A knowledge tracking method based on the evolution of knowledge relationships using dynamic graph neural networks

Abstract

This invention relates to a knowledge tracking method based on the evolution of knowledge relationships using a dynamic graph neural network, belonging to the field of knowledge tracking technology, and solves the problem of inaccurate tracking of knowledge states in existing technologies. The method includes: constructing a graph structure based on students, questions, and knowledge points at each historical moment; constructing and training a dynamic graph neural network, which includes a feature extraction module, a prediction module, and a knowledge state calculation module; using the feature extraction module to obtain time-aware features of the current student, current question, and current knowledge point; inputting the time-aware features of the current student and current question, as well as the current student's knowledge state at the current knowledge point, into the prediction module to predict the probability of a correct answer; and inputting the time-aware features of the current student, current question, and current knowledge point, as well as the interaction result of the current student's answer to the current question, into the knowledge state calculation module to update the current student's knowledge state. This achieves accurate tracking of the student's knowledge state.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge tracking, and in particular to a knowledge tracking method based on dynamic graph neural network. BACKGROUND

[0002] Intelligent teaching guidance systems based on artificial intelligence technology have been widely used in online education systems, constantly promoting the intelligentization of online education. Knowledge tracking, as an important technology in intelligent teaching guidance systems, can help educators understand students' learning situation and knowledge mastery, thereby providing personalized learning guidance. The main task of knowledge tracking is to evaluate the students' continuously developing knowledge state, i.e., knowledge mastery, based on the students' historical answer sequences, so as to predict the students' future learning performance.

[0003] In the field of smart education, as students continue to learn, their knowledge point relationship network will change continuously, mainly reflected in: students will incorporate newly learned knowledge points into the network, and reshape the relationship between previously mastered knowledge points and other knowledge points. At the same time, influenced by the knowledge transfer effect in educational psychology, the learning benefits of a student on a certain knowledge point will be transferred to its associated knowledge points, and the direction of the student's knowledge transfer is constrained by the relationship network of knowledge points. Traditional knowledge tracking methods model the relationship network of knowledge points statically, and this static assumption ignores the dynamic changes of knowledge relationships, which can lead to incorrect knowledge transfer, especially in long-term learning sequences, which can accumulate errors and cannot accurately track the students' knowledge state. SUMMARY

[0004] In view of the above analysis, the embodiments of the present application aim to provide a knowledge tracking method based on dynamic graph neural network to solve the problem that the existing method cannot accurately track the students' knowledge state.

[0005] In one aspect, the embodiments of the present application provide a knowledge tracking method based on dynamic graph neural network, comprising the following steps:

[0006] Constructing a graph structure for each historical time point based on the students, questions and knowledge points corresponding to the questions at each historical time point;

[0007] Constructing and training a dynamic graph neural network to obtain a trained dynamic graph neural network, the trained dynamic graph neural network comprising a feature extraction module, a prediction module and a knowledge state calculation module;

[0008] Obtaining time-aware features of the current student, the current question and the current knowledge point at the current time point based on the graph structure of the historical time point using the feature extraction module;

[0009] inputting time-aware features of the current student and the current question, and a knowledge state of the current student on the current knowledge point into the prediction module to predict a probability of the current student correctly answering the current question;

[0010] inputting the time-aware features of the current student, the current question and the current knowledge point, and an interaction result of the current student answering the current question into the knowledge state calculation module, and updating the knowledge state of the current student based on a knowledge point relationship matrix.

[0011] Based on the further improvement of the above method, the feature extraction module comprises:

[0012] a sequence generation module configured to generate a historical question sequence and a historical knowledge point sequence corresponding to the current student at a current time based on a graph structure at a historical time, and a historical student sequence corresponding to the current question;

[0013] a student representation module configured to encode and aggregate features of the historical question sequence corresponding to the current student to obtain time-aware features of the current student;

[0014] a question representation module configured to encode and aggregate features of the historical student sequence corresponding to the current question to obtain time-aware features of the current question;

[0015] a knowledge point representation module configured to encode and aggregate features of the historical knowledge point sequence corresponding to the current student and the current knowledge point to obtain time-aware features of the current knowledge point.

[0016] Based on the further improvement of the above method, the knowledge point representation module comprises:

[0017] a basic feature extraction module configured to extract basic features of the historical knowledge point sequence and basic features of the current knowledge point;

[0018] a feature encoding module configured to encode the basic features of the historical knowledge point sequence and the basic features of the current knowledge point to obtain encoded features of the historical knowledge point sequence and encoded features of the current knowledge point;

[0019] a second aggregation module configured to aggregate the encoded features of the historical knowledge point sequence and the encoded features of the current knowledge point by using an attention mechanism based on a gated time decay kernel function to obtain time-aware features of the current knowledge point.

