Deep knowledge tracking method based on graph neural network

Abstract

The invention discloses a deep knowledge tracking method based on a graph neural network, and belongs to the field of wisdom education. The method specifically comprises the following steps: firstly, cleaning a real data set, and storing a binary relationship between a problem and a skill in each piece of interactive data into a dictionary; then, all problems of the same student form a sequence according to incremental time steps, corresponding exercise sequence diagrams and interaction sequence diagrams are generated according to a binary relation, then, a corresponding graph neural network is constructed, and representation of each graph in the exercise sequence diagrams and the interaction sequence diagrams is obtained respectively; splicing the exercise map representation of the t + 1 time step and the interaction map representation of the t time step, and predicting a learning result of the t + 1 time step of the student through a full-connection neural network; and finally, the graph neural network is embedded into an existing Transform model, and parameters are updated through back propagation. According to the method, the dynamic graph representation of the problem and the interaction sequence is realized, the knowledge state is updated by utilizing an attention mechanism, and the performance and the interpretability of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the field of smart education, and specifically relates to a deep knowledge tracking method based on graph neural network. Background Art

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

[0003] The knowledge tracking task aims to evaluate students’ ability status and predict future performance by analyzing the interaction data between students and educational resources. Online education platforms can provide customized learning resources and suggestions to improve learning outcomes.

[0004] In a real learning environment, students' answers are quite sparse, especially in online learning. The platform is filled with a large number of exercises, but most of the exercises actually only involve a small number of interactions. Moreover, the input of the knowledge tracking model often uses skills as the unique identifier and lacks problem-related information, which limits the performance of the model. Summary of the invention

[0005] In response to the problems of missing question information and limited performance in existing deep knowledge tracking methods, the present invention proposes a deep knowledge tracking method based on graph neural network, which realizes dynamic graph representation of question sequence and interaction sequence and updates the knowledge status using attention mechanism, and improves the performance and interpretability of the model.

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

[0007] Step 1: Clean the public real data set, retain four types of data: question data, skill data, interaction data between the two, and student data, delete all data with empty necessary values, and number the remaining data after deletion.

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

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

[0010] Step 2: According to the question number and skill number, the binary relationship between the question and the skill in each interactive data is stored in the dictionary.

[0011] For the interaction data marked as d1, it contains a question q1 and one or more skills d1=(q1,{s1,s2,...,sk}), the interaction data marked as d2, including question q1 and one or more skills d2 = (q1, {s k+1 ,s k+2 ,...,s h}), the dictionary storage is represented as: QS[q1]={s1,s2,...,s k ,s k+1 ,...,s h};

[0012] where question q1 is the key in the dictionary and skills list {s1,s2,...,s k ,s k+1 ,...,s h} is the value corresponding to the key in the dictionary.

[0013] Step 3: According to the student ID, all questions of the same student at different time steps are sequenced in ascending chronological order.

[0014] Each student has one question in one time step; the length of each sequence selection time step 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 graph data of questions and skills respectively according to the binary relationship stored in the dictionary, and generate the corresponding exercise sequence diagram and interaction sequence diagram.

[0017] The generation process of 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 to form an interaction sequence;

[0019] Then, based on the interaction sequence and the dictionary, an exercise sequence graph is constructed, which contains two types of nodes: question and skill. The edge between the question-skill nodes contains two attributes: the number of correct answers and the number of incorrect answers to the current question.

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

[0021] Step 5: Using the exercise sequence graph and interaction sequence graph of all time steps of student A, use the graph convolution method to construct the corresponding graph neural network respectively, and obtain the representation of each graph in the exercise sequence graph and interaction sequence graph 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 layers of graph convolution:

[0026]

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

[0028]

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

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

[0031] In addition, different graph neural networks can be selected for representation based on the data set.

[0032] Among them, 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 aggregation iterations, 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 at layer k-1; {v} is the node set; σ is the nonlinear activation function; W (k) is the weight of the kth layer; is the feature vector of node i at layer k; h graph is the global representation of the pooled image; w fc and b fc are the weight vector and bias vector of the fully connected layer respectively. N represents the total number of nodes in the exercise sequence diagram / interaction sequence diagram, and the total number of nodes in the two diagrams is different.

[0033] Step 6: Concatenate the exercise graph representation at time step t+1 with the interaction graph representation at time step t, and use the fully connected neural network to predict the learning performance of student A at time step t+1 to evaluate the prediction performance of the graph neural network.

