A knowledge tracking method, device, equipment and storage medium

Abstract

The present invention relates to a knowledge tracking method, device, equipment and storage medium, wherein the knowledge tracking method comprises: counting input data and classifying them; constructing a basic graph according to the correlation between questions and knowledge points and the corresponding embedding matrix; combining the basic graph with the online augmentation function to obtain an online augmentation graph; generating an online representation according to the online augmentation graph; predicting the initial features of the questions according to the online representation; fusing the features of multiple attributes to obtain interactive features; extracting the knowledge state according to the interactive features and the predicted initial features of the questions to obtain the implicit knowledge state; refining the implicit knowledge state to obtain the model prediction value. The knowledge tracking method of the present invention can take into account various learning features and can more accurately predict the future learning situation of students.

Description

Technical Field

[0001] The present invention relates to the fields of smart education and intelligent tutoring, and in particular to a knowledge tracking method. Background Art

[0002] In recent years, with the development and popularization of online education platforms, online education has become one of the mainstream teaching methods, and more and more students, teachers and parents are paying attention to the teaching system of online education plus artificial intelligence. Compared with traditional offline education, the most significant advantage of online education is that it can retain learners' detailed learning trajectories and provide conditions for investigating the effectiveness of learners' behavior under different trajectories. Learners can check their current knowledge points and learn weak knowledge points in a targeted manner to achieve the goal of mastering all knowledge points of the course; teachers can also conduct personalized teaching and adjust teaching goals purposefully according to students' knowledge points.

[0003] Providing personalized teaching guidance to learners requires accurate assessment of the learners' knowledge status, and knowledge tracking technology plays a key role in this. Knowledge tracking technology evaluates students' learning level by analyzing their previous answer records and predicts their future performance. Specifically, knowledge tracking technology builds a dynamic model that can reflect students' knowledge level over time based on their historical answer data and is used to judge students' understanding of each knowledge point. Such a dynamic model is crucial for monitoring students' learning progress.

[0004] Knowledge tracking based on graph neural networks is the most advanced knowledge tracking technology. The knowledge tracking model used in it uses additive aggregation or average aggregation when aggregating features, which fails to reflect the importance of different nodes. Recurrent neural networks and their derivative models are widely used as basic architectures. However, these models often have the problem of difficulty in capturing long-term dependencies. In addition, for learners, learning features include knowledge points, questions, and answers. In the research on knowledge tracking, the use of learning features is relatively small, and the large number of learning features in the data set are not reasonably utilized. In addition, the question embedding method has data sparsity or ambiguity in interpretation. Summary of the invention

[0005] Based on this, the purpose of the present invention is to provide a knowledge tracking method that can make full use of students' various learning characteristics and more accurately predict students' future learning performance.

[0006] A knowledge tracing method, comprising:

[0007] Counting the input data and classifying them into question data, knowledge point data, feature data of multiple attributes, and answer data; wherein the feature data of multiple attributes include average reaction time feature, average accuracy feature, question type feature, and average number of attempts feature;

[0008] Construct a basic graph based on the relationship between questions and knowledge points and the corresponding embedding matrix;

[0009] Combining the basic graph with the online augmentation function to obtain the online augmented graph;

[0010] Generate online representation based on online augmented graph;

[0011] Obtain the predicted initial features of the question based on the online representation prediction;

[0012] The features of multiple attributes are fused to obtain interactive features;

[0013] Extract the knowledge state based on the interactive features and the predicted initial features of the question to obtain the implicit knowledge state;

[0014] Refine the implicit knowledge state to obtain the model prediction value.

[0015] Furthermore, through the first connection layer The average reaction time feature, the average accuracy feature and the question type feature are combined to obtain the fusion feature:

[0016]

[0017] Through the second connection layer Linearly fuse the fusion features with the predicted initial features of the question to obtain the question fusion features:

[0018]

[0019] Through the third link layer The interactive features are obtained by connecting the question fusion features with the answer features and the average number of attempts features:

[0020]

[0021] in , , , , , , , They are average reaction time feature, average accuracy feature, question type feature, fusion feature, answer feature, average number of attempts feature, question fusion feature, and interaction feature.