[0020] Based on the further improvement of the above method, the second aggregation module obtains the time-aware features corresponding to the current knowledge point in the following manner:

[0021] ; ;

[0022] ;

[0023] ;

[0024] in, This represents the encoded feature at time j in the sequence of historical knowledge points. This represents the encoding features of the current knowledge point. This represents the attention mechanism. This represents the Softmax function. Indicates average pooling. This represents the Sigmoid function. , and These are learnable parameters. This represents the gated time decay kernel function, where the current time is the t-th time. This indicates the time-awareness characteristics of the current knowledge point. Let represent the context vector output by the attention mechanism at time j. The context vectors output by the attention mechanism from time 1 to time t-1 together constitute . The superscript T indicates transpose, and d indicates... Dimensions.

[0025] Based on a further improvement of the above method, the knowledge state calculation module includes a knowledge relationship calculation module and a knowledge state update module;

[0026] The knowledge relationship calculation module is used to obtain the knowledge relationship matrix of the current knowledge point based on the time-aware features of the current student and the current knowledge point.

[0027] The knowledge state update module is used to calculate the knowledge state of the current student after answering the current question based on the time-aware characteristics of the current student and the current question, as well as the knowledge relationship matrix of the current knowledge point.

[0028] Based on a further improvement of the above method, the knowledge relationship calculation module obtains the knowledge relationship matrix of the current knowledge point in the following manner:

[0029] Input the time-aware features of the current student and the current knowledge point into the multilayer perceptron to obtain the embedding features of the current knowledge point;

[0030] Obtain the encoding features of each knowledge point;

[0031] A masking matrix is ​​generated based on the current students' historical answer data;

[0032] The knowledge relationship matrix of the current knowledge point is obtained based on the embedding features of the current knowledge point, the encoding features of each knowledge point, and the masking matrix.

[0033] Based on the above method, a further improvement is made to obtain the knowledge point relationship matrix of the current knowledge point using the following formula:

[0034] ;

[0035] ;

[0036] in, This represents the embedding features of the current knowledge point. A matrix representing the encoded features of all knowledge points. The dimension representing the encoded features of each knowledge point. Represents the masking matrix. This represents batch matrix multiplication. Represents the Hadamard product. This represents the ReLU activation function. This represents the knowledge point relationship matrix for the current knowledge point. Indicates an intermediate variable.

[0037] Based on a further improvement of the above method, the knowledge state update module obtains the knowledge state of the current student after answering the current question in the following way:

[0038] Calculate the amount of knowledge change of the current student on the current knowledge point after answering the current question based on the time perception characteristics of the current student and the current question, as well as the interaction results between the current student and the current question;

[0039] Based on the knowledge point relationship matrix of the current knowledge point and the amount of knowledge change on the current knowledge point, the knowledge transfer calculation is performed to obtain the overall change in the student's knowledge state after answering the current question.

[0040] Based on the overall change in the student's knowledge state before answering the current question and after answering the current question, we obtain the student's knowledge state after answering the current question.

[0041] Based on a further improvement of the above method, the student representation module includes:

[0042] The basic feature extraction module is used to extract basic features from historical problem sequences.

[0043] The feature encoding module is used to encode the basic features of the historical problem sequence to obtain the encoded features of the historical problem sequence;

[0044] The first aggregation module is used to aggregate the encoded features of historical problem sequences using gated loop units to obtain the current student's time perception features.

[0045] Based on further improvements to the above methods, the fundamental characteristics of historical problem sequences include sequence time characteristics;

[0046] The feature encoding module obtains the encoded features of the sequence time features in the following way:

[0047] ;

[0048] ; ;

[0049] ;

[0050] in, Encoded features representing the temporal characteristics of a sequence. This represents the ReLU activation function. , , Represents the weight parameters. , , Indicates bias. This represents the time latent feature of the short-term and long-term effects at time j. The time feature at time j in the sequence time feature is represented by the time latent features of the long and short effects from time 1 to time t-1. , This represents the time threshold, where the current time is the t-th time.

[0051] Compared with existing technologies, this invention constructs a graph structure for each historical moment, extracts the temporal perception features of the current student, question, and knowledge point based on the graph structure of the historical moment, and then predicts the probability of the current student correctly answering the current question based on the temporal perception features of the current student and the current question, as well as the current student's knowledge state on the current knowledge point. After the current student answers the current question, the temporal perception features of the current student, the current question, and the current knowledge point, as well as the interaction result of the current student's answer to the current question, are input into the knowledge state calculation module. This updates the current student's knowledge state based on the knowledge point relationship matrix, simulates the changes in the knowledge point relationship network during the student's learning process, and thus determines the direction of knowledge transfer during the student's learning process. This accurately calculates the student's knowledge state and avoids the problem of incorrect knowledge transfer caused by traditional knowledge tracing modeling with static graph structures, which leads to incorrect assessment of the student's knowledge state. This achieves accurate tracking of the student's knowledge state.

