Deep knowledge tracing method based on graph neural networks

Abstract

This invention discloses a deep knowledge tracking method based on graph neural networks, belonging to the field of smart education. Specifically, the method involves: first, cleaning the real dataset and storing the binary relationship between questions and skills in each interaction data point in a dictionary. Then, for the same student, all their questions are sequenced in ascending time steps, and corresponding exercise sequence graphs and interaction sequence graphs are generated based on the binary relationships. Next, a corresponding graph neural network is constructed to obtain the representation of each graph in the exercise sequence graph and interaction sequence graph. The exercise graph representation at time step t+1 is concatenated with the interaction graph representation at time step t, and the student's learning result at time step t+1 is predicted using a fully connected neural network. Finally, the graph neural network is embedded into an existing Transformer model, and parameters are updated through backpropagation. This invention achieves dynamic graph representation of questions and interaction sequences and utilizes an attention mechanism to update the knowledge state, improving the model's performance and interpretability.

Description

Technical Field

[0001] This invention belongs to the field of smart education, specifically relating to a deep knowledge tracking method based on graph neural networks. Background Technology

[0002] In the context of the information age, the emergence of online education has changed the way students learn, and traditional education models have been challenged. Diverse and personalized learning methods have gradually replaced traditional teaching models; research on knowledge tracking technology has become a key means to achieve personalized learning.

[0003] The Knowledge Tracking Task aims to assess students' abilities and predict future performance by analyzing data on their interactions with educational resources. Online education platforms can then use this data to provide customized learning resources and suggestions, thereby improving learning outcomes.

[0004] In real-world learning environments, students' answers exhibit significant sparsity, especially in online learning where platforms are filled with numerous exercises, but most of these exercises involve only a limited number of interactions. Furthermore, the input to knowledge tracking models often uses skills as the sole identifier, lacking relevant question information, which limits the model's performance. Summary of the Invention

[0005] To address the limitations of existing deep knowledge tracing methods, such as missing question information and performance constraints, this invention proposes a deep knowledge tracing method based on graph neural networks. This method achieves dynamic graph representation of question sequences and interaction sequences, updates knowledge states using an attention mechanism, and improves the performance and interpretability of the model.

[0006] The deep knowledge tracing method based on graph neural networks includes the following steps:

[0007] Step 1: Clean the publicly available real dataset, retaining four types of data: problem data, skill data, interaction data between the two, and student data. Delete all data with empty values ​​and number the remaining data after deletion.

[0008] If a question and a skill appear in an interaction, it means that the question includes the skill.

[0009] The extracted student IDs, question identifiers, and skill identifiers are numbered starting from 0 to obtain new numerical identifiers.

[0010] Step 2: Based on the question number and skill number, store the binary relationship between the question and skill in each piece of interactive data into a dictionary.

[0011] For the interactive data identified as d1, which contains a question q1 and one or more skills d1 = (q1, {s1, s2, ..., s...}k}), the interactive data, identified as d2, contains the question q1 and one or more skills d2=(q1,{s k+1 ,s k+2 ,...,s h If}), then the dictionary storage representation is: QS[q1]={s1,s2,...,s k ,s k+1 ,...,s h};

[0012] Where q1 is a key in the dictionary, and the skill list {s1,s2,...,s...} is... k ,s k+1 ,...,s h} represents the value corresponding to the key in the dictionary.

[0013] Step 3: Based on the student ID, assemble all the questions for the same student at different time steps into a sequence in ascending order of time.

[0014] Each student has one question at a time step; the length of the selected time step for each sequence is l.

[0015] The sequence includes the sequence length, question number, skill number, and answer result.

[0016] Step 4: For student A, construct l dynamic problem and skill graph data for each of their sequences based on the binary relations stored in the dictionary, and generate corresponding exercise sequence graphs and interaction sequence graphs.