[0034] The specific steps are:

[0035] Step 601: Represent the question q at the current time step t t The interaction graph representation h with the previous time step t-1Splice and get the new vector z t ;

[0036] As shown in the following formula:

[0037]

[0038] Step 602: vector z t Input to the first linear layer, apply ReLU activation function and Dropout regularization in sequence, input the result to the second linear layer, and output the result 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, b1 and b2 are the bias vectors of the first and second linear layers, and p is the probability of each neural unit being discarded 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 indicators 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 diagram and the interaction sequence diagram as input respectively. Perform training and predict whether the student’s answer is correct or not at each time step, and update the graph neural network parameters and Transformer test model parameters through back propagation.

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

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

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

[0051] 3. The deep knowledge tracking method based on graph neural network of the present invention embeds graph neural network into the input of existing knowledge tracking model, inputs dynamic graph sequence data, trains the model, and greatly improves the performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of a deep knowledge tracking method based on a graph neural network of the present invention;

[0053] Figure 2 It is a schematic diagram of the deep knowledge tracking method based on graph neural network of the present invention;

[0054] Figure 3 It is a structure of a sequence diagram of exercises constructed in an embodiment of the present invention;

[0055] Figure 4 It is an interactive sequence diagram structure constructed in an embodiment of the present invention;

[0056] Figure 5 It is a flowchart of a neural network representation sequence diagram according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of the present invention.

[0058] The present invention proposes a deep knowledge tracking method based on graph neural network. In the process of 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 the students' questions and interaction sequences into a dynamic graph structure sequence, and using the graph neural network to capture the relationship between questions and skills, accurate question graph representations and interaction graph representations are obtained. These graph representations are spliced ​​and input into the fully connected layer, which can directly predict the students' answer results. In addition, the graph neural network can be embedded in the input layer of the existing knowledge tracking model, and predictions are made by inputting the graph sequence. Practice has proved that higher prediction performance can be obtained by inputting the dynamic graph sequence into the graph neural network and the fully connected layer for prediction; embedding the graph neural network into the existing knowledge tracking model and using the dynamic graph structure sequence as input can also improve the performance of the existing deep knowledge tracking method.

[0059] The deep knowledge tracking method based on graph neural network, such as Figure 1 As shown, the following steps are included:

[0060] Step 1: Clean the public real data set, filter out the interactive data with key information about problems and skills, retain four types of data: problem data, skill data, interactive data, and student data, delete all data with empty necessary values, and number the identifiers of the remaining data after deletion.

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

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

[0063] Step 2: According to the question number and skill number, the binary relationship between the question and the skill in each interactive data is stored in the dictionary.

[0064] As long as the problem and skill appear in an interaction, it means that the problem contains the skill, that is, the skill can help solve the problem. Summarize the binary relationship between the problem and the skill in all interaction data and store it in a dictionary. The key is the problem number and the value is the skill number list. The details are as follows:

[0065] For the interaction data marked as d1, it contains a question q1 and one or more skills d1=(q1,{s1,s2,...,s k}), the interaction data marked as d2, including question q1 and one or more skills d2 = (q1, {s k+1 ,s k+2 ,...,s h}), the dictionary storage is represented as: QS[q1]={s1,s2,...,s k ,s k+1 ,...,s h};

[0066] where question q1 is the key in the dictionary and skills list {s1,s2,...,s k ,s k+1 ,...,s h} is the value corresponding to the key in the dictionary.

[0067] Step 3: According to the student ID, all questions of the same student at different time steps are sequenced in ascending chronological order.

[0068] Each student has one question in one time step; the length of each sequence selection time step 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 relationship stored in the dictionary, and generate the corresponding exercise sequence diagram and interaction sequence diagram.

[0071] Both the exercise sequence diagram and the interaction sequence diagram contain question and skill nodes. The graph for each time step in the exercise sequence diagram is an undirected graph connecting the question of the current time step with the skills it covers, and the number of wrong answers and correct answers attributes of the current question are filled on the edge connecting the question and skill nodes.

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

[0073] The generation process of 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 to form an interaction sequence;

[0075] Then, based on the interaction sequence and the dictionary, an exercise sequence graph is constructed, which contains two types of nodes: question and skill. The edge between the question-skill nodes contains two attributes: the number of correct answers and the number of incorrect answers to the current question.

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

[0077] Step 5: Using the exercise sequence graph and interaction sequence graph of all time steps of student A, use the graph convolution method to construct the corresponding graph neural network respectively, and obtain the representation of each graph in the exercise sequence graph and interaction sequence graph respectively;

[0078] For the exercise sequence diagram and the interaction sequence diagram, 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 layers of graph convolution:

[0082]

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

[0084]

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

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

[0087] In addition, different graph neural networks can be selected for representation based on the data set.