[0052] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0053] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0054] Figure 1 This is a flowchart of a knowledge relationship evolution knowledge tracking method based on dynamic graph neural networks, according to an embodiment of the present invention. Detailed Implementation

[0055] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0056] A specific embodiment of the present invention discloses a knowledge relationship evolution knowledge tracking method based on dynamic graph neural networks, such as... Figure 1 As shown, it includes the following steps:

[0057] Construct a graph structure for each historical moment based on the students, questions, and corresponding knowledge points for each question;

[0058] A trained dynamic graph neural network is obtained by constructing and training the dynamic graph neural network, which includes a feature extraction module, a prediction module, and a knowledge state calculation module.

[0059] Based on the graph structure of historical moments, the feature extraction module is used to obtain the time-aware features of the current student, the current question, and the current knowledge point at the current moment.

[0060] The prediction module is input into the time-aware characteristics of the current student and the current question, as well as the current student's knowledge status on the current knowledge point, to predict the probability that the current student will answer the current question correctly.

[0061] The knowledge state calculation module inputs the time-aware features of the current student, the current question, and the current knowledge point, as well as the interaction result of the current student's answer to the current question, into the knowledge state calculation module. The knowledge state calculation module updates the current student's knowledge state based on the knowledge point relationship matrix.

[0062] It should be noted that the current knowledge point is the knowledge point corresponding to the current question.

[0063] Compared with existing technologies, the knowledge relationship evolution knowledge tracking method based on dynamic graph neural networks provided in this embodiment constructs a graph structure for each historical moment. Based on the graph structure of historical moments, it extracts the time-aware features of the current student, question, and knowledge point at the current moment. Then, based on the time-aware features of the current student and the current question, as well as the current student's knowledge state on the current knowledge point, it predicts the probability that the current student will correctly answer the current question. After the current student answers the current question, the time-aware features of the current student, the current question, and the current knowledge point, as well as the interaction result of the current student's answer to the current question, are input into the knowledge state calculation module. This updates the current student's knowledge state based on the knowledge point relationship matrix, simulates the changes in the knowledge point relationship network during the student's learning process, and thus determines the direction of knowledge transfer during the student's learning process. It accurately calculates the student's knowledge state and avoids the problem of incorrect knowledge transfer caused by traditional knowledge tracking modeling with static graph structures, which leads to incorrect assessment of the student's knowledge state. This achieves accurate tracking of the student's knowledge state.

[0064] It should be noted that after the current student answers the current question, the current student's knowledge status is updated, which is used as the input for the prediction module before the student answers the question next time.

[0065] In practice, student responses are represented as quintuples. ,in Represents the student index. Indicates the problem index. This indicates the index of the knowledge points corresponding to the question. This represents the student's answer; 1 indicates a correct answer, and 0 indicates an incorrect answer. This represents the timestamp of the interaction. A graph structure is constructed for each historical moment based on the student, question, and corresponding knowledge point. All historical moment answer data form a dynamic graph structure. This dynamic graph structure has three types of nodes: student, question, and knowledge point. When a student answers a question, an edge is created between the student and the question, and an edge is created between the answered question and its associated knowledge point. Each node has an initial feature (student nodes contain the student's index, question nodes contain both the question and knowledge point indices, and knowledge point nodes contain the knowledge point index). The s-th student node is represented as... The q-th problem node is represented as The k-th knowledge point node is represented as The initial feature dimension of the node is Each interaction edge has an initial feature (the initial feature of the edge contains information about whether the student's answer is correct). The initial feature of the interaction edge between the s-th student and the q-th question is represented as follows: , dimension .

[0066] In practice, the initial knowledge state of the s-th student is represented as a randomly initialized matrix. The knowledge state corresponding to the kth knowledge point is ,in The dimension representing the knowledge state, K represents the number of knowledge points.

[0067] The constructed dynamic graph neural network includes a feature extraction module, a prediction module, and a knowledge state calculation module. The dynamic graph neural network is trained to obtain a trained dynamic graph neural network.

[0068] The interaction edges between student nodes and question nodes can only be determined after the student answers the question. Knowledge tracking obtains the student's knowledge status based on their historical answering behavior, thereby predicting the student's future answering performance. In implementation, firstly, based on the graph structure of historical moments, a feature extraction module is used to obtain the time-aware features of the current student, the current question, and the current knowledge point at the current moment.