[0017] The process of generating Student A's exercise sequence diagram and interaction sequence diagram is as follows:

[0018] First, for l time steps, the interaction between student A and each exercise is obtained according to the exercise index, and then combined with the exercise index binary to form an interaction sequence;

[0019] Then, based on the interaction sequence and the dictionary, a sequence graph of exercises is constructed. This graph contains two types of nodes: questions and skills. The edges between two nodes, a question and a skill, contain two attributes: the number of correct answers and the number of incorrect answers for the current question.

[0020] Finally, based on the exercise sequence graph, the exercise sequence graphs of all previous time steps of the student are superimposed, and a third attribute, whether the current answer is correct or not, is added to the edges between the question-skill nodes to form an interaction sequence graph.

[0021] Step 5: Using the problem sequence graphs and interaction sequence graphs of student A at all time steps, construct the corresponding graph neural networks using graph convolution methods to obtain the representation of each graph in the problem sequence graphs and interaction sequence graphs respectively;

[0022] The architectures of the two graph convolutional neural networks are as follows:

[0023] Initial node embedding:

[0024] h (0) =X

[0025] Aggregate neighbors through K-layer graph convolution:

[0026]

[0027] Pooling is used to obtain the global representation of the graph:

[0028]

[0029] The final graph representation is obtained through a fully connected layer:

[0030] h graph =σ(w fc h graph +b fc )

[0031] Furthermore, different graph neural networks can be selected for representation based on the dataset.

[0032] Where X represents the node features of the graph; h (0) This represents the feature vector of the first layer; It is the feature vector of node v after k aggregation iterations, which is the feature vector of the kth layer; It is a neighboring node of node v. It is a neighboring node of node u; is the feature vector of node u in layer k-1; {v} is the set of nodes; σ is the non-linear activation function; W (k) These are the weights of the k-th layer; h is the feature vector of node i in layer k; graph It is the global representation of the graph after pooling; w fc and b fc These are the weight vector and bias vector of the fully connected layer, respectively. N represents the total number of nodes in the exercise sequence graph / interaction sequence graph; the total number of nodes in the two graphs is different.

[0033] Step 6: Concatenate the problem graph representation at time step t+1 with the interaction graph representation at time step t, and use a fully connected neural network to predict the student's learning performance at time step A t+1, thus evaluating the predictive performance of the graph neural network.

[0034] The specific steps are as follows:

[0035] Step 601: Represent the question q at the current time step t. t Interaction graph representation h with the previous time step t-1By concatenating the vectors, a new vector z is obtained. t ;

[0036] As shown in the formula below:

[0037]

[0038] Step 602, convert vector z t The input is fed into the first linear layer, and after applying the ReLU activation function and Dropout regularization in sequence, the result is fed into the second linear layer, and the output result is a2:

[0039] a1=W1z t +b1

[0040] h1 = ReLU(a1)

[0041] h′1=Dropout(h1,p)

[0042] a2=W2h′1+b2

[0043] W1 and W2 are the weight matrices of the first and second linear layers, respectively; b1 and b2 are the bias vectors of the first and second linear layers, respectively; and p is the probability of each neuron being dropped in Dropout regularization.

[0044] Step 603: Input the result a2 into the Sigmoid activation function to obtain the predicted value y:

[0045]

[0046] Step 604: Use AUC, ACC, MAE, and RMSE to evaluate the predicted value y.

[0047] Step 7: Embed the graph neural network into the existing Transformer model, and use the representations of the dynamic question sequence graph and the interaction sequence graph as inputs. Train and predict whether the student answers correctly at each time step, and update the graph neural network parameters and the Transformer model parameters through backpropagation.

[0048] The present invention has the following technical effects:

[0049] 1. The present invention is a deep knowledge tracking method based on graph neural networks, which constructs a dynamic graph structure sequence from the question sequence and interaction sequence of students' answers, effectively solving the problem of data scarcity of students' answers in online platforms.