[0022] Furthermore, the knowledge state is calculated based on the interaction features, the predicted initial features of the question, and the parameter matrix of the model:

[0023] , , ;

[0024] in, , , is the parameter matrix of the model;

[0025] The dot product attention mechanism is used to extract the knowledge state of the student's past answer history to obtain the implicit knowledge state:

[0026]

[0027] Where d is the dimension of the input vector, It is an implicit knowledge state.

[0028] Furthermore, before the formal operation, training enhancement is performed by the following method to obtain the optimal initial prediction features: some nodes are randomly selected from the node set E of the basic graph, and a modified subset is formed according to the proportion of their deletion probability , get the set of all current graph nodes :

[0029]

[0030] Randomly select some edges from the edge set V of the base graph and form a modified subset in proportion to their probability of being deleted , get the set of all current graph edges :

[0031]

[0032] in, is the edge of the base graph, is a node of the base graph, is the probability of eliminating an edge, is the probability of eliminating a node.

[0033] Furthermore, an online representation is generated according to the online augmented graph through an online autoencoding method. Before the formal operation, training enhancement is performed through the following target autoencoding method to obtain the optimal prediction initial features: the target augmented graph is obtained by combining the target augmentation function, the target representation is output according to the target augmented graph, and the target representation is compared with the predicted initial features of the question, and the online parameters of the online autoencoder are corrected to make the online representation closer to the target representation.

[0034] Furthermore, after inputting the data, we apply the known random graph augmentation function Augment the base graph to obtain a target augmented graph , where the node features , the adjacency matrix , is the number of nodes, is the feature vector dimension, is the number of edges; generate the target representation based on the target augmented graph :

[0035]

[0036] in are the parameters of the target autoencoder, To determine the target parameters of the target autoencoder parameter result; through a series of linear transformations and nonlinear activation functions, it is realized from arrive The target is represented by Projected into the new representation space, we get ,in, To represent truth;

[0037] According to the cosine similarity and The gradient relationship is:

[0038]

[0039]

[0040] in, Online parameters that determine the online autoencoder parameter results; is the learning rate, and the final update is based on the target representation relative to The gradient of is calculated using the Adam optimization method; The loss function calculated for cosine similarity;

[0041] Update online parameters , so that the prediction target of each node is represented by Closer to the true target representation .

[0042] Furthermore, after obtaining the model prediction value, the following steps are used to update the parameters:

[0043] The computation nodes are embedded as:

[0044]

[0045] Where W represents the weight matrix; is the attention coefficient, which indicates the attention of node j to node i, and is obtained by the node attention formula:

[0046]

[0047] in, represents the set of adjacent nodes of node i, represents a nonlinear activation function, represents the weight matrix, T represents the transpose, represents the concatenation operation;

[0048] The attention factor With node features Normalize and linearly combine to obtain the high-level features of each node:

[0049]

[0050] in, For advanced features, represents a nonlinear activation function;

[0051] A multi-head attention mechanism is added to concatenate the results of multiple attention calculations and obtain the updated features through the average operation of the final layer:

[0052]

[0053] in, To update the features.

[0054] The present invention also provides a knowledge tracking device, comprising:

[0055] A data statistics module is used to count the input data and classify the data into question data, knowledge point data, feature data of multiple attributes, and answer data; wherein the feature data of multiple attributes include average reaction time feature, average accuracy feature, question type feature, and average number of attempts feature;

[0056] Graph construction module, used to construct the basic graph according to the relationship between questions and knowledge points and the corresponding embedding matrix;

[0057] An enhancement module, used to combine the basic graph with the online augmentation function to obtain an online augmented graph;

[0058] The encoding module is used to generate online representations based on the online augmented graph;

[0059] A prediction module is used to predict the initial features of the question based on the online representation;

[0060] Feature fusion module, used to fuse features of multiple attributes to obtain interactive features;

[0061] The knowledge evolution module is used to extract the knowledge state according to the interaction features and the predicted initial features of the question to obtain the implicit knowledge state;

[0062] The refinement module is used to refine the implicit knowledge state to obtain the model prediction value.