[0069] Specifically, the feature extraction module includes:

[0070] The sequence generation module is used to generate the historical question sequence and historical knowledge point sequence corresponding to the current student at the current moment, as well as the historical student sequence corresponding to the current question, based on the graph structure of historical moments.

[0071] The student representation module is used to perform feature encoding and feature aggregation on the historical question sequence corresponding to the current student to obtain the current student's time-aware features.

[0072] The problem representation module is used to perform feature encoding and feature aggregation on the historical student sequence corresponding to the current problem to obtain the time-aware features of the current problem;

[0073] The knowledge point representation module is used to encode and aggregate the features of the historical knowledge point sequence corresponding to the current student and the current knowledge point to obtain the time-aware features of the current knowledge point.

[0074] For the current student (e.g., the s-th student), extract all the question nodes that have interacted with this student and whose interaction timestamps are less than the current time (the t-th time) from the historical graph structure, forming the historical question sequence corresponding to this student at the current time. Generate a sequence of historical knowledge points based on the knowledge points corresponding to each question in the historical question sequence. .

[0075] For the current question (the q-th question) node, extract all student nodes that have interacted with this question from the historical graph structure, and whose interaction timestamps are less than the current time. This constitutes the historical student sequence corresponding to this question at the current time. .

[0076] The current knowledge point (the kth knowledge point) is the knowledge point corresponding to the current question (the qth question).

[0077] During implementation, a maximum sequence length can be set, for example, to 200. If the maximum sequence length is set to 200, sequences shorter than 200 are padded with 0, and sequences longer than 200 only select the closest interactive node in time, thus unifying the sequence length.

[0078] Based on the obtained sequence, time-aware features of the current student, current question, and current knowledge point are generated.

[0079] Specifically, for the current student, the time perception features of the current student are obtained by feature encoding and feature aggregation of the corresponding historical question sequence.

[0080] Specifically, the student representation module includes:

[0081] The basic feature extraction module is used to extract basic features from historical problem sequences.

[0082] The feature encoding module is used to encode the basic features of the historical problem sequence to obtain the encoded features of the historical problem sequence;

[0083] The first aggregation module is used to aggregate the encoded features of historical problem sequences using gated loop units to obtain the current student's time perception features.

[0084] Specifically, the basic characteristics of historical problem sequences include sequence node characteristics, sequence edge characteristics, sequence time characteristics, and sequence structure characteristics.

[0085] Sequence node features It is a feature sequence composed of the initial features of each problem node in the historical problem sequence.

[0086] Sequence edge features It is a feature sequence consisting of the initial features of the interaction between the current student and each problem node in the historical problem sequence.

[0087] Time features of the sequence It is a feature sequence consisting of the time interval between the timestamp of the interaction between the current student and each question node in the historical question sequence and the current moment.

[0088] Considering that current students may have answered the same questions repeatedly in the past, sequence structure features are introduced. Set the position of the repeated practice in the historical problem sequence to 1, and the others to 0. For example: the current problem is... The students' history answer sequence is as follows The corresponding sequence structure feature is (1,0,1,0).

[0089] After obtaining the sequence node features, sequence edge features, sequence time features, and sequence structure features, these sequence features are encoded to obtain the encoded features. In practice, the encoding process can use a linear function. The specific calculation process is as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] in, Encoded features representing the characteristics of sequence nodes Encoded features representing sequence edge features Encoded features representing sequence structure characteristics , , Represents the weight parameters. , , Indicates bias.

[0094] Considering the different effects produced by long and short time periods, the temporal features of the sequence are encoded using the following formula:

[0095] ;

[0096] ; ;

[0097] ;

[0098] in, Encoded features representing the temporal characteristics of a sequence. This represents the ReLU activation function. , , Represents the weight parameters. , , Indicates bias. This represents the time latent feature of the short-term and long-term effects at time j. The time feature at time j in the sequence time feature is represented by the time latent features of the long and short effects from time 1 to time t-1. , This represents the time threshold, used to separate long and short time effects, with the current time being the t-th time.

[0099] Then, gated recurrent units are used to aggregate the encoded features of the historical question sequence to obtain the current student's time-perceived features. .

[0100] During implementation, considering that students not only show sensitivity to repetitive questions during practice, but also to previously learned knowledge points, the sequence structure characteristics of historical knowledge points were incorporated. Also included in the aggregation process. Historical knowledge point sequence structure characteristics. See below.