[0050] 2. The present invention is a deep knowledge tracking method based on graph neural networks. The dynamic graph sequence can be directly used to predict students' answer results through graph neural networks and fully connected neural networks, and has high performance.

[0051] 3. The present invention is a deep knowledge tracing method based on graph neural networks. It embeds graph neural networks into the input of existing knowledge tracing models, inputs dynamic graph sequence data, and trains the model, which greatly improves the performance of the model. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the deep knowledge tracing method based on graph neural networks of the present invention.

[0053] Figure 2 This is a schematic diagram of the deep knowledge tracing method based on graph neural networks of the present invention;

[0054] Figure 3 This is the exercise sequence graph structure constructed in the embodiments of the present invention;

[0055] Figure 4 This is the interactive sequence graph structure constructed in the embodiments of the present invention;

[0056] Figure 5 This is a schematic diagram of the neural network representation sequence flow according to an embodiment of the present invention. Detailed Implementation

[0057] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some, not all, embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0058] This invention proposes a deep knowledge tracing method based on graph neural networks. During the interaction between students and exercises, each question is related to a skill; one question may correspond to multiple skills, or multiple questions may correspond to one skill. By constructing a dynamic graph structure sequence from the student's questions and interaction sequences, and utilizing graph neural networks to capture the relationship between questions and skills, accurate question graph representations and interaction graph representations are obtained. These graph representations are concatenated and input into a fully connected layer to directly predict the student's answers. Furthermore, graph neural networks can be embedded into the input layer of existing knowledge tracing models, using the input graph sequence for prediction. Practice has shown that inputting dynamic graph sequences into graph neural networks and fully connected layers for prediction yields high prediction performance; embedding graph neural networks into existing knowledge tracing models and using dynamic graph structure sequences as input can also improve the performance of existing deep knowledge tracing methods.

[0059] The deep knowledge tracing method based on graph neural networks, such as Figure 1 As shown, it includes the following steps:

[0060] Step 1: Clean the publicly available real dataset, filter out the interactive data with key information on questions and skills, retain four types of data: question data, skill data, interactive data, and student data, delete all data with empty necessary values, and number the remaining data after deletion.

[0061] If a question and a skill appear in an interaction, it means that the question includes the skill.

[0062] The extracted student IDs, question identifiers, and skill identifiers are numbered starting from 0 to obtain new numerical identifiers.

[0063] Step 2: Based on the question number and skill number, store the binary relationship between the question and skill in each piece of interactive data into a dictionary.

[0064] If a question and a skill appear in an interaction, it means that the question contains the skill, meaning the skill helps solve the problem. The binary relationships between questions and skills in all interaction data are summarized and stored in a dictionary, where the key is the question number and the value is a list of skill numbers. Specifically:

[0065] For the interactive data identified as d1, which contains a question q1 and one or more skills d1 = (q1, {s1, s2, ..., s...} k}), the interactive data, identified as d2, contains the question q1 and one or more skills d2=(q1,{s k+1 ,s k+2 ,...,s h If}), then the dictionary storage representation is: QS[q1]={s1,s2,...,s k ,s k+1 ,...,s h};

[0066] Where q1 is a key in the dictionary, and the skill list {s1,s2,...,s...} is... k ,s k+1 ,...,s h} represents the value corresponding to the key in the dictionary.

[0067] Step 3: Based on the student ID, assemble all the questions for the same student at different time steps into a sequence in ascending order of time.

[0068] Each student has one question at a time step; the length of the selected time step for each sequence is l.

[0069] Each sequence consists of four elements: sequence length, question number, skill number, and answer result.

[0070] Step 4: For student A, construct dynamic graph data of questions and skills for each time step based on the binary relations stored in the dictionary, and generate corresponding exercise sequence graphs and interaction sequence graphs.

[0071] Both the exercise sequence graph and the interaction sequence graph contain questions and skill nodes. The graph for each time step in the exercise sequence graph is an undirected graph that connects the question at the current time step with the skills it covers. The edges connecting the questions and skill nodes are filled with the attributes of the number of incorrect answers and the number of correct answers for the current question.