[0063] The present invention also provides a knowledge tracking device, comprising a memory and at least one processor, wherein the memory stores instructions, and the memory and the at least one processor are interconnected via lines; the at least one processor calls the instructions in the memory so that the knowledge tracking device executes the steps of the knowledge tracking method as described above.

[0064] The present invention also provides a computer-readable storage medium for knowledge tracking, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps of the knowledge tracking method described above are implemented.

[0065] For better understanding and implementation, the present invention is described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a schematic diagram of the structure of the knowledge tracking device of the present invention.

[0067] Figure 2 Schematic diagram of the knowledge tracking method flow of the present invention. DETAILED DESCRIPTION

[0068] The applicant carefully analyzed the existing knowledge tracking model and found that it focused on the application of the attention mechanism, but made less use of learning features and was unable to distinguish the specific content of some features. The applicant tried to add more learning features, use graph contrast learning to clearly represent features such as questions, linearly fuse multiple features of questions, and obtain more accurate learning predictions through attention extraction.

[0069] Based on the above invention concept, please refer to Figure 1 The present invention provides a knowledge tracking device, including a preprocessing module, a data statistics module, a graph construction module, an enhancement module, an encoding module, a prediction module, a feature fusion module, a knowledge evolution module and a refinement module. Figure 2 The operation process of the knowledge tracking device of the present invention is demonstrated.

[0070] The preprocessing module is used to obtain online answering data, filter out invalid information and error information, and remove interactive data with an interaction volume of less than 5 to obtain model data. By removing useless and messy information, the amount of model calculation can be reduced, and the data is more accurate during statistics, thereby improving the accuracy of subsequent predictions.

[0071] The data statistics module is used to count the model data and classify the data into question data, knowledge point data, multiple attribute feature data and answer data. The multiple attribute feature data are used to measure the difficulty of the question, including average reaction time, average number of attempts, average accuracy and question type.

[0072] The graph construction module is used to construct a basic graph based on the relationship between the question and the knowledge point and the corresponding embedding matrix. The relationship between the question and the knowledge point can be represented by a binary graph. It is represented by Q, where Q is the question and S is the knowledge point. is a set of edge information, k is the number of edges, e is the edge, and the node on the edge is v. The sentences and paragraphs corresponding to the question are mapped to a 384-dimensional dense vector space. After obtaining the 384-dimensional dense vector, a linear layer is added to reduce the 384-dimensional dense vector to 256 dimensions, and the question embedding matrix is ​​obtained. ,in , D represents the embedding size, and the topic embedding matrix Each row in corresponds to a question. The sentences and paragraphs corresponding to the knowledge points are also mapped to a 384-dimensional dense vector space. After obtaining the 384-dimensional dense vector, a linear layer is added to reduce the 384-dimensional dense vector to 256 dimensions, and the knowledge point embedding matrix is ​​obtained. ,in , D represents the embedding size, and the knowledge point embedding matrix Each row in corresponds to a knowledge point. Embedding the question embedding matrix and the knowledge point embedding matrix into the binary graph can get the basic graph , where the node features , the adjacency matrix , N is the number of nodes, O is the dimension of the feature vector, and L is the number of edges.

[0073] Similarly, for other attribute features that measure the difficulty of the question, the corresponding time embedding matrix can be obtained respectively , frequency embedding matrix , Accuracy embedding matrix , type embedding matrix , where M is the maximum value of the average reaction time, P is the maximum value of the average number of attempts, C is the maximum value of the average accuracy, and F is the number of question types.