[0101] That is, the sequence data input to the gated recurrent unit includes , , , and

[0102] The current time and the student's time perception characteristics obtained through the gated loop unit are represented as follows: .

[0103] In implementation, the problem representation module has a similar structure to the student representation module. It extracts basic features from the historical student sequence (including sequence node features, sequence edge features, sequence time features, and sequence structure features), encodes these basic features to obtain the encoded features of the historical student sequence, and aggregates these encoded features using a gated recurrent unit to obtain the time-aware features of the current problem. The difference between the problem representation module and the student representation module lies in the fact that the data input to the gated loop unit does not include the knowledge point sequence structure features. That is, the sequence data input to the gated loop unit of the problem representation module includes the encoded features of the sequence node features. Encoding features of sequence edge features Encoding features of sequence structure Encoding features of sequence time features .

[0104] Specifically, the knowledge point representation module includes:

[0105] The basic feature extraction module is used to extract the basic features of historical knowledge point sequences and the basic features of the current knowledge point.

[0106] The feature encoding module is used to encode the basic features of the historical knowledge point sequence and the basic features of the current knowledge point to obtain the encoded features of the historical knowledge point sequence and the encoded features of the current knowledge point.

[0107] The second aggregation module is used to aggregate the encoding features of historical knowledge point sequences and the encoding features of current knowledge points using an attention mechanism based on a gated time decay kernel function to obtain the time-aware features of the current knowledge point.

[0108] During implementation, the basic feature extraction module of the knowledge point representation module extracts the basic features of the historical knowledge point sequence in the same way as the basic feature extraction module of the student representation module, including sequence node features. Sequence edge features and sequence structure features The sequence edge features of the knowledge point representation module are the same as those of the student representation module.

[0109] In addition, the basic extraction module of the knowledge point representation module also extracts the basic features of the current knowledge point. The basic features of the current knowledge point include the initial features and sequence structure features. Among them, the sequence structure features are used to indicate whether the current student has learned the current knowledge point; if they have, it is 1, and if not, it is 0.

[0110] The feature encoding module of the knowledge point representation module also uses linear functions to encode the basic features:

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] Encoded features representing the characteristics of sequence nodes Encoded features representing sequence edge features Encoded features representing sequence structure characteristics , , and Represents the weight parameters. , , and Indicates bias. This represents the encoding features of the current knowledge point. This represents the initial characteristics of the current knowledge point. This represents the sequence structure features of the current knowledge point. The encoded features representing the initial features of the current knowledge point. Encoding features that represent the sequence structure features of the current knowledge point.

[0119] It should be noted that the sequence structure features of the current knowledge point include information on whether the current student has studied the current knowledge point.

[0120] Encoded features representing the characteristics of sequence nodes Encoded features representing sequence edge features The encoding features representing the structural characteristics of a sequence constitute the encoding features of a sequence of historical knowledge points. .

[0121] After obtaining the encoded features, an attention mechanism based on the gated time decay kernel function is used to obtain the time-aware features of the current knowledge point.

[0122] Specifically, the encoded features of the current knowledge point are used as the query in the attention mechanism, and the encoded features of each moment in the historical knowledge point sequence are used as the key and value of the attention mechanism. To enhance the influence of the local time window on knowledge points, this application introduces a gated time decay kernel function into the attention mechanism.

[0123] Specifically, the second aggregation module obtains the time-aware features corresponding to the knowledge points in the following way:

[0124] ; ;

[0125] ; ;

[0126] in, This represents the encoded feature at time j in the sequence of historical knowledge points. This represents the encoding features of the current knowledge point. This represents the attention mechanism. This represents the Softmax function. Indicates average pooling. This represents the Sigmoid function. , and These are learnable parameters. This represents the gated time decay kernel function, where the current time is the t-th time. This indicates the time-awareness characteristics of the current knowledge point. Let represent the context vector output by the attention mechanism at time j. The context vectors output by the attention mechanism from time 1 to time t-1 together constitute . d represents Dimensions.

[0127] Encoding characteristics of historical knowledge point sequences In the example, the encoding feature at time j is: .

[0128] Among them, parameters , and These are used to control the steepness of the function's descent, determine the time boundary of the local window, and adjust the decay capability of time, respectively.

[0129] By introducing a gated time decay kernel function, the learning pattern of knowledge points being more likely to be generated within a local short time window during the student's learning process is simulated, thereby improving the accuracy of knowledge relationship construction.

[0130] Specifically, the knowledge state calculation module includes a knowledge relationship calculation module and a knowledge state update module; the knowledge relationship calculation module is used to obtain the knowledge relationship matrix of the current knowledge point based on the time-aware features of the current student and the current knowledge point;

[0131] The knowledge status update module is used to calculate the knowledge status of the current student after answering the current question, based on the time-aware characteristics of the current student and the current question, as well as the knowledge relationship matrix of the current knowledge point.

[0132] To achieve more accurate knowledge tracking, it's necessary to calculate the impact of a student's change in the current knowledge point's state on other knowledge points, i.e., the knowledge relationship matrix of the current knowledge point. This allows for the construction of a personalized knowledge point relationship network for each student, modeling the changes in these relationships after each question. Based on these knowledge point relationships, the knowledge gained after each question is mapped to a change in the overall knowledge state.

[0133] Specifically, the knowledge relationship calculation module obtains the knowledge relationship matrix of the current knowledge point in the following way:

[0134] Input the time-aware features of the current student and the current knowledge point into the multilayer perceptron to obtain the embedding features of the current knowledge point;

[0135] Obtain the encoding features of each knowledge point;

[0136] A masking matrix is ​​generated based on the current students' historical answer data;

[0137] The knowledge relationship matrix of the current knowledge point is obtained based on the embedding features of the current knowledge point, the encoding features of each knowledge point, and the masking matrix.

[0138] Intuitively, students' recent learning status influences their understanding of knowledge points. Therefore, the temporal perception features of knowledge points are added to the temporal perception features of student nodes, and the resulting data is processed through a multilayer perceptron to obtain the embedding features of the knowledge concepts. The specific calculation process is as follows:

[0139] ;

[0140] in, This indicates the embedded features of the current knowledge point. This indicates the current time perception characteristics of students. This indicates the time-awareness characteristics of the current knowledge point. and The weight parameters represent the multilayer perceptron. and This represents the bias of the multilayer perceptron. This represents the Tanh activation function.

[0141] During implementation, the initial features of all knowledge points are encoded to obtain the encoded features, represented as follows: In implementation, a linear encoder can be used to encode the initial features of all knowledge points. This can be represented as:

[0142] ;

[0143] in, This represents the initial feature matrix of all knowledge points. The encoding feature matrix representing all knowledge points. and These represent the weight parameters and bias of the linear encoder, respectively.

[0144] A scaled dot product is performed on the embedded features of the current knowledge point and the encoded features of all knowledge points to represent the association between the current knowledge point and other knowledge points. This is expressed as:

[0145] ;

[0146] in, This represents the embedding features of the current knowledge point. This represents batch matrix multiplication. The dimension representing the encoded features of each knowledge point. This represents the ReLU activation function. It indicates the relationship between the current knowledge point and other knowledge points.

[0147] Meanwhile, to mask the connections between knowledge points that students have not yet encountered, a masking matrix is ​​generated by removing duplicate knowledge points from the students' historical practice sequences. This function operates on the relationships between the current knowledge point and other knowledge points, outputting a personalized knowledge point relationship matrix for the student at that time step. This indicates the number of knowledge points. It is represented as:

[0148] ;

[0149] in, Represents the Hadamard product. This represents the knowledge point relationship matrix for the current knowledge point.

[0150] After calculating the personalized knowledge point relationship matrix at the current moment, the knowledge state of the current student after answering the current question is obtained based on the time perception characteristics of the current student and the current question, as well as the knowledge relationship matrix of the current knowledge points.

[0151] Specifically, the knowledge status update module obtains the current student's knowledge status after answering the current question in the following way:

[0152] Calculate the amount of knowledge change of the current student on the current knowledge point after answering the current question based on the time perception characteristics of the current student and the current question, as well as the interaction results between the current student and the current question;

[0153] Based on the knowledge point relationship matrix of the current knowledge point and the amount of knowledge change on the current knowledge point, the knowledge transfer calculation is performed to obtain the overall change in the student's knowledge state after answering the current question.

[0154] Based on the overall change in the student's knowledge state before answering the current question and after answering the current question, we obtain the student's knowledge state after answering the current question.

[0155] Specifically, the knowledge status calculation module calculates the change in a student's knowledge of the current knowledge point after answering the current question using the following method:

[0156] ;

[0157] ;

[0158] ;

[0159] ;

[0160] in, This indicates the amount of knowledge change a student has made regarding the current knowledge point. Represents the embedding features of the current student. Represents the embedding features of the current problem. This represents the initial characteristics of the interaction edge between the current student and the current question (including information on whether the student answered the question correctly). The encoded features representing the interaction edges between the current student and the current question. , , and Represents the weight parameters. , , and Indicates bias.