[0072] For each time step in the interaction sequence graph, it is an undirected graph of all the questions involved in the current time step and the skills they cover. The edges connecting the questions and the corresponding skill nodes are filled with the attributes of the number of correct answers, the number of incorrect answers, and the correct / incorrect answer result of the last time for the current question.

[0073] The process of generating Student A's exercise sequence diagram and interaction sequence diagram is as follows:

[0074] First, for l time steps, the interaction between student A and each exercise is obtained according to the exercise index, and then combined with the exercise index binary to form an interaction sequence;

[0075] Then, based on the interaction sequence and the dictionary, a sequence graph of exercises is constructed. This graph contains two types of nodes: questions and skills. The edges between two nodes, a question and a skill, contain two attributes: the number of correct answers and the number of incorrect answers for the current question.

[0076] Finally, based on the exercise sequence graph, the exercise sequence graphs of all previous time steps of the student are superimposed, and a third attribute, whether the current answer is correct or not, is added to the edges between the question-skill nodes to form an interaction sequence graph.

[0077] Step 5: Using the problem sequence graphs and interaction sequence graphs of student A at all time steps, construct the corresponding graph neural networks using graph convolution methods to obtain the representation of each graph in the problem sequence graphs and interaction sequence graphs respectively;

[0078] For the exercise sequence graph and the interaction sequence graph, two graph neural network architectures are constructed to represent them respectively. The architectures of the two graph convolutional neural networks are as follows:

[0079] Initial node embedding:

[0080] h (0) =X

[0081] Aggregate neighbors through K-layer graph convolution:

[0082]

[0083] Pooling is used to obtain the global representation of the graph:

[0084]

[0085] The final graph representation is obtained through a fully connected layer:

[0086] h graph =σ(w fc h graph +b fc )

[0087] Furthermore, different graph neural networks can be selected for representation based on the dataset.

[0088] Where X represents the node features of the graph; h (0) This represents the feature vector of the first layer; It is the feature vector of node v after k aggregation iterations, which is the feature vector of the kth layer; It is a neighboring node of node v. It is a neighboring node of node u; is the feature vector of node u in layer k-1; {v} is the set of nodes; σ is the non-linear activation function; W (k) These are the weights of the k-th layer; h is the feature vector of node i in layer k; graph It is the global representation of the graph after pooling; w fc and b fc These are the weight vector and bias vector of the fully connected layer, respectively. N represents the total number of nodes in the exercise sequence graph / interaction sequence graph; the total number of nodes in the two graphs is different.

[0089] Step 6: Concatenate the problem graph representation at time step t+1 with the interaction graph representation at time step t, and use a fully connected neural network to predict the student's learning performance at time step A t+1, thus evaluating the predictive performance of the graph neural network.

[0090] The specific steps are as follows:

[0091] Step 601: Represent the question q at the current time step t. t Interaction graph representation h with the previous time step t-1 By concatenating the vectors, a new vector z is obtained. t ;

[0092] As shown in the formula below:

[0093]

[0094] Step 602, convert vector z t The input is fed into the first linear layer, and after applying the ReLU activation function and Dropout regularization in sequence, the result is fed into the second linear layer, and the output result is a2:

[0095] a1=W1z t +b1

[0096] h1 = ReLU(a1)

[0097] h′1=Dropout(h1,p)

[0098] a2=W2h′1+b2

[0099] W1 and W2 are the weight matrices of the first and second linear layers, respectively; b1 and b2 are the bias vectors of the first and second linear layers, respectively; and p is the probability of each neuron being dropped in Dropout regularization.

[0100] Step 603: Input the result a2 into the Sigmoid activation function to obtain the predicted value y:

[0101]

[0102] Step 604: Use AUC, ACC, MAE, and RMSE to evaluate the predicted value y.