[0074] The enhancement module is used to combine the basic graph with the online augmentation function to obtain an online augmented graph. Augment the base graph to obtain the online augmented graph , where the node features , the adjacency matrix , N is the number of nodes, O is the dimension of the feature vector, and L is the number of edges. Compared with the basic graph, the augmented online augmented graph greatly enhances the ability to express information, thereby improving the accuracy of subsequent predictions.

[0075] The online encoding module is used to generate an online representation based on the online augmented graph. The online encoding module uses an online encoder To encode, For online parameters, continuous real-time operation is obtained online representation :

[0076] Specifically, GAT may be used as an encoder.

[0077] Understandably, online representation The actual initial features of the question in one augmented form may vary in different augmented forms.

[0078] The prediction module is used to predict the initial features of the topic based on the online representation of the encoding module. The prediction module includes a node-level predictor that receives the information from the online encoder. Online representation , and output the target representation based on the online representation , where the target representation contains the predicted initial features of the title , Represents the i-th row vector of question Q.

[0079] The feature fusion module is used to fuse the features of multiple attributes to obtain interactive features. Specifically, through the first connection layer The average reaction time feature, the average accuracy feature and the question type feature are combined to obtain the fusion feature:

[0080]

[0081] Through the second connection layer Linearly fuse the fusion features with the predicted initial features of the question to obtain the question fusion features:

[0082]

[0083] Finally, through the third connection layer The interactive features are obtained by connecting the question fusion features with the answer features and the average number of attempts features:

[0084]

[0085] in , , , , , , , They are average reaction time feature, average accuracy feature, question type feature, fusion feature, answer feature, average number of attempts feature, question fusion feature, and interaction feature.

[0086] The knowledge evolution module is used to extract the knowledge state according to the interaction features and the predicted initial features of the question to obtain the implicit knowledge state. Specifically, the knowledge state is calculated according to the interaction features, the predicted initial features of the question and the parameter matrix of the model:

[0087] , ,

[0088] in, , , is the parameter matrix of the model;

[0089] The dot product attention mechanism is used to extract the knowledge state of the student's past answer history to obtain the implicit knowledge state:

[0090]

[0091] Where d is the dimension of the input vector, It is an implicit knowledge state.

[0092] The refinement module is used to refine the implicit knowledge state to obtain the model prediction value. Specifically, the implicit knowledge state obtained by the knowledge evolution module is refined through two connection layers to obtain the overall optimization function of the knowledge state:

[0093]

[0094] in, , , , , and All are trainable parameters, and T represents transpose.

[0095] The loss function of the model is the binary cross entropy (BCE) between the model prediction value and the true value of the data:

[0096]

[0097] in, is a binary label 0 or 1, Is the output of The probability of the label.

[0098] Before the official operation, the prediction module can also be trained and enhanced to obtain the optimal prediction topic features.

[0099] In one embodiment, the edges or nodes to be deleted are randomly selected from the base graph and then combined with the online augmentation function to obtain the online augmented graph. Specifically, in form, some nodes are randomly selected from E of the base graph, and a modified subset is formed according to the proportion of their probability of being deleted. , get the set of all current graph nodes :

[0100]

[0101] Randomly select some edges from the base graph V, in proportion to their probability of being deleted, to form a modified subset , get the set of all current graph edges :

[0102]

[0103] in, is the edge of the base graph, v is the node of the base graph, is the probability of eliminating an edge, is the probability of eliminating a node, and Reflects the importance of edges and nodes respectively. In this way, the function ensures that unimportant edges or points are eliminated first while ensuring that the important structure of the graph is not damaged.

[0104] In another embodiment, a target encoding module is further included, which is used to obtain a target augmentation map in combination with the target augmentation function, output a target representation according to the target augmentation map, and correct the online parameters of the online encoding module by comparing the target representation with the predicted initial features of the question. Make the online representation closer to the target representation.

[0105] The target encoding module uses a target encoder To construct the target representation, its target parameters are , with the online encoder Online parameters Different. After inputting the graph data, apply the random graph augmentation function Augment the base graph to obtain the target augmented graph , where the node features , the adjacency matrix , N is the number of nodes, O is the feature vector dimension, and L is the number of edges. The target encoder generates a target representation based on the target augmented graph :

[0106]

[0107] Specifically, the GAT graph attention autoencoder is used as the target encoder as well as the online encoder.

[0108] Then, the prediction module Acting on the target representation , projecting it onto ,get ,in, is the true representation. Prediction module Here it plays a role of conversion, embedding the original node into Map to a new representation space and get the real representation in the new representation space , which is more suitable for subsequent loss calculation and model optimization. In actual implementation, the prediction module Using the existing multi-layer perceptron (MLP), the update parameters are , through a series of linear transformations and nonlinear activation functions to achieve arrive 's mapping.

[0109] According to the cosine similarity and The gradient relationship is:

[0110]

[0111]

[0112] Update the encoding module Online parameters , so that the prediction target of each node is represented by Closer to the true target representation ,in, is the learning rate, and the final update is based only on the target representation relative to The gradient calculation of , using the Adam optimization method, The loss function calculated for cosine similarity.

[0113] After obtaining the model prediction value, the following methods can be used to update the model parameters to improve the accuracy of the prediction value:

[0114] The computation nodes are embedded as:

[0115]

[0116] Where W represents the weight matrix; is the attention coefficient, which indicates the attention of node j to node i, and is obtained by the node attention formula:

[0117]

[0118] in, represents the set of adjacent nodes of node i, represents a nonlinear activation function, represents the weight matrix, T represents the transpose, Represents a concatenation operation.

[0119] Finally, the attention coefficient With node features Normalize and linearly combine to obtain the high-level features of each node:

[0120] in, For advanced features, represents a non-linear activation function.

[0121] When considering the multi-head attention mechanism, we can also concatenate the results of multiple attention calculations to calculate the feature representation of each node, and obtain the updated features through the average operation of the final layer:

[0122]

[0123] in, To update the features.

[0124] In order to verify the actual effect of the knowledge tracking device of the present invention, the following test method is used for evaluation.

[0125] First, we obtain three real open source datasets as the test data source, namely ASSISTments2009, ASSISTments2012, and ASSISTments2017. Please refer to Table 1, which shows the statistical data of the three datasets.

[0126] Table 1

[0127]

[0128] Among them, the ASSISTments2009 dataset is about math exercises, collected from the free online tutoring ASSISTments platform in the 2009-2010 academic year. It has been widely used as a standard benchmark for KT methods in the past decade. It includes 3,374,115 interactions, 4,661 students, 17,737 questions, and 123 knowledge points.

[0129] The ASSISTments2012 dataset is about math exercises, collected from the free online tutoring ASSISTments platform in the 2012-2013 academic year. It includes 2,630,080 interactions, 28,834 students, 50,989 questions, and 198 knowledge points.

[0130] The ASSISTments2017 dataset is about math exercises, collected from the free online tutoring ASSISTments platform in the 2017-2018 academic year. It includes 942,816 interactions, 1,709 students, 3,163 questions, and 102 knowledge points.

[0131] The Adam optimizer was used to train the model, and the learning rate was set to 0.001. To prevent overfitting, the dropout was set to 0.6. All models were implemented using the PyTorch framework, and the hyperparameter settings were kept consistent with the open source code. The evaluation indicators area under the curve (AUC) and accuracy (ACC) were used as evaluation indicators.

[0132] Table 2

[0133] Model ASSISTments2009 ASSISTments2012 ASSISTments2017 DKT 0.7538 0.7267 0.7187 GKT 0.7532 0.7534 0.7267 SAKT 0.7423 0.7027 0.7053 SAINT 0.7563 0.7344 0.7124 simpleKT 0.7878 0.7705 0.7591 GCLFKT 0.8225 0.8112 0.8409

[0134] Table 3

[0135] Model ASSISTments2009 ASSISTments2012 ASSISTments2017 DKT 0.7036 0.7325 0.6867 GKT 0.7012 0.7356 0.7052 SAKT 0.6932 0.7226 0.7085 SAINT 0.7042 0.7126 0.7114 simpleKT 0.7234 0.7505 0.7079 GCLFKT 0.7561 0.7963 0.7568

[0136] Please refer to Table 2 and Table 3, where Table 2 is a comparison table of AUC values ​​of the GCLFKT model of the present invention and other models on three data sets, and Table 3 is a comparison table of ACC values ​​of the GCLFKT model of the present invention and other models on three data sets. The test results show that the results of GCLFKT on the three data sets are better than those of other models. Compared with DKT and GKT, the AUC index is improved by about 7%. This is because GCLFKT uses graph contrast learning to obtain better question feature representation. Compared with SAKT, SAINT and simpleKT, the AUC index is improved by about 8%, 7% and 4%. This is because GCLFKT integrates rich question features and can better distinguish questions with the same knowledge points than other models.

[0137] Based on the above-mentioned knowledge tracking method, the present invention also provides a knowledge tracking device, including a memory and at least one processor, the memory stores instructions, and the memory and the at least one processor are interconnected through lines; the at least one processor calls the instructions in the memory so that the knowledge tracking device executes the steps of the knowledge tracking method described in any one of the above-mentioned embodiments.

[0138] Based on the above-mentioned knowledge tracking method, the present invention further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the knowledge tracking method described in any one of the above-mentioned embodiments are implemented.

[0139] The present invention may take the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing program code. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include but are not limited to: phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0140] The above-mentioned embodiments only express several implementation methods of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, and the present invention is also intended to include these modifications and modifications.

Claims

1. A knowledge tracking method, characterized in that: include: Counting the input data and classifying them into question data, knowledge point data, feature data of multiple attributes, and answer data; wherein the feature data of multiple attributes include average reaction time feature, average accuracy feature, question type feature, and average number of attempts feature; Construct a basic graph based on the relationship between questions and knowledge points and the corresponding embedding matrix; Combining the basic graph with the online augmentation function to obtain the online augmented graph; Generate online representation based on online augmented graph; Obtain the predicted initial features of the question based on the online representation prediction; The features of multiple attributes are combined to obtain interactive features: through the first connection layer The average reaction time feature, the average accuracy feature and the question type feature are combined to obtain the fusion feature: Through the second connection layer Linearly fuse the fusion features with the predicted initial features of the question to obtain the question fusion features: Through the third link layer The interactive features are obtained by connecting the question fusion features with the answer features and the average number of attempts features: in , , , , , , , They are average reaction time feature, average accuracy feature, question type feature, fusion feature, answer feature, average number of attempts feature, question fusion feature, and interaction feature. is the initial feature for predicting the topic; The knowledge state is extracted based on the interaction features, the predicted initial features of the question, and the parameter matrix of the model to obtain the implicit knowledge state: , , ; in, , , is the parameter matrix of the model; The dot product attention mechanism is used to extract the knowledge state of the student's past answer history to obtain the implicit knowledge state: Where d is the dimension of the input vector, is the implicit knowledge state, for The transpose of Refine the implicit knowledge state to obtain the model prediction value.

2. The knowledge tracking method according to claim 1, characterized in that: Before the official operation, the following method is used to perform training enhancement to obtain the optimal initial prediction features: some nodes are randomly selected from the node set V of the base graph, and a modified subset is formed according to the proportion of their deletion probability , get the set of all current graph nodes : Randomly select some edges from the edge set E of the base graph and form a modified subset in proportion to their probability of being deleted , get the set of all current graph edges : in, is the edge of the base graph, is a node of the base graph, is the probability of eliminating an edge, is the probability of eliminating a node.

3. The knowledge tracking method according to claim 2, characterized in that: Generate online representation based on online augmented graph through online autoencoding method; Before the official operation, the following target autoencoding method is used for training enhancement to obtain the optimal initial prediction features: The target augmentation function is combined to obtain the target augmentation graph, and the target representation is output according to the target augmentation graph. By comparing the target representation with the predicted initial features of the question, the online parameters of the online autoencoder are corrected to make the online representation closer to the target representation.

4. The knowledge tracking method according to any one of claims 2 or 3, characterized in that After inputting the data, we apply the known random graph augmentation function Augment the base graph to obtain a target augmented graph , where the node features , the adjacency matrix , is the number of nodes, is the feature vector dimension, is the number of edges; generate the target representation based on the target augmented graph : in are the parameters of the target autoencoder, To determine the target parameters of the target autoencoder parameter result; through a series of linear transformations and nonlinear activation functions, it is realized from arrive The target is represented by Projected into the new representation space, we get ,in, is the predicted target representation obtained based on the online representation prediction, is a trainable parameter, is a constant; According to the cosine similarity and The gradient relationship is: in, Online parameters that determine the online autoencoder parameter results; is the learning rate, and the final update is based on the target representation relative to The gradient of is calculated using the Adam optimization method; The loss function calculated for cosine similarity; Update online parameters , so that the prediction target of each node is represented by Closer to the true target representation .

5. The knowledge tracking method according to claim 4, characterized in that: After obtaining the model prediction value, the following steps are used to update the parameters: The computation nodes are embedded as: Where W represents the weight matrix; is the attention coefficient, which indicates the attention of node j to node i, and is obtained by the node attention formula: in, represents the set of adjacent nodes of node i, represents a nonlinear activation function, represents the weight matrix, T represents the transpose, represents the concatenation operation; The attention factor With node features Normalize and linearly combine to obtain the high-level features of each node: in, For advanced features, represents a nonlinear activation function; A multi-head attention mechanism is added to concatenate the results of multiple attention calculations and obtain the updated features through the average operation of the final layer: in, To update the features.

6. A knowledge tracking device, characterized in that: include: A data statistics module is used to count the input data and classify the data into question data, knowledge point data, feature data of multiple attributes, and answer data; wherein the feature data of multiple attributes include average reaction time feature, average accuracy feature, question type feature, and average number of attempts feature; Graph construction module, used to construct the basic graph according to the relationship between questions and knowledge points and the corresponding embedding matrix; An enhancement module, used to combine the basic graph with the online augmentation function to obtain an online augmented graph; The encoding module is used to generate online representations based on the online augmented graph; A prediction module is used to predict the initial features of the question based on the online representation; The feature fusion module is used to fuse the features of multiple attributes to obtain interactive features: through the first connection layer The average reaction time feature, the average accuracy feature and the question type feature are combined to obtain the fusion feature: Through the second connection layer Linearly fuse the fusion features with the predicted initial features of the question to obtain the question fusion features: Through the third link layer The interactive features are obtained by connecting the question fusion features with the answer features and the average number of attempts features: in , , , , , , , They are average reaction time feature, average accuracy feature, question type feature, fusion feature, answer feature, average number of attempts feature, question fusion feature, and interaction feature. is the initial feature for predicting the topic; The knowledge evolution module is used to extract the knowledge state based on the interaction features, the predicted initial features of the question, and the parameter matrix of the model to obtain the implicit knowledge state: , , ; in, , , is the parameter matrix of the model; The dot product attention mechanism is used to extract the knowledge state of the student's past answer history to obtain the implicit knowledge state: Where d is the dimension of the input vector, is the implicit knowledge state, for The transpose of The refinement module is used to refine the implicit knowledge state to obtain the model prediction value.

7. A knowledge tracking device, characterized in that: It includes a memory and at least one processor, the memory stores instructions, and the memory and the at least one processor are interconnected through lines; the at least one processor calls the instructions in the memory to enable the knowledge tracking device to perform the steps of the knowledge tracking method as described in any one of claims 1-5.

8. A computer-readable storage medium for knowledge tracking, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the knowledge tracking method according to any one of claims 1 to 5.

Images