[0161] The knowledge students acquire through practice not only changes their mastery of the current knowledge point but also transfers to related knowledge concepts. This knowledge transfer path is determined by the student's personalized dynamic knowledge relationship matrix. Therefore, based on the knowledge point relationship matrix of the current knowledge point and the amount of knowledge change at the current knowledge point, knowledge transfer calculations can be performed to obtain the amount of knowledge change the student has across all knowledge points after answering the current question.

[0162] Specifically, it is expressed as follows:

[0163] ;

[0164] Obtain the overall change in knowledge state Then, by adding this information to the student's knowledge state before answering the current question, we can obtain the student's knowledge state after answering the current question, thus enabling precise tracking of the student's knowledge state.

[0165] During implementation, based on students' knowledge status, it can be found which knowledge points students are not proficient in, and then targeted exercises can be given to improve students' learning efficiency and state.

[0166] In practice, in order to obtain a well-trained dynamic graph neural network and achieve accurate knowledge state assessment and tracking, the dynamic graph neural network also includes a prediction module, which is used to predict the probability of a student correctly answering the current question based on the time-aware characteristics of the student and the question, as well as the student's knowledge state before answering the question.

[0167] Specifically, the prediction module includes:

[0168] The encoding layer is used to encode the time-aware features of students and questions respectively to obtain the embedded features of students and questions;

[0169] Multilayer perceptrons are used to predict the probability that a student will answer the current question correctly, based on the embedded features of the student and the question, as well as the student's knowledge state of the current knowledge point before answering the current question.

[0170] In implementation, the encoding layer can use a linear encoder to encode the time-aware features of the current student and the current question to obtain the embedded features of the student and the question. , .

[0171] Since knowledge status can accurately reflect students' mastery of knowledge concepts, the prediction process incorporates students' knowledge status of the knowledge points involved in the question before answering it. The data is then concatenated with the embedded features of the student and question nodes as input to the multilayer perceptron, and the output is the predicted probability of the student answering the question correctly. The specific calculation process is as follows:

[0172] ;

[0173] ;

[0174] in, This indicates a splicing operation. This represents the Sigmoid function. This represents the predicted probability that the current student has correctly answered the current question, as output by the prediction module. and Represents the weight parameters. and Indicates bias.

[0175] In practice, when training the constructed dynamic graph neural network, multiple students' answer data can be collected to construct a training sample set. The dynamic graph neural network is then trained based on this training sample set. During implementation, the parameters of the dynamic graph neural network are updated by minimizing the cross-entropy loss between the prediction probability of the prediction module and the student's actual answer. After training converges, the trained dynamic graph neural network is obtained.

[0176] In implementation, the constructed training sample set includes each student's answer data over a period of time, constructing a graph structure corresponding to each time step, with each training sample represented as a quintuple. For each sample, the feature extraction module generates the historical question sequence and historical knowledge point sequence corresponding to the student node, as well as the historical student sequence corresponding to the question node, based on the graph structure of that historical time step. This yields the time-aware features of the student, question, and knowledge point. The prediction module then predicts the probability of the student answering the question correctly, and the knowledge state calculation module updates the student's knowledge state after the student answers. The loss is calculated based on the prediction results and the actual answer results, and the parameters of the dynamic graph neural network are updated based on the loss.

[0177] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0178] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A knowledge relation evolution knowledge tracking method based on a dynamic graph neural network, characterized in that, The method comprises the following steps: constructing a graph structure of each historical moment based on students, questions and knowledge points corresponding to the questions of each historical moment; constructing and training a dynamic graph neural network to obtain a trained dynamic graph neural network, the trained dynamic graph neural network comprising a feature extraction module, a prediction module and a knowledge state calculation module; obtaining time-aware features of a current student, a current question and a current knowledge point at a current moment based on the graph structure of the historical moment by using the feature extraction module; inputting the time-aware features of the current student and the current question and a knowledge state of the current student on the current knowledge point into the prediction module to predict a probability of the current student correctly answering the current question; inputting the time-aware features of the current student, the current question and the current knowledge point and an interaction result of the current student answering the current question into the knowledge state calculation module, the knowledge state calculation module updating the knowledge state of the current student based on a knowledge point relationship matrix; the knowledge state calculation module comprises a knowledge relationship calculation module and a knowledge state updating module; the knowledge relationship calculation module is configured to obtain the knowledge point relationship matrix of the current knowledge point based on the time-aware features of the current student and the current knowledge point; the knowledge state updating module is configured to calculate the knowledge state of the current student after answering the current question based on the time-aware features of the current student and the current question and the knowledge point relationship matrix of the current knowledge point; the knowledge relationship calculation module obtains the knowledge point relationship matrix of the current knowledge point in the following manner: inputting the time-aware features of the current student and the current knowledge point into a multi-layer perception machine to obtain embedding features of the current knowledge point; obtaining encoding features of each knowledge point; generating a mask matrix based on historical answering data of the current student; obtaining the knowledge point relationship matrix of the current knowledge point based on the embedding features of the current knowledge point, the encoding features of each knowledge point and the mask matrix; the knowledge state updating module obtains the knowledge state of the current student after answering the current question in the following manner: calculating a knowledge change amount of the current student on the current knowledge point after answering the current question according to the time-aware features of the current student and the current question and the interaction result of the current student and the current question; performing knowledge transfer calculation based on the knowledge point relationship matrix of the current knowledge point and the knowledge change amount on the current knowledge point to obtain a total change amount of the knowledge state of the current student after answering the current question; obtaining the knowledge state of the current student after answering the current question based on the knowledge state of the current student before answering the current question and the total change amount of the knowledge state after answering the current question.

2. The knowledge relation evolution knowledge tracking method based on a dynamic graph neural network according to claim 1, characterized in that, the feature extraction module comprises: a sequence generation module configured to generate a historical question sequence and a historical knowledge point sequence corresponding to the current student at the current moment and a historical student sequence corresponding to the current question based on the graph structure of the historical moment; a student representation module configured to perform feature encoding and feature aggregation on the historical question sequence corresponding to the current student to obtain time-aware features of the current student; a question representation module configured to perform feature encoding and feature aggregation on the historical student sequence corresponding to the current question to obtain time-aware features of the current question; The knowledge point representation module is configured to encode and aggregate features of a historical knowledge point sequence corresponding to the current student and the current knowledge point to obtain time-aware features of the current knowledge point.

3. The knowledge relation evolution knowledge tracking method based on a dynamic graph neural network according to claim 2, characterized in that, The knowledge point representation module comprises: a basic feature extraction module configured to extract basic features of the historical knowledge point sequence and basic features of the current knowledge point; a feature encoding module configured to encode the basic features of the historical knowledge point sequence and the basic features of the current knowledge point to obtain encoded features of the historical knowledge point sequence and encoded features of the current knowledge point; a second aggregation module configured to aggregate the encoded features of the historical knowledge point sequence and the encoded features of the current knowledge point by using an attention mechanism based on a gated time decay kernel function to obtain the time-aware features of the current knowledge point.

4. The knowledge relation evolution knowledge tracking method based on a dynamic graph neural network according to claim 3, characterized in that, The second aggregation module obtains the time-aware features corresponding to the current knowledge point in the following manner: ； ； ； ； wherein, denotes the encoding feature of the jth time point in the historical knowledge point sequence, denotes the encoding feature of the current knowledge point, denotes the attention mechanism, denotes the Softmax function, denotes the average pooling, denotes the Sigmoid function, , and is a learnable parameter, denotes the gating time decay kernel function, the current time point is the tth time point, denotes the time-aware feature of the current knowledge point, denotes the context vector output by the attention mechanism at the jth time point, the context vectors output by the attention mechanisms at the 1st to the t-1th time points jointly constitute , the superscript T denotes transposition, d denotes the dimension of 5. The knowledge relation evolution knowledge tracking method based on a dynamic graph neural network according to claim 1, characterized in that, The knowledge point relationship matrix of the current knowledge point is obtained in the following formula: ； ； wherein, represents an embedding feature of the current knowledge point, represents a matrix of encoding features of all knowledge points, represents a dimension of the encoding feature of each knowledge point, represents a masking matrix, represents a batch matrix multiplication, represents a Hadamard product, represents a Relu activation function, represents a knowledge point relationship matrix of the current knowledge point, represents an intermediate variable.

6. The knowledge relation evolution knowledge tracking method based on a dynamic graph neural network according to claim 2, characterized in that, The student representation module comprises: a basic feature extraction module configured to extract basic features of the historical question sequence; a feature encoding module configured to encode the basic features of the historical question sequence to obtain encoded features of the historical question sequence; a first aggregation module configured to aggregate the encoded features of the historical question sequence by using a gated recurrent unit to obtain time-aware features of the current student.

7. The knowledge relation evolution knowledge tracking method based on a dynamic graph neural network according to claim 6, characterized in that, The basic features of the historical question sequence comprise sequence time features; The feature encoding module obtains the encoded features of the sequence time features in the following manner: ； ； ； ； wherein, denotes an encoding feature representing sequence time features, denotes a Relu activation function, denotes a weight parameter, denotes a bias, denotes a long-short effect time hidden feature at the jth time point, denotes a time feature at the jth time point in the sequence time features, and the long-short effect time hidden features at the 1st to the t-1th time points collectively constitute , denotes a time threshold, and the current time point is the tth time point.​​​​

Images