[0103] Step 7: Embed the graph neural network into the existing knowledge tracking and prediction Transformer model, and use the representations of the dynamic question sequence graph and the interaction sequence graph as inputs. Train and predict whether the student answers correctly at each time step, and update the graph neural network parameters and the Transformer model parameters through backpropagation.

[0104] Generally, in knowledge tracing tasks, the input is the student's interaction sequence X. T =(x1,x2,...,x T ), x T It is usually a tuple (e t ,r t ), e t r represents the student's exercise index at time step t. t Represents students and exercises e t The interaction is correct; 1 for correct and 0 for incorrect.

[0105] Traditional deep learning knowledge tracing models take a sequence of questions and a sequence of interactions as input. Assume there are a total of N questions, Q. N =(q1,q2,…,q NIf the question sequence is processed into an N-dimensional one-hot encoding, and the interaction sequence is processed into a 2N-dimensional one-hot encoding, then the input one-hot vector will be set to 1 at index i. For the interaction sequence, in order to distinguish between correct and incorrect answers, the dimension will be expanded to 2N. Specifically, when a student answers a question correctly, the vector at index i will be set to 1, and the rest will be set to 0. When a student answers a question incorrectly, the vector at index i+N will be set to 1, and the rest will be set to 0.

[0106] In this invention, the one-hot encoding in the original model is replaced by a graph neural network, and the input of an existing knowledge tracing prediction model is embedded. The dynamic question sequence graph and the interaction sequence graph are used as the input of the model, and the student's answer is predicted at each time. The graph neural network parameters and the knowledge tracing prediction model parameters are updated through backpropagation.

[0107] All parameters were trained using the standard cross-entropy loss function:

[0108]

[0109] Where, r t p represents the correctness of the student's answer to the question. t This represents the probability that the model predicts a correct answer to a question. This calculation method trains the model by minimizing the difference between the two.

[0110] Example:

[0111] This embodiment is applied to knowledge tracking tasks within a smart education system environment. In knowledge tracking tasks within the smart education field, this method can utilize graph neural networks to mine higher-order information between questions and skills, fully representing question vectors and student interaction vectors. It overcomes the limitation of current knowledge tracking methods, which rely on different skill numbers as unique identifiers for student answers due to data scarcity. Figure 2 The diagram illustrates the overall workflow of this invention. On one hand, the dynamic sequence graph constructed using this method, processed by graph neural networks and fully connected neural networks, can directly predict students' answers, achieving high performance. On the other hand, embedding graph neural networks into existing knowledge tracking models, and inputting the dynamic graph sequence constructed using this method for training and prediction, significantly improves the performance of existing models.

[0112] The specific steps are as follows:

[0113] Step 1: Perform data cleaning on four publicly available real datasets: the Assist2009 and Assist2017 datasets from the online tutoring platform Assistments, the Ednet dataset from the English education mobile application Santa, and the Junyi dataset from a Taiwanese online platform. Delete interactive data that is missing key information such as student identifiers, answer times, question identifiers, and skill identifiers, and filter out interactive data with key information about questions and skills. Then, number the student identifiers, question identifiers, and skill identifiers starting from 0 to obtain new numerical identifiers.

[0114] 5000 students and their interactions are randomly selected from a large dataset. All items with empty values ​​are deleted. The selected student information and question information, including ID, question identifier, and skill identifier, are numbered starting from 0 to obtain new numerical identifiers.

[0115] Step 2: Obtain the binary relationship between questions and skills from each piece of interaction data.

[0116] For each piece of data, there is an inclusion relationship between the questions and corresponding skills in that data. Extract the questions and skills from all the interaction data and store them in a dictionary;

[0117] Step 3: Group the data obtained in Step 1 and Step 2 according to student ID, and divide them into sequence data according to the increasing time order.

[0118] The data obtained in step 1 is grouped according to student ID and divided into answer sequence data of length T in ascending time order. If the number of interactions of a student is less than the length T, it is padded with zeros, and the actual length of the sequence before padding is recorded. If the number of interactions of a student is greater than the length T, it is pruned to generate multiple sequences.

[0119] Step 4: Based on the relationships obtained in Step 2, construct dynamic graph data of questions and skills for each student's time step using the sequence data, and generate exercise sequence graphs and interaction sequence graphs.

[0120] For each time step, a knowledge state graph is constructed that includes the question and answer structures of all previous interactions. This knowledge state graph structure will be different at each time step and can serve as a dynamic representation of the student's knowledge state.

[0121] The problem sequence graph represents the association between problems and skills as an undirected graph, where nodes represent problems and skills, and edges represent the relationships between them. Each problem q and its corresponding skill s are constructed into a graph, serving as the input for predicting the problem. If a problem involves multiple skills, then there are edges connecting the problem node to each of these skill nodes.

[0122] In the interaction sequence diagram, a knowledge state diagram containing the question and answer structures of all previous interactions is constructed for each time step. This knowledge state diagram structure will be different at each time step and can serve as a dynamic representation of the student's knowledge state.

[0123] For each student's answer sequence of length T, given N questions and L skills:

[0124] Problem set Q N =(q1,q2,…,q N ) and skill set S L =(s1,s2,…,s L ); and the dictionary mapping from problem to skill: QS = {q1:[s1],q2:[s1,s2],…,q N :[s l-2 ,s l ]).

[0125] According to exercise sequence E T =(e1,e2,…,e T For each time step, a graph is constructed to represent the relationship between questions and skills. The association between questions and skills is represented as an undirected graph, where nodes represent questions and skills, and edges represent the relationships between them. Each question q and its corresponding skill s are constructed into a graph, which serves as the input for predicting the question. If a question involves multiple skills, then there are edges connecting the question node to these skill nodes.

[0126] Specifically: For each time step, construct a graph structure connecting the current time step's problem and its corresponding skill. Problems and skills are linked, and additional information (the number of incorrect and correct answers to the current problem) can be added to the edges. Assume that the problem corresponding to exercise e1 in time step 1 is q1, the problem corresponding to exercise e2 in time step 2 is q2, and the problem corresponding to exercise e3 in time step 3 is still q2; the problem-to-skill mapping relationship QS = {q1:[s1],q2:[s1,s2],q3:[s1,s2]}. Then, a graph structure can be constructed as follows: Figure 3 The exercise sequence diagram shown has blue nodes representing skills and yellow nodes representing questions. Each node has two attributes: the first attribute is the number of times the current question was answered incorrectly, and the second attribute is the number of times it was answered correctly. In exercise e3, question q2 was answered correctly in the previous question, so the first attribute is 1 and the second attribute is 0.

[0127] A graph relating questions and skills up to the current time step is constructed, connecting questions to their associated skills and adding rich attribute information to the edges, such as: number of correct answers, number of incorrect answers, and the result of the correct or incorrect answer. The knowledge state graph at time step t is an update of the graph at time step t-1. If question q or its associated skill s at time step t has not appeared in the graph at time step t-1, a new node for question q and its corresponding node for skill s will be added, along with corresponding edges and their attribute information. If question q or its associated skill s has appeared in the graph at the previous time step, no new nodes will be added; only the edge attributes will be updated. Three attributes are set: number of correct answers, number of incorrect answers, and whether the current answer is correct (0 indicates incorrect, 1 indicates correct). For example, [0, 0, 0] indicates that the number of correct answers is 0, the number of incorrect answers is 0, and the current answer is incorrect. Assuming that in time step 1 the incorrect answer is question q1, the incorrect answer count on the edge attribute is set to 1, and the correct / incorrect answer status is set to 0; in time step 2 the correct answer is question q2, a new node q2 is added with its correct answer count set to 1, the correct / incorrect answer status is set to 1, and the incorrect answer count on the edge attribute of q1 is updated by incrementing by 1; in time step 3 the correct answer is question q2, the correct answer count on the edge attribute of q2 is updated to 2, and the correct / incorrect answer status is set to 1. Then, the following structure can be constructed: Figure 4 The interaction sequence diagram is shown.

[0128] Step 5: Construct graph neural network architectures for the exercise sequence graph and the interaction sequence graph respectively. Use these two graph neural networks to obtain the representation of each graph in the exercise sequence graph and the interaction sequence graph to represent the problem and the student's knowledge state.

[0129] like Figure 5 As shown, by embedding nodes, aggregating neighbors, pooling, and using a fully connected layer, a graph representation of each time step is obtained to represent the problem and the student's knowledge state.

[0130] The architecture of a graph neural network is as follows:

[0131] Initial node embedding:

[0132] h (0) =X

[0133] Aggregate neighbors through K-layer graph convolution:

[0134]

[0135] Pooling is used to obtain the global representation of the graph:

[0136]

[0137] The final graph representation is obtained through a fully connected layer:

[0138] h graph =σ(w fch graph +b fc )

[0139] Furthermore, different graph neural networks can be selected for representation based on the dataset.

[0140] Step 6: Concatenate the exercise graph representation with the interaction graph representation, and use a fully connected layer of a neural network to predict the student's learning performance at a future time step, and evaluate the prediction performance based on the graph neural network.

[0141] By extracting the problem representation at time step t+1 and the interaction representation at time step t, and concatenating them together as input to the fully connected layer, the prediction result for time step t+1 can be obtained.

[0142] Step 7: Input the graph representation into the existing knowledge tracking and prediction model, and evaluate the performance of the method of embedding the graph neural network into the existing model and training and predicting using the input graph sequence data.

[0143] To further improve prediction accuracy, a self-attention mechanism is employed to update the hidden knowledge state. Borrowing from the context-aware concept of Transformers, the question sequence is processed with self-attention to obtain a new question sequence representation, and the interaction sequence is processed with self-attention to obtain a more accurate knowledge state. These are then simultaneously input into a multi-head attention system to update the knowledge state representation of the relevant prediction questions, thereby predicting the results. Considering the forgetting factor, an exponential decay over time is embedded in the attention weight calculation process, making the model pay more attention to the knowledge recently learned by students.

[0144] Graph neural networks are embedded into the input of existing deep knowledge tracing methods. Dynamic graph sequence data is input, and the representation obtained by the graph neural network is further used in conjunction with existing deep knowledge tracing methods to predict whether a student can answer a question correctly at each time step. The prediction results of the method embedding the graph neural network into the existing knowledge tracing model are evaluated using AUC, ACC, MAE, and RMSE.

Claims

1. A deep knowledge tracing method based on graph neural networks, characterized in that, The specific steps are as follows: Step one, clean the public real data set, keep the problem data, skill data, and the interaction data and student data of the two, and delete all the necessary numerical data, and number the remaining data after deletion; Step two, according to the problem number and skill number, the binary relationship between the problem and the skill in each interaction data is stored in the dictionary; Step three, according to the student id, all the problems of the same student at different time steps are composed into a sequence according to the increasing time order; Step four, for student A, the sequence is constructed according to the binary relationship stored in the dictionary, and l dynamic problem and skill graphs are generated respectively, and the corresponding exercise sequence graph and interaction sequence graph are generated; The generation process of student A's exercise sequence graph and interaction sequence graph is as follows: First, for l time steps, the interaction between student A and each exercise is obtained according to the exercise index, and is combined with the exercise index binary to arrange into an interaction sequence; Then, according to the interaction sequence and the dictionary, the exercise sequence graph is constructed, which contains problem and skill nodes, and the edges between problem-skill nodes contain two attributes: the number of correct answers and the number of wrong answers of the current problem; Finally, based on the exercise sequence graph, superimpose the exercise sequence graph of all previous time steps of this student, and add a third attribute of whether the current answer is correct or not on the edge between the problem-skill nodes to form the interaction sequence graph; Step five, use the graph convolution method to construct the corresponding graph neural network using the exercise sequence graph and the interaction sequence graph of all time steps of student A, and obtain the representation of each graph in the exercise sequence graph and the interaction sequence graph; Step six, splice the exercise graph representation of t+1 time step and the interaction graph representation of t time step, and predict the learning performance of student A at t+1 time step through a fully connected neural network to evaluate the prediction performance of the graph neural network; Step seven, embed the graph neural network into the existing Transformer model, and input the representation of the dynamic problem sequence graph and the interaction sequence graph respectively, train and predict whether the student answers correctly at each time step, and update the parameters of the graph neural network and the Transformer model through back propagation. 2.The deep knowledge tracing method based on graph neural network of claim 1, wherein, In step one, if the problem and the skill appear in an interaction, it means that the problem contains the skill; The extracted student id, problem identifier and skill identifier are numbered from 0 to obtain new digital number identifier. 3.The deep knowledge tracing method based on graph neural network of claim 1, wherein, In the step two, for the interaction data identified as d1, it contains one question q1 and one or more skills d1 = (q1, {s1, s2,..., s k}), for the interaction data identified as d2, it contains question q1 and one or more skills d2 = (q1, {s k+1 ,s k+2 ,...,s h}), then the dictionary is stored as: QS[q1] = {s1, s2,..., s k ,s k+1 ,...,s h}; where the question q1 is a key in the dictionary, the skill list {s1, s2,..., s k , k+1 s h} is the value corresponding to the key in the dictionary. 4.The deep knowledge tracing method based on graph neural network of claim 1, wherein, In step three, each student has one problem in a time step; the length of each sequence is l, and the sequence includes sequence length, problem number, skill number and answer result. 5.The deep knowledge tracing method based on graph neural network of claim 1, wherein, In step three, the architecture of the two graph convolutional neural networks is as follows: Initial node embedding: h (0) = X Aggregate neighbors through K layers of graph convolution: Pool to get the global representation of the graph: Get the final graph representation through the fully connected layer: h graph = σ(w fc h graph + b fc ) In addition, different graph neural networks can be selected for representation according to the data set; where X is the node feature of the graph; h (0) represents the feature vector of the first layer; is the feature vector of node v after k times of aggregation iteration, that is, the feature vector of the kth layer; is the neighbor node of node v, is the neighbor node of node u; is the feature vector of node u in the k-1th layer; {v} is a node set; σ is a nonlinear activation function; W (k) is the weight of the kth layer; is the feature vector of node i in the kth layer; h graph is the global representation of the pooled graph; w fc and b fc are the weight vector and bias vector of the full connection layer respectively; N represents the total number of nodes of the question sequence graph / interaction sequence graph, and the total number of nodes of the two graphs is different. 6.The deep knowledge tracing method based on graph neural network of claim 1, wherein, The specific steps of step six are as follows: Step 601, represent the question of the current t time step with q t The interaction graph of the previous time step represents h t-1 Splice to get a new vector z t ; As shown in the following formula: Step 602, vector z t is input to a first linear layer, and after ReLU activation function and Dropout regularization are applied in sequence, the result is input to a second linear layer, and the output result a2 is output: a1 = W1z t + b1 h1 = ReLU (a1) h'1 = Dropout (h1, p) a2 = W2h'1 + b2 W1 and W2 are weight matrices of the first linear layer and the second linear layer respectively, b1 and b2 are bias vectors of the first linear layer and the second linear layer respectively, and p is a probability that each neural unit is discarded in the Dropout regularization; Step 603: input the result a2 into a Sigmoid activation function to obtain a predicted value y: Step 604: evaluate the predicted value y by using AUC, ACC, MAE and RMSE indexes.

Images