[0088] Among them, 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 aggregation iterations, 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 at layer k-1; {v} is the node set; σ is the nonlinear activation function; W (k) is the weight of the kth layer; is the feature vector of node i at layer k; h graph is the global representation of the pooled image; w fc and b fc are the weight vector and bias vector of the fully connected layer respectively. N represents the total number of nodes in the exercise sequence diagram / interaction sequence diagram, and the total number of nodes in the two diagrams is different.

[0089] Step 6: Concatenate the exercise graph representation at time step t+1 with the interaction graph representation at time step t, and use the fully connected neural network to predict the learning performance of student A at time step t+1 to evaluate the prediction performance of the graph neural network.

[0090] The specific steps are:

[0091] Step 601: Represent the question q at the current time step t t The interaction graph representation h with the previous time step t-1 Splice and get the new vector z t ;

[0092] As shown in the following formula:

[0093]

[0094] Step 602: vector z t Input to the first linear layer, apply ReLU activation function and Dropout regularization in sequence, input the result to the second linear layer, and output the result 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, b1 and b2 are the bias vectors of the first and second linear layers, and p is the probability of each neural unit being discarded 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 indicators 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 diagram and the interaction sequence diagram as input respectively. Perform training and predict whether the student’s answer is correct or not at each time step, and update the graph neural network parameters and Transformer prediction model parameters through back propagation.

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

[0105] Traditional deep learning knowledge tracing models input question sequences and interaction sequences. Assuming the total number of questions is N Q N =(q1,q2,…,q N), the question sequence will be processed into a one-hot encoding of N dimensions, and the interaction sequence will be processed into a one-hot encoding of 2N dimensions. For the question sequence, if the question number is i, the input one-hot vector will be set to 1 at the position with index i. For the interaction sequence, in order to distinguish correct answers from incorrect answers, the dimension will be expanded to 2N. Specifically, when the student answers the question correctly, the i position is also set to 1, and the rest are 0. When the student answers the question incorrectly, the i+N position is set to 1, and the rest are 0.

[0106] In the present invention, the one-hot encoding in the original model is replaced by encoding through a graph neural network, the input of the existing knowledge tracking prediction model is embedded, the dynamic question sequence diagram and the interaction sequence diagram are used as the input of the model, and the correctness of the student’s answer is predicted at each time, and the graph neural network parameters and the knowledge tracking prediction model parameters are updated through back propagation;

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

[0108]

[0109] Among them, r t represents the correctness of the students’ answers to the questions, p t Represents the probability of the correct answer to the exercise predicted by the model. This calculation method trains the model by minimizing the difference between the two.

[0110] Example:

[0111] This embodiment is applied to the knowledge tracking task in the environment of smart education system. In the knowledge tracking task in the field of smart education, this method can use graph neural network to mine high-level information between questions and skills, and fully characterize exercise vectors and student interaction vectors. It solves the limitation of current knowledge tracking methods that use different skill numbers as unique identifiers of student answers due to data scarcity. Figure 2 The overall process framework of the present invention is shown. On the one hand, the dynamic sequence diagram constructed by this method can be directly predicted by using graph neural network and fully connected neural network processing, achieving higher performance. On the other hand, the graph neural network is embedded into the existing knowledge tracking model, and the dynamic graph sequence constructed by this method is input for training and prediction, which greatly improves the performance of the existing model.

[0112] The specific steps are as follows:

[0113] Step 1: Perform data cleaning on four public real data sets, namely, the Assist2009 and Assist2017 data sets of the online guidance platform Assistments, the Ednet data set of the English education mobile application software Santa, and the Junyi data set of the Taiwan online platform. Delete the interactive data that lacks the key information of student ID, answer time, question ID, and skill ID, and filter out the interactive data with the key information of question and skill; and number the student ID, question ID, and skill ID from 0 to obtain a new digital number ID.

[0114] 5,000 students and their interactions are randomly extracted from a larger data set, and all items with empty necessary values ​​are deleted. The extracted student information and problem information including id, problem identifier, and skill identifier are numbered starting from 0 to obtain new digital identifiers.

[0115] Step 2: From each piece of interaction data, obtain the binary relationship between the question and the skill.

[0116] For each piece of data, there is an inclusion relationship between the question and the corresponding skill in the data. The relationship between the question and skill in all the interactive data is extracted and stored in the dictionary.

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

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

[0119] Step 4: Based on the relationship obtained in step 2, the sequence data is used to construct dynamic graph data of problems and skills for each student at each time step, and generate exercise sequence diagrams and interaction sequence diagrams.

