Knowledge Tracing and Test Question Answering Prediction Method Based on Heterogeneous Graph Attention Network

Abstract

The present invention relates to the field of educational data mining, and discloses a knowledge tracing and test question answering prediction method based on a heterogeneous graph attention network, including: a node embedding layer for mapping nodes of different types into low-dimensional vector representations; a graph attention layer, including: a linear transformation layer for converting the dimension of the input features into a form suitable for multi-head attention calculation; a type-specific learnable parameter tensor for calculating attention coefficients between nodes connected by edges of different types; a LeakyReLU activation function for enhancing the expressive power of the model; a type-specific GAT layer, a ModuleList composed of multiple GraphAttentionLayers, for performing multi-layer attention mechanism processing on node features; a GCN layer, using a GCNConv layer as a feature enhancement module for fusing graph structure information to update node features; an output layer, including a linear layer for mapping high-dimensional features to a single predicted value, and the predicted value is the predicted value of the correct rate of a student answering a question.

Description

Technical Field

[0001] The present invention relates to the field of educational data mining, and specifically to a knowledge tracing and test question answering prediction method based on a heterogeneous graph attention network. Background Art

[0002] In the field of educational data mining, early research mainly focused on simple statistical analysis and association rule mining, making it difficult to deeply mine complex graph structure relationships. Recently, graph neural networks have developed rapidly. However, existing heterogeneous graph models mostly adopt a unified way to process nodes and edges, without fully considering the differences in node types (students, questions, knowledge, timestamps) and various edge types (such as student answering questions, knowledge association, etc.) in the educational scenario;

[0003] Disadvantages of the prior art: When traditional graph neural network models process heterogeneous graph data, they often cannot effectively capture the complex relationships between different types of nodes and edges, especially in the heterogeneous graph composed of students, questions, knowledge, and timestamps in the educational field. This results in insufficient accuracy in tasks such as predicting the correct rate of students answering questions. At the same time, the model structure is single, making it difficult to adapt to the analysis needs in different scenarios of individual students and all students, and it has poor scalability and flexibility. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a knowledge tracing and test question answering prediction method based on a heterogeneous graph attention network, including:

[0005] A node embedding layer for mapping different types of nodes into low-dimensional vector representations, including:

[0006] A student node embedding layer for mapping the discrete number of a student into a low-dimensional vector representation with a dimension of student_embed_dim;

[0007] A question node embedding layer for representing the attribute features of a question and mapping the discrete number of the question into a vector with a dimension of question_embed_dim;

[0008] A knowledge node embedding layer for converting the discrete number of knowledge into a vector with a dimension of knowledge_embed_dim;

[0009] A timestamp node embedding layer for converting timestamp information into a vector with a dimension of timestamp_embed_dim;

[0010] An edge embedding layer for embedding different types of edges in the heterogeneous graph with a dimension of edge_embed_dim;

[0011] A graph attention layer, including:

[0012] A linear transformation layer that converts the dimension of the input features into a form suitable for multi-head attention calculation;

[0013] Type-specific learnable parameter tensors for calculating attention coefficients between nodes with different types of edge connections;

[0014] The LeakyReLU activation function for enhancing the expressive power of the model;

[0015] Type-specific GAT layers, a ModuleList composed of multiple GraphAttentionLayer, for performing multi-layer attention mechanism processing on node features;

[0016] GCN layers that use GCNConv layers as feature enhancement modules for fusing graph structure information to update node features;

[0017] An output layer, including a linear layer, that maps high-dimensional features to a single predicted value, and the predicted value is the predicted value of the correct rate of the student's answer to the question.

[0018] Furthermore, it also includes a graph convolutional layer for aggregating the information of all students, whose input and output dimensions are both the student embedding dimension student_embed_dim. By constructing the graph structure relationship between students and performing graph convolution operations, the feature representation of each student can fuse the relevant information of other students.

[0019] Furthermore, it also includes parameters for constructing the student association relationship, which are used to represent the knowledge similarity weight and the question similarity weight. When constructing the adjacency matrix between students, it fuses the adjacency matrices constructed according to knowledge point association and question association.

[0020] Furthermore, the forward propagation process of the model includes:

[0021] S1, Data preprocessing and embedding layer operations, which process the graph data of a single student or all students according to the mode, extract the features of various types of nodes and the information of edges, and convert the discrete numbers into vector representations through the corresponding embedding layers;

[0022] S2, Graph attention layer processing, which sequentially passes the embedding representations through multiple graph attention layers for feature extraction and update;

[0023] S3, Aggregating the information of all students, by constructing the association adjacency matrices of students on knowledge points and questions, and performing graph convolution operations to obtain the aggregated student feature representations;

[0024] S4. The GCN layer and the output layer operate. After aggregating the information of all students, the embedded representation is further enhanced in features through the GCNConv layer to construct the embedded representation of node pairs for prediction, and then input into the output layer to obtain the predicted accuracy value.

[0025] Further, the process of constructing the student association relationship includes:

[0026] S11. The build_knowledge_adjacency_matrix method is used to calculate the similarity based on the knowledge embeddings in the student embeddings, and determine the adjacency relationship according to the threshold setting to obtain the adjacency matrix based on the knowledge embedding similarity.

[0027] S12. The build_question_adjacency_matrix method is used to calculate the similarity based on the question embeddings in the student embeddings, and determine the adjacency relationship according to the threshold setting to obtain the adjacency matrix based on the question embedding similarity.

[0028] S13. The compute_similarity method is used to calculate the cosine similarity between the input embedding vectors and return the similarity matrix.

[0029] The beneficial effects of the present invention are as follows:

[0030] Mode flexibility: It can flexibly switch between the single-student and all-student analysis modes. The single-student mode provides fine-grained support for personalized learning diagnosis, and the all-student mode helps in educational policy formulation and overall teaching evaluation, meeting the needs of diverse educational scenarios, which is difficult to achieve in the prior art.

[0031] Feature expression enhancement: The multi-layer GAT and GCN cooperate to make the node feature expression richer, effectively mining the deep semantics of the educational heterogeneous graph, which is superior to most single-structure models. Brief Description of the Drawings

[0032] Figure 1 It is a schematic flowchart of the knowledge tracing and test question answering prediction method based on the heterogeneous graph attention network;

[0033] Figure 2 It is an implementation schematic diagram of the knowledge tracing and test question answering prediction method based on the heterogeneous graph attention network. Detailed Embodiments

[0034] The technical solution of the present invention will be further described in detail below with reference to the drawings, but the protection scope of the present invention is not limited to the following description.

[0035] The features and performance of the present invention will be further described in detail below in combination with embodiments.

[0036] As Figure 1As shown in the figure, a knowledge tracing and test question answering prediction method based on a heterogeneous graph attention network includes:

[0037] A node embedding layer for mapping nodes of different types into low-dimensional vector representations, including:

[0038] A student node embedding layer for mapping the discrete numbers of students into low-dimensional vector representations with a dimension of student_embed_dim;

[0039] A question node embedding layer for representing the attribute features of questions, mapping the discrete numbers of questions into vectors with a dimension of question_embed_dim;

[0040] A knowledge node embedding layer for converting the discrete numbers of knowledge into vectors with a dimension of knowledge_embed_dim;

[0041] A timestamp node embedding layer for converting timestamp information into vectors with a dimension of timestamp_embed_dim;

[0042] An edge embedding layer for embedding different types of edges in the heterogeneous graph, with a dimension of edge_embed_dim;

[0043] A graph attention layer, including:

[0044] A linear transformation layer for converting the dimension of the input features into a form suitable for multi-head attention calculation;

[0045] Type-specific learnable parameter tensors for calculating the attention coefficients between nodes connected by different types of edges;

[0046] A LeakyReLU activation function for enhancing the expressive power of the model;

[0047] Type-specific GAT layers, a ModuleList composed of multiple GraphAttentionLayer, for performing multi-layer attention mechanism processing on node features;

[0048] A GCN layer, using a GCNConv layer as a feature enhancement module, for fusing graph structure information to update node features;

[0049] An output layer, including a linear layer, for mapping high-dimensional features to a single predicted value, and the predicted value is the predicted value of the correct rate of students answering questions.

[0050] It also includes a graph convolutional layer for aggregating the information of all students, with both the input and output dimensions being the student embedding dimension student_embed_dim. By constructing the graph structure relationship among students and performing graph convolution operations, the feature representation of each student can fuse the relevant information of other students.

[0051] It also includes parameters for constructing the student association relationship, which are used to represent the knowledge similarity weight and the question similarity weight. When constructing the adjacency matrix among students, it fuses the adjacency matrices constructed based on knowledge association and question association.

[0052] The forward propagation process of the model includes:

[0053] S1. Data preprocessing and embedding layer operation. According to the mode, it processes the graph data of a single student or all students, extracts the features of various types of nodes and the information of edges, and converts the discrete numbers into vector representations through the corresponding embedding layer.

[0054] S2. Graph attention layer processing. It sequentially passes the embedding representation through multiple graph attention layers for feature extraction and update.

[0055] S3. Aggregating the information of all students. By constructing the association adjacency matrix of students in terms of knowledge points and questions and performing graph convolution operations, the aggregated student feature representation is obtained.

[0056] S4. GCN layer and output layer operation. After the embedding representation after aggregating the information of all students passes through the GCNConv layer for feature enhancement, the node pair embedding representation for prediction is constructed and input into the output layer to obtain the predicted correct rate value.

[0057] The process of constructing the student association relationship includes:

[0058] S11. The build_knowledge_adjacency_matrix method is used to calculate the similarity based on the knowledge embedding in the student embedding and determine the adjacency relationship according to the threshold setting, so as to obtain the adjacency matrix based on the knowledge embedding similarity.

[0059] S12. The build_question_adjacency_matrix method is used to calculate the similarity based on the question embedding in the student embedding and determine the adjacency relationship according to the threshold setting, so as to obtain the adjacency matrix based on the question embedding similarity.

[0060] S13. The compute_similarity method is used to calculate the cosine similarity between the input embedding vectors and return the similarity matrix.

[0061] Specifically, as Figure 2 shown, the present invention includes a node embedding layer:

[0062] student_embed: The student node embedding layer maps the discrete number of students into a low-dimensional vector representation with a dimension of student_embed_dim.

[0063] question_embed: The question node embedding layer maps the discrete number of questions into a vector with a dimension of question_embed_dim, which is used to represent the attribute features of questions.

[0064] knowledge_embed: The knowledge node embedding layer converts the discrete number of knowledge into a vector with a dimension of knowledge_embed_dim.

[0065] timestamp_embed: The timestamp node embedding layer converts the timestamp information into a vector with a dimension of timestamp_embed_dim.

[0066] edge_embed: The edge embedding layer embeds different types of edges in the heterogeneous graph with a dimension of edge_embed_dim.

[0067] Graph Attention Layer:

[0068] Core Components and Calculation Process:

[0069] W: The linear transformation layer converts the dimension of the input feature x from in_channels to num_heads * out_channels for multi-head attention calculation. Its role is to perform a preliminary feature transformation on the input features to make them adapt to the requirements of the multi-head attention mechanism and provide appropriate feature dimensions for subsequent attention calculation and message passing.

[0070] type_specific_a: The type-specific learnable parameter tensor with a dimension of (edge_types, num_heads, 2 * out_channels). It selects specific attention parameters according to the edge type and is used to calculate the attention coefficients between nodes connected by different types of edges. This parameter enables the model to learn different attention patterns for different types of edge relationships, thereby more carefully processing the diverse edge information in the heterogeneous graph.

[0071] leaky_relu: The LeakyReLU activation function is used to introduce non-linearity, enhance the model's expressive power, and perform activation transformation on the intermediate results during the attention calculation process.

[0072] Forward Propagation Process:

[0073] First, use the add_self_loops function to add self-loops to the input edge_index, ensuring that nodes can consider their own feature information during message passing, making the graph structure information more complete and rich.

[0074] Perform a feature transformation on the input feature x to obtain h through the W layer and adjust the shape for subsequent multi-head attention calculation. This step converts and reshapes the original node features, preparing for the parallel calculation of the multi-head attention mechanism.

[0075] Obtain the node indices row and col connected by the edges according to edge_index, and select the corresponding type of attention parameters from type_specific_a based on the edge type edge_type. Concatenate the node features h according to the edge connection situation to get a_input, and incorporate the influence factor performance_factor based on the performance of all students to consider information related to the rules of all students, making the attention calculation more targeted and effective. Obtain the attention scores attention_scores through a series of tensor operations, leaky_relu activation, and softmax operations, and apply dropout for random inactivation to prevent overfitting.

[0076] Finally, by calling the propagate method inherited from the MessagePassing base class, aggregate the features of neighbor nodes according to the attention scores, complete the message passing process, and finally return the processed and shape-adjusted output features.

[0077] Type-specific GAT layer:

[0078] A ModuleList composed of multiple GraphAttentionLayers is used for multi-layer attention mechanism processing of node features. By stacking multiple graph attention layers, the model can gradually extract higher-level and more abstract feature representations, deeply mining the complex relationships and potential patterns between nodes in the heterogeneous graph. Each layer performs further attention calculation and feature update based on the output features of the previous layer, enabling the node features to fuse more hierarchical graph structure information and neighbor node information, enhancing the model's feature extraction ability and expression ability for heterogeneous graph data.

[0079] GCN layer:

[0080] The GCNConv layer, as a feature enhancement module, further uses graph convolution operations to fuse graph structure information to update node features based on the features processed by the graph attention layer.

[0081] Output layer:

[0082] A linear layer output_layer with an input dimension of total_embed_dim * num_heads * 2 and an output dimension of 1. The role of this layer is to perform a final linear transformation on the node pair features processed by the previous multi-layer graph neural network, mapping the high-dimensional features to a single predicted value, that is, the predicted accuracy rate of the student's answer to the question. By adjusting the weight and bias parameters, the output layer can output a predicted probability value between 0 and 1 according to the input feature information, indicating the likelihood of the student answering the question correctly, thus achieving the prediction goal of the model.

[0083] A graph convolutional layer for aggregating information of all students:

[0084] A graph convolutional layer specifically designed to aggregate information related to all students, with both input and output dimensions being the student embedding dimension student_embed_dim. Its purpose is to fully consider the performance patterns of all students on the same knowledge points and questions in the model. By constructing the graph structure relationship between students and performing graph convolutional operations, the feature representation of each student can fuse the relevant information of other students.

[0085] Parameters for constructing student association relationships: learnable parameter matrices, which are used to represent the knowledge similarity weight and the question similarity weight respectively. When constructing the adjacency matrix between students, these weight parameters are used to fuse the adjacency matrices constructed based on knowledge association and question association. By adjusting the weight sizes, the model can flexibly control the importance degree of different aspects of association in constructing the relationship between students, thus more accurately capturing the complex association patterns between students based on knowledge and questions, making the process of aggregating information of all students more adaptable and effective.

[0086] The forward propagation process of the model

[0087] Data preprocessing and embedding layer operations:

[0088] According to different values of is_single_student_mode, the graph data of a single student or the graph data of all students is processed respectively. The features of various types of nodes such as students, questions, knowledge, and timestamps, as well as the edge information, are extracted from the input data dictionary data. Then, the discrete numbers of various types of nodes are converted into vector representations through the corresponding embedding layers, and these vectors are concatenated along the last dimension to obtain the combined embedding representation combined_embeds.

[0089] Graph attention layer processing:

[0090] The combined_embeds are successively passed through multiple graph attention layers (gat_layer) for feature extraction and update. In each graph attention layer, according to the calculation process of the graph attention layer, feature transformation, attention calculation, and message passing are performed, enabling the node features to gradually incorporate more graph structure-related information and neighbor node information, thereby extracting more representative and discriminative feature representations. Through the stacking of multiple graph attention mechanisms.

[0091] Aggregate all student information:

[0092] Call the aggregate_all_students_info method to aggregate all student information. This method first extracts the student node embeddings student_embeds from the combined_embeds, and then constructs an adjacency matrix based on the associations of students in knowledge points and questions. Specifically, the adjacency matrix based on knowledge embedding similarity is obtained through the build_knowledge_adjacency_matrix method, and the adjacency matrix based on question embedding similarity is obtained through the build_question_adjacency_matrix method. These two adjacency matrices are fused according to knowledge_similarity_weight and question_similarity_weight to obtain the final adjacency matrix adj_matrix between students. Then, the adjacency matrix is converted into the edge index format edge_index, and the student embeddings student_embeds and the edge index edge_index are passed into the student_agg_gcn graph convolutional layer for graph convolutional operations, enabling the information between students to spread and aggregate in the constructed graph structure, obtaining the aggregated student feature representation aggregated_embed. Finally, the aggregated features are extended to the same shape as the input combined_embeds and added to the original combined_embeds, integrating the information related to all students into the features. This process fully utilizes the performance patterns of all students on the same knowledge points and questions by constructing a reasonable student association graph structure and performing graph convolutional operations, providing a more comprehensive information reference for predicting the answering accuracy of individual students and enhancing the prediction ability and generalization performance of the model.

[0093] GCN layer and output layer operations:

[0094] After aggregating all students' information, the combined_embeds are further enhanced in features through the GCNConv layer. The graph convolution operation is used to further update the node features, enabling the node features to fuse more extensive graph structure information. Then, according to the requirements of the prediction task, the node pair embedding representation pair_embeds for prediction is constructed. In the single-student mode, the features of the student node and the question node are concatenated in order through specific indexing operations; in the all-students mode, the corresponding node pairs are generated according to the numbers of all students and all questions. Finally, the pair_embeds are input into the output layer output_layer, and the predicted correct rate value is obtained through linear transformation and returned. The output layer maps the high-dimensional features processed by the previous multi-layer graph neural network to a single predicted value, realizing the conversion from feature extraction to the final prediction result and completing the model's prediction task of the correct rate of students' answering questions.

[0095] Methods related to constructing the adjacency matrix

[0096] build_knowledge_adjacency_matrix method:

[0097] First, the corresponding knowledge embeds are extracted from the student embeds student_embeds. The knowledge embeds are in a certain fixed interval of the student embed features (implemented through the extract_knowledge_embeds method), and this interval needs to be adjusted and determined according to the feature arrangement in the actual data.

[0098] Then, the similarity between the knowledge embeds is calculated. The cosine similarity is calculated using the compute_similarity method to obtain the similarity matrix similarity_matrix between the knowledge embeds.

[0099] Finally, the adjacency relationship is determined by setting a threshold according to the similarity. The positions corresponding to the elements that meet the threshold condition are set as the adjacency relationship (value is 1), otherwise as the non-adjacency relationship (value is 0), to obtain the adjacency matrix adj_matrix based on the knowledge embed similarity. The adjacency matrix constructed in this way can reflect the association relationship between students based on their knowledge mastery, providing a graph structure information basis based on knowledge factors for subsequent graph convolution operations.

[0100] build_question_adjacency_matrix method:

[0101] Similar to the way of constructing the knowledge adjacency matrix, first, the question embeddings question_embeds are extracted from the student embeddings, and then the similarity between the question embeddings is calculated to obtain the similarity matrix. Then, according to different threshold settings, the adjacency relationship is determined to obtain the adjacency matrix based on the similarity of the question embeddings.

[0102] compute_similarity method:

[0103] Calculate the cosine similarity between the input embedding vectors and return the similarity matrix. The specific implementation process is to traverse all pairs of embedding vectors. For each pair of embedding vectors, use the F.cosine_similarity function to calculate their cosine similarity and fill the result into the corresponding position of the similarity matrix.

[0104] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications, and environments, and can be changed within the scope of the concept described herein through the above teachings or the technology or knowledge in related fields. And any changes and modifications made by those skilled in the art without departing from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. A knowledge tracing and test question answering prediction method based on a heterogeneous graph attention network, characterized in that Including: A node embedding layer, which is used to map nodes of different types into low-dimensional vector representations, including: A student node embedding layer, which is used to map the discrete numbers of students into low-dimensional vector representations with a dimension of student_embed_dim; A question node embedding layer, which is used to represent the attribute features of questions and map the discrete numbers of questions into vectors with a dimension of question_embed_dim; A knowledge node embedding layer, which converts the discrete numbers of knowledge into vectors with a dimension of knowledge_embed_dim; A timestamp node embedding layer, which converts timestamp information into vectors with a dimension of timestamp_embed_dim; An edge embedding layer, which embeds different types of edges in the heterogeneous graph with a dimension of edge_embed_dim; A graph attention layer, including: A linear transformation layer, which converts the dimension of the input features into a form suitable for multi-head attention calculation; A type-specific learnable parameter tensor, which is used to calculate the attention coefficients between nodes connected by different types of edges; A LeakyReLU activation function, which is used to enhance the expressