaptive learning systems. It can help educators understand students' learning progress and weak links, and carry out targeted teaching interventions.

[0003] The current KT model mainly solves two major problems: one is to track and monitor students' knowledge status, and the other is to predict students' learning performance.

[0004] With the development of information technology, knowledge tracking (KT) has received more and more attention in the field of education. As the core technology for monitoring and predicting learners' knowledge status, KT tracks and predicts students' knowledge status in real time by analyzing learning behavior data. This technology is of great value in personalized teaching and adaptive learning systems. It can help educators understand students' learning progress and weak links, and conduct targeted teaching interventions. The current KT model mainly solves two major problems: one is to track and monitor students' knowledge status, and the other is to predict students' learning performance.

[0005] Most traditional KT methods use RNN or attention mechanism to track knowledge state. Bayesian Knowledge Tracing (BKT) adopts Hidden Markov Model (HMM) to update binary knowledge state (mastered / not mastered) according to answer history. Deep Knowledge Tracing (DKT) uses RNN or LSTM to represent knowledge state with hidden layers. Dynamic Key-Value Memory Network (DKVMN) uses LSTM-based dynamic storage mechanism for key-value pair storage and update. Self-Attention Knowledge Tracing (SAKT) applies the self-attention mechanism of Transformer, while Context-Aware Attention Knowledge Tracing (AKT) combines context information with attention mechanism. Graph-based Knowledge Tracing (GKT) adopts Graph Neural Network (GNN) to describe the relationship between students and knowledge points.

[0006] In terms of predicting student learning performance, Knowledge Tracking Machine (KTM), which integrates matrix decomposition and factorization machine (FM) to predict student grades; Convolutional Knowledge Tracking (CKT), which uses CNN to capture knowledge state evolution; DeepIRT, which combines deep learning with item response theory (IRT) to model ability and question characteristics to improve accuracy. The simple and powerful model REKT consists of two linear regression units (Forget-Response-Update, FRU) and has excellent prediction performance. Fine-grained knowledge tracking (FKT) combines response time, multi-task learning and temporal distance attention mechanism, and uses an encoder-decoder predictor model to improve prediction accuracy.

[0007] Existing KT models focus on modeling students' knowledge status based on concepts or problems, ignoring the characteristics of problems, concepts, and the overall impact on students' hidden knowledge status. Faced with complex knowledge structures, it is a challenge to effectively capture and dynamically update students' knowledge mastery while improving prediction accuracy. Summary of the invention

[0008] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art. This study models students' knowledge from three perspectives: problem, concept, and overall, and comprehensively captures the evolution of their knowledge state. We use LSTM to process interaction sequences and capture changes in knowledge state. The IRT model is used to predict students' performance on different knowledge points, quantify learner abilities and problem characteristics, and enhance model interpretability. In addition, the knowledge state is dynamically updated through the self-attention mechanism, so that the model can capture the long-term dependencies and relationships between knowledge points and learning contexts.

[0009] The purpose of the present invention is to achieve the following technical measures: a knowledge tracking method based on multi-perspective dynamic evolution, comprising the following steps:

[0010] Step 1: Obtain the input sequence based on the question feature, the input sequence based on the concept feature, and the input sequence based on the global feature at the current time t+1, and compare the three obtained sequences with the interaction feature E at the current time. t+1 splicing, and input them into the LSTM model respectively to obtain the corresponding output features, and splicing the three output features as the representation of the final knowledge state at the current moment H t+1 ;

[0011] Step 2: Use IRT to calculate the probability L of the student answering the question correctly, combined with the final knowledge state q ,Combining the students’ answers with the inconsistent behaviors of concept mastery, we can obtain the predicted values ​​of students’ learning behaviors;

[0012] In step 2, the behavior of students’ answers being inconsistent with their mastery of the concepts includes two types: guessing the correct answer without truly mastering it and making mistakes despite knowing the answer. First, the knowledge state H t+1 With the interaction feature E t+1 Splice them together and pass them through MLP to get two feature sequences G and S representing the behavior, G = MLP G (H t+1 ∪E t+1 ), S = MLP S (H t+1 ∪E t+1 ), where MLP G , MLP S There are two independent MLP networks with different weight matrices and bias vectors. σ is the activation function, and the following response formula y is used. t+1 =L q *(1-S)+(1-L q )*G, we can get the predicted value of the final student learning behavior.

[0013] Furthermore, in step 1, we first set the problem feature matrix Q∈R n×d , concept feature matrix C∈R m×d , the answer feature matrix R∈R 2×d , where n is the number of questions, m is the number of concepts, and d is the feature dimension, then the interactive feature is represented as Among them, Qq t+1 and Cc t+1 are the problem q at time t+1 in Q and C respectively t+1 and the concept c at time t+1 t+1 The line where diffq t+1 Yes t+1 The difficulty captures the degree to which the problem varies relative to its concept, diffq t+1 and are all trainable parameters.

[0014] Furthermore, in step 1, for the input sequence based on the problem feature, the state sequence containing the current problem q is extracted from the problem state sequence t+1 The interactive units are concatenated to obtain the input sequence based on the problem features;

[0015] For the input sequence based on concept features, the state sequence containing the current concept c is extracted from the concept state sequence. t+1 The interactive units are concatenated to obtain the input sequence based on the concept features;

[0016] For the input sequence based on global features, the global state sequence is directly used as the input sequence based on global features;

[0017] For any interaction unit X before the current time t+1 t =(q t ,c t ,r t ), the feature is calculated as X t =E t +Rr t , where Rr t is the rth t OK, q t ,c t ,r t They represent the question, concept and answer at time t respectively.

[0018] Furthermore, in order to enable the LSTM model to track the learner's knowledge mastery in real time, a residual connection is made between the output features of the LSTM model and the knowledge state obtained by the dynamic update module at the previous moment to obtain the knowledge state based on the problem features at the current moment. Knowledge state based on concept characteristics and the knowledge state based on global features Finally, these three knowledge states are concatenated to obtain the representation of the final knowledge state, namely

[0019] The dynamic update module is implemented through the self-attention mechanism.

[0020] Furthermore, the specific processing process of using the dynamic update module to obtain the knowledge state at the current moment is as follows;

[0021] For the problem state, K in the attention mechanism q and V q Set to knowledge state based on problem characteristics The query Q q Set to the interaction unit E at time t+1 t+1 , put them into the self-attention mechanism to get S q =MultiHead(Q q ,K q ,V q ); then a feedforward neural network is used to incorporate the results of the self-attention mechanism into the model, and the interaction between different underlying dimensions is considered, using the following response formula, FFN(S q )=ReLU(S q W (1) +b (1) )W (2) +b (2) ,in These are all learnable parameters during training. Finally, a residual connection is made to obtain an updated representation of the problem state sequence.

[0022] For the conceptual state, the key K c Sum value V c Set as knowledge state based on concept features The query Q c Set to the interaction unit E at time t+1 t+1 , put them into the self-attention mechanism to get S c =MultiHead(Q c ,K c ,V c ); then use the feedforward neural network to process, and finally get the updated concept state sequence representation through the residual connection

[0023] For the global state, the key K o Sum value V o Set to global feature-based knowledge state The query Q o Set to the interaction unit E at time t+1 t+1 , put them into the self-attention mechanism to get S o =MultiHead(Q o ,K o ,V o ); then use the feedforward neural network to process, and finally, get the updated global state sequence representation through the residual connection in and They represent the problem state sequence, concept state sequence and global state sequence at the current moment respectively.

[0024] Furthermore, the probability of a student answering a question correctly is calculated as follows:

[0025] First, calculate the student's mastery level of concept i at time t+1 The final knowledge state H t+1 With the interaction feature E t+1 Put them together and pass them through MLP to get the level of conceptual mastery;

[0026] Then, the problems to be processed at time t+1 are processed through MLP to calculate the discriminant parameters α of specific problems under concept i. i , the difficulty of a specific problem under concept i is β i Finally, the IRT model is applied, using the following response formula Where α = σ[MLP α (qt+1 )],β=σ[MLP β (q t+1 )]; MLP α and MLP β There are two independent MLP networks with different weight matrices and bias vectors. σ is the activation function, and N c Represents the total number of concepts, and the probability L that students correctly answer questions under concept i is obtained qi .

[0027] Furthermore, in step 2, y t+1 Calculate the cross loss entropy with the actual answer result, and the loss function is recorded as Where T represents the total number of time steps, r t+1 Represents the actual response result of the student at time step t+1.

[0028] The present invention also provides a knowledge tracking system based on multi-perspective dynamic evolution, comprising:

[0029] one or more processors;

[0030] A storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the knowledge tracking method based on multi-perspective dynamic evolution as described in the above scheme.

[0031] The present invention is based on a knowledge tracking method based on dynamic evolution from multiple perspectives. It uses a model that effectively integrates problem-based, concept-based, and holistic knowledge states to comprehensively capture the evolution of their knowledge states to improve the accuracy of student performance prediction. By utilizing and combining LSTM networks, self-attention mechanisms, and IRT, it is possible to capture the complex interactions between different knowledge dimensions, thereby gaining a more comprehensive understanding of the student's learning process. The present invention can scientifically and comprehensively predict the student's learning situation, thereby assisting teachers in conducting precise teaching. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The figure is a flowchart of an embodiment of the present invention.

[0033] Figure 2 This is a flow chart of feature sequence extraction in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0035] like Figure 1As shown, a knowledge tracking method based on multi-perspective dynamic evolution includes the following steps:

[0036] Step 1: Obtain the problem-based feature sequence, concept-based feature sequence, and global feature sequence at the current time t+1, concatenate the three sequences with the interaction features at the current time, and input them into the LSTM model to obtain the corresponding output features. Concatenate the three output features as the representation of the final knowledge state at the current time.

[0037] Step 2, using IRT, combined with the final knowledge status, calculate the probability of students answering questions correctly, and combine the students' answers with the inconsistent behavior of concept mastery to obtain the predicted value of students' learning behavior;

[0038] Step 3: Incorporate the feature information of the current interaction unit into the interaction history information, and use the self-attention mechanism to dynamically update the three states at the current time t+1. The three knowledge states include question state, concept state, and global state, which are used for the next batch of sequences to obtain knowledge state representation.

[0039] (1) Representation of features and knowledge states

[0040] First, it is crucial to accurately represent each interaction unit and connect the relationship between questions, concepts, and answers. We use the following feature matrix: Question Q∈R n×d , concept C∈R m×d , the answer is R∈R 2×d , where n is the number of questions, m is the number of concepts, and d is the feature dimension. The feature representation at time t+1 is Here, Qq t+1 and Cc t+1 are the problem q at time t+1 in Q and C respectively t+1 and the concept c at time t+1 t+1 The line where diffq t+1 Yes t+1 The difficulty captures the degree to which the problem varies relative to its concept, diffq t+1 and are all trainable parameters. For any interaction unit X before the current moment t =(q t ,c t ,r t ), q t ,c t ,r t They represent the question, concept and answer at time t respectively, and the feature is calculated as X t =E t +Rr t , where Rrt is the answer r at time t in R t The row where you are located.

[0041] To better model students' knowledge status, we use three types of feature sequences: input sequences based on problem features, input sequences based on concept features, and input sequences based on global features. We process input sequences based on problem features to track learners' knowledge of specific problems to show their mastery of specific problems, process input sequences based on concept features to track learners' knowledge of specific concepts to reflect their mastery of problems involving these concepts, and process input sequences based on global features to track the overall knowledge in the entire interaction history to show the overall performance.

[0042] (1-1) Extraction of feature sequences

[0043] The process of feature sequence extraction is as follows Figure 2 For example, Figure 2 In the problem feature input sequence At time t+1, the problem number of the interaction unit we need to predict is 1, so we select all the interaction units containing the problem number 1 in the problem state sequence, as shown in the figure X1, X α , these interaction units are concatenated using the union ∪ method as the input sequence based on the problem features but The same, input sequence based on concept features Extraction, the interactive unit concept number at time t+1 is 2, we select all interactive units containing question number 2 in the concept state sequence, such as X2, X3, X t , then the input sequence based on the concept feature Finally, we select the entire global state sequence as the input sequence based on global features but

[0044] (1-2) Modeling of knowledge status

[0045] We use LSTM to process interaction sequences and capture knowledge states, allowing the model to track learners’ knowledge mastery in real time. Figure 1 In part (a), for the knowledge state representation based on question features, we first concatenate the input sequence based on question features and the interaction features at time t+1 as the input of the LSTM model. Then, the output of the LSTM model The problem state obtained by the dynamic update module at the previous moment (i.e., calculated using the same method as in step 3) By making a residual connection, we can get the knowledge state based on the problem characteristics at time t+1 but Knowledge state based on concept characteristics and knowledge state representation based on global features The same calculation method is used. in, and The conceptual state and global state obtained by the dynamic update module at the last moment.

[0046] Finally, we concatenate the three knowledge states to get the representation of the final knowledge state, namely

[0047]

[0048] (2) Gradual learning process

[0049] In order to combine the problem characteristics with the concept mastery level in the model to obtain students' ability to solve specific problems and predict students' performance on different knowledge points, we use the IRT model, such as Figure 1 Part (b) can quantify learner ability and problem characteristics and enhance model interpretability. We also consider the behavior of students’ answers that are inconsistent with their concept mastery and apply it to the probability calculation formula.

[0050] (2-1) Calculate students’ level of mastery of the problem

[0051] First, calculate the student's mastery level of concept i at time t+1 We will know the state H t+1 With the interaction feature E t+1 Put them together and pass them through MLP to get the level of conceptual mastery.

[0052] Then we use MLP to calculate the discriminant parameters α of specific problems under concept i for the problems that need to be processed at time t+1. i , the difficulty of a specific problem under concept i is β i , α=σ[MLP α (q t+1 )],β=σ[MLP β (q t+1 )], where MLP α , MLP β are two independent MLP networks with different weight matrices and bias vectors, and σ is the activation function. Finally, we apply the IRT model and use the following response formula in N cRepresents the total number of concepts, and the probability L that students correctly answer questions under concept i is obtained qi .

[0053] (2-2) Eliminate random guessing and predict students’ learning behavior

[0054] Based on the analysis of question-answering behavior in the classic BKT, we also consider behaviors in which students’ answers are inconsistent with their conceptual mastery. Behaviors in which students’ answers are inconsistent with their conceptual mastery are divided into two types: guessing the correct answer without truly mastering it (P(G)) and making mistakes despite knowing the answer (P(S). The sum of these probabilities gives the overall probability of the student’s correct answer.

[0055] First, we transform the knowledge state H t+1 With the interaction feature E t+1 Splice them together and pass them through MLP to get two feature sequences G and S representing the behavior, G = MLP G (H t+1 ∪E t+1 ),S=MLP S (H t+1 ∪E t+1 ), where MLP G , MLP S There are two independent MLP networks with different weight matrices and bias vectors. σ is the activation function. Then use the following response formula y t+1 =L q *(1-S)+(1-L q )*G, we can get the predicted value of the final student learning behavior.

[0056] In model training, calculating loss is an important step in evaluating prediction accuracy. We choose to use the cross entropy loss function to calculate the loss. Its adaptability to classification tasks and sensitivity to wrong predictions make the model pay more attention to difficult-to-classify samples during training, ultimately improving the overall performance. t+1 Calculate the cross loss entropy with the actual answer result, and the loss function is recorded as Where T represents the total number of time steps, r t+1 Represents the actual response result of the student at time step t+1.

[0057] (3) Dynamic update process

[0058] In order to enable the model to capture the long-term dependencies between knowledge points and learning contexts, we incorporate the feature information of the current interaction unit into the interaction history information and use the self-attention mechanism to dynamically update the knowledge state and adjust the weights of the knowledge points, such as Figure 1 Part (c).

[0059] At each time step, we update the current question state sequence, concept state sequence, and global state sequence, which can be used to acquire representation sequences for the next batch of sequences and perform residual connection operations in the process of knowledge state acquisition.

[0060] For the problem state, we use K in the attention mechanism q and V q Set as the output of the knowledge state acquisition module, that is, the knowledge state based on the problem characteristics The query Q q Set to the interaction unit E at time t+1 t+1 , put them into the self-attention mechanism to get S q =MultiHead(Q q ,K q ,V q ). We then use a feedforward neural network to incorporate the results of the self-attention mechanism into the model and consider the interactions between different underlying dimensions, using the following response formula, FFN(S q )=ReLU(S q W (1) +b (1) )W (2) +b (2) ,in are all learnable parameters during training. Finally, we make a residual connection to obtain the updated problem state sequence representation

[0061] For the concept state, we will key K c Sum value V c Set as the output of the knowledge state acquisition module, that is, the knowledge state based on concept features The query Q c Set to the interaction unit E at time t+1 t+1 , put them into the self-attention mechanism to get S c =MultiHead(Q c ,K c ,V c ). Then use the feedforward neural network FFN (S c )=ReLU(S c W (1) +b (1) )W (2) +b (2) Finally, the updated concept state sequence representation is obtained through residual connection

[0062] For the global state, we have the key K o Sum value Vo Set to the output of the knowledge state acquisition module, that is, the knowledge state based on global features The query Q o Set to the interaction unit E at time t+1 t+1 , put them into the self-attention mechanism to get S o =MultiHead(Q o ,K o ,V o ). Then use the feedforward neural network FFN (S o )=ReLU(S o W (1) +b (1) )W (2) +b (2) ,Finally, the updated global state sequence representation is obtained through residual connection

[0063] Table 1 below is the comparison results between this method and the baseline method. We select classic or state-of-the-art methods as baseline methods. The top of the table is the dataset, and the rightmost is the existing model. The data in the table are AUC and ACC indicators, and the best results are highlighted in bold. It can be seen from the table that our method DMKT always outperforms other baseline models in all datasets, demonstrating its significant advantage in improving the accuracy of response prediction.

[0064] Table 1 Comparison results between the method of the present invention and the existing method

[0065]

[0066] On the other hand, an embodiment of the present invention further provides a knowledge tracking system based on multi-perspective dynamic evolution, comprising:

[0067] one or more processors;

[0068] A storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the knowledge tracking method based on multi-perspective dynamic evolution as described in the above scheme.

[0069] The contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A knowledge tracking method based on multi-perspective dynamic evolution, characterized by The method comprises the following steps: Step 1: Obtain the input sequence based on the question feature, the input sequence based on the concept feature, and the input sequence based on the global feature at the current time t+1, and compare the three obtained sequences with the interaction feature E at the current time. t+1 splicing, and input them into the LSTM model respectively to obtain the corresponding output features, and splicing the three output features as the representation of the final knowledge state at the current moment H t+1 ; Step 2: Use IRT to calculate the probability L of the student answering the question correctly, combined with the final knowledge state q ,Combining the students’ answers with the inconsistent behaviors of concept mastery, we can obtain the predicted values ​​of students’ learning behaviors; In step 2, the behavior of students’ answers being inconsistent with their mastery of the concepts includes two types: guessing the correct answer without truly mastering it and making mistakes despite knowing the answer. First, the knowledge state H t+1 With the interaction feature E t+1 Splice them together and pass them through MLP to get two feature sequences G and S representing the behavior, G = MLP G (H t+1 ∪E t+1 ), S = MLP S (H t+1 ∪E t+1 ), where MLP G , MLP S There are two independent MLP networks with different weight matrices and bias vectors. σ is the activation function, and the following response formula y is used. t+1 =L q *(1-S)+(1-L q )*G, we can get the predicted value of the final student learning behavior.

2. The knowledge tracking method based on multi-perspective dynamic evolution according to claim 1 is characterized in that: In step 1, first set the problem feature matrix Q∈R n×d , concept feature matrix C∈R m×d , the answer feature matrix R∈R 2×d , where n is the number of questions, m is the number of concepts, and d is the feature dimension, then the interactive feature is represented as Among them, Qq t+1 and are the problem q at time t+1 in Q and C respectively t+1 and the concept c at time t+1 t+1 The line where diffq t+1 Yes t+1 The difficulty captures the degree to which the problem varies relative to its concept, diffq t+1 and are all trainable parameters.

3. The knowledge tracking method based on multi-perspective dynamic evolution according to claim 2 is characterized in that: In step 1, for the input sequence based on the problem feature, extract the sequence containing the current problem q from the problem state sequence t+1 The interactive units are concatenated to obtain the input sequence based on the problem features; For the input sequence based on concept features, the state sequence containing the current concept c is extracted from the concept state sequence. t+1 The interactive units are concatenated to obtain the input sequence based on the concept features; For the input sequence based on global features, the global state sequence is directly used as the input sequence based on global features; For any interaction unit X before the current time t+1 t =(q t ,c t ,r t ), the feature is calculated as X t =E t +Rr t , where Rr t is the rth t OK, q t ,c t ,r t They represent the question, concept and answer at time t respectively.

4. The knowledge tracking method based on multi-perspective dynamic evolution according to claim 1 is characterized in that: In order to enable the LSTM model to track the learner's knowledge mastery in real time, a residual connection is made between the output features of the LSTM model and the knowledge state obtained by the dynamic update module at the previous moment to obtain the knowledge state based on the problem features at the current moment. Knowledge state based on concept characteristics and the knowledge state based on global features Finally, these three knowledge states are concatenated to obtain the representation of the final knowledge state, namely The dynamic update module is implemented through the self-attention mechanism.

5. The knowledge tracking method based on multi-perspective dynamic evolution according to claim 1 is characterized in that: The specific process of using the dynamic update module to obtain the current knowledge state is as follows: For the problem state, K in the attention mechanism q and V q Set to knowledge state based on problem characteristics The query Q q Set to the interaction unit E at time t+1 t+1 , put them into the self-attention mechanism to get S q =MultiHead(Q q ,K q ,V q ); then a feedforward neural network is used to incorporate the results of the self-attention mechanism into the model, and the interactions between different underlying dimensions are considered, using the following response formula, FFN(S q )=ReLU(S q W (1) +b (1) )W (2) +b (2) ,in These are all learnable parameters during training. Finally, a residual connection is made to obtain an updated representation of the problem state sequence. For the conceptual state, the key K c Sum value V c Set as knowledge state based on concept features The query Q c Set to the interaction unit E at time t+1 t+1 , put them into the self-attention mechanism to get S c =MultiHead(Q c ,K c ,V c ); then use the feedforward neural network to process, and finally get the updated concept state sequence representation through the residual connection For the global state, the key K o Sum value V o Set to global feature-based knowledge state The query Q o Set to the interaction unit E at time t+1 t+1 , put them into the self-attention mechanism to get S o =MultiHead(Q o ,K o ,V o ); then use the feedforward neural network to process, and finally, get the updated global state sequence representation through the residual connection in and They represent the problem state sequence, concept state sequence and global state sequence at the current moment respectively.

6. The knowledge tracking method based on multi-perspective dynamic evolution according to claim 1 is characterized by: Calculating the probability of a student answering a question correctly is implemented as follows: First, calculate the student's mastery level of concept i at time t+1 The final knowledge state H t+1 With the interaction feature E t+1 Put them together and pass them through MLP to get the level of conceptual mastery; Then, the problems to be processed at time t+1 are processed through MLP to calculate the discriminant parameters α of specific problems under concept i. i , the difficulty of a specific problem under concept i is β i Finally, the IRT model is applied, using the following response formula Where α = σ[MLP α (q t+1 )],β=σ[MLP β (q t+1 )]; MLP α and MLP β There are two independent MLP networks with different weight matrices and bias vectors. σ is the activation function, and N c Represents the total number of concepts, and the probability L that students correctly answer questions under concept i is obtained qi .

7. The knowledge tracking method based on multi-perspective dynamic evolution according to claim 1 is characterized by: Step 2 also includes converting y t+1 Calculate the cross loss entropy with the actual answer result, and the loss function is recorded as Where T represents the total number of time steps, r t+1 Represents the actual response result of the student at time step t+1.

8. A knowledge tracking system based on multi-perspective dynamic evolution, characterized in that: include: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the knowledge tracking method based on multi-perspective dynamic evolution as described in any one of claims 1 to 7.

Images