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

[0121] The relationship between questions and skills is represented in the exercise sequence diagram as an undirected graph, where nodes represent questions and skills, and edges represent the relationship between questions and skills. Each question q and its corresponding skill s are constructed into a graph as the input for predicting questions. If a question involves multiple skills, there will be edges connecting the question node and these skill nodes.

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

[0123] For each student's T-length answer sequence, N questions are known, and L skills are:

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

[0125] According to Exercise Sequence E T =(e1,e2,…,e T ), build a graph for each time step of the problem and skill. Represent the relationship between the problem and the skill as an undirected graph, where nodes represent problems and skills, and edges represent the relationship between problems and skills. Build a graph for each problem q and its corresponding skill s as the input for predicting the problem. If a problem involves multiple skills, there will be edges connecting the problem node and these skill nodes.

[0126] Specifically: For each time step, the current time step question and its corresponding skill are constructed into a graph structure. The question and skill are connected, and their edges can add additional information (the number of wrong answers and correct answers to the current question). Suppose the question corresponding to exercise e1 at time step 1 is q1, the question corresponding to exercise e2 at time step 2 is q2, and the question corresponding to exercise e3 at time step 3 is still q2; the problem-to-skill mapping relationship QS = {q1: [s1], q2: [s1, s2], q3: [s1, s2]}. Then we can construct the following Figure 3 In the exercise sequence diagram shown, blue nodes represent skills, yellow nodes represent questions, and there are two attributes on the edge. The first attribute is the number of wrong answers to the current question, and the second attribute is the number of correct answers. In exercise e3, question q2 was answered correctly in the last answer, so the first attribute is 1 and the second attribute is 0.

[0127] Construct a graph of questions and skills before the current moment. The questions are connected to their related skills, and add rich attribute information on the edges, such as the number of correct answers, the number of incorrect answers, and the correct and incorrect answer results. The knowledge state graph constructed at time step t is an update of the graph at time step t-1. If the question q at time step t or its related skill s has not appeared in the graph at time step t-1, a new node for question q and its corresponding skill s node will be added. Add the corresponding edges and attribute information on the edges. If the question q or the related skill s has appeared in the graph of the previous time step, no new nodes will be added, and only the attributes of the edges will be updated. Three attributes are set, namely the number of correct answers, the number of incorrect answers, and whether the current answer is correct or not (0 means incorrect, 1 means correct). For example, [0, 0, 0] means the number of correct answers is 0, the number of incorrect answers is 0, and the current answer is incorrect. Assume that the incorrect answer to question q1 at time step 1 is set to 1, the number of wrong answers of the edge attribute is set to 1, and the current correct or incorrect answer is set to 0; at time step 2, the correct answer to question q2 is set to 1, the correct or incorrect answer of the current answer is set to 1, and the number of wrong answers of the edge attribute of q1 is updated by 1; at time step 3, the correct answer to question q2 is set to 2, and the current correct or incorrect answer is set to 1, then the following can be constructed: Figure 4 The interaction sequence diagram shown.

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

[0129] like Figure 5 As shown in the figure, through node embedding, neighbor aggregation, pooling, and fully connected layers, a graph representation of each time step is obtained to represent the problem and the student's knowledge state.

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

[0131] Initial node embedding:

[0132] h (0) =X

[0133] Aggregate neighbors through K layers of graph convolution:

[0134]

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

[0136]

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

[0138] h graph =σ(w fch graph +b fc )

[0139] In addition, different graph neural networks can be selected for representation based on the data set.

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

[0141] The problem representation at time step t+1 and the interaction representation at time step t are extracted and concatenated together as the input of the fully connected layer to obtain the prediction result for time step t+1.

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

[0143] In order to further improve the prediction accuracy, the self-attention mechanism is used to further update the hidden knowledge state. Drawing on the context-aware idea of ​​Transformer, the question sequence is processed by self-attention to obtain a new question sequence representation, and the interaction sequence is processed by self-attention to obtain a more accurate knowledge state. Then it is simultaneously input into the multi-head attention to update the knowledge state representation of the relevant prediction questions, so as to predict the results. Taking into account the forgetting factor, in the process of attention weight calculation, the exponential decay with time interval is embedded, so that the model pays more attention to the knowledge that students have recently acquired.

[0144] Embed the graph neural network into the input of the existing deep knowledge tracing method. Input the dynamic graph sequence data, and the representation obtained by the graph neural network is further used by the existing deep knowledge tracing method to predict whether the student can answer the question correctly at each time step. The prediction results of the embedding graph neural network into the existing knowledge tracing model method are evaluated by AUC, ACC, MAE, and RMSE.

Claims

1. A deep knowledge tracking method based on graph neural network, characterized by: The specific steps are as follows: Step 1: Clean the public real data set, retain the four types of data: question data, skill data, interaction data between the two, and student data, delete all data with empty values, and number the remaining data after deletion; Step 2: According to the question number and skill number, the binary relationship between the question and the skill in each interactive data is stored in the dictionary; Step 3: According to the student ID, all questions of the same student at different time steps are sorted into a sequence in ascending time order; Step 4: For student A, construct a dynamic graph of questions and skills based on the binary relationship stored in the dictionary, and generate the corresponding exercise sequence graph and interaction sequence graph; The generation process of Student A's exercise sequence diagram and interaction sequence diagram is as follows: 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 to form an interaction sequence; Then, based on the interaction sequence and dictionary, an exercise sequence graph is constructed. The graph contains two types of nodes: question and skill. The edge between the question-skill nodes contains two attributes: the number of correct answers and the number of incorrect answers to the current question. Finally, based on the exercise sequence diagram, the exercise sequence diagrams of all previous time steps of the student are superimposed, and a third attribute of whether the current answer is correct or not is added to the edge between the question-skill node to form an interaction sequence diagram; Step 5: Using the exercise sequence graph and interaction sequence graph of all time steps of student A, use the graph convolution method to construct the corresponding graph neural network respectively, and obtain the representation of each graph in the exercise sequence graph and interaction sequence graph respectively; Step 6: Concatenate the exercise graph representation at time step t+1 with the interaction graph representation at time step t, and use the fully connected neural network to predict the learning performance of student A at time step t+1 to evaluate the prediction performance of the graph neural network. Step 7: Embed the graph neural network into the existing Transformer model, and use the representations of the dynamic question sequence diagram and the interaction sequence diagram as input respectively. Perform training and predict whether the student’s answer is correct or not at each time step, and update the graph neural network parameters and Transformer test model parameters through back propagation.

2. The deep knowledge tracking method based on graph neural network according to claim 1, characterized in that: In step 1, if the question and the skill appear in one interaction, it means that the question contains the skill; The extracted student ID, question ID and skill ID are numbered starting from 0 to obtain new digital IDs.

3. The deep knowledge tracking method based on graph neural network according to claim 1, characterized in that: In step 2, the interactive data identified as d1 includes a question q1 and one or more skills d1=(q1,{s1,s2,...,s k }), the interaction data marked as d2, including question q1 and one or more skills d2 = (q1, {s k+1 ,s k+2 ,...,s h }), the dictionary storage is represented as: QS[q1] = {s1, s2, ..., s k ,s k+1 ,...,s h }; where question q1 is the key in the dictionary and skills list {s1,s2,...,s k ,s k+1 ,...,s h } is the value corresponding to the key in the dictionary.

4. The deep knowledge tracking method based on graph neural network according to claim 1, characterized in that: In step three, each student has one question in one time step; the length of each sequence selected time step is l, and the sequence includes sequence length, question number, skill number and answer result.

5. The deep knowledge tracking method based on graph neural network according to claim 1, characterized in that: In step 3, the architectures of the two graph convolutional neural networks are as follows: Initial node embedding: h (0) =X Aggregate neighbors through K layers of graph convolution: Pooling is performed to obtain a global representation of the graph: The final graph representation is obtained 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 based on the data set; Among them, 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 aggregation iterations, 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 at layer k-1; {v} is the node set; σ is the nonlinear activation function; W (k) is the weight of the kth layer; is the feature vector of node i at layer k; h graph is the global representation of the pooled image; w fc and b fc 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, and the total number of nodes in the two graphs is different.

6. The deep knowledge tracking method based on graph neural network according to claim 1, characterized in that: The specific steps of step six are: Step 601: Represent the question q at the current time step t t The interaction graph representation h with the previous time step t-1 Splice and get the new vector z t ; As shown in the following formula: Step 602: vector z t Input to the first linear layer, apply ReLU activation function and Dropout regularization in sequence, input the result to the second linear layer, and output the result a2: a1=W1z t +b1 h1=ReLU(a1) h′1=Dropout(h1,p) a2=W2h′1+b2 W1 and W2 are the weight matrices of the first and second linear layers, b1 and b2 are the bias vectors of the first and second linear layers, and p is the probability of each neural unit being discarded in Dropout regularization; Step 603: Input the result a2 into the Sigmoid activation function to obtain the predicted value y: Step 604: Use AUC, ACC, MAE, and RMSE indicators to evaluate the predicted value y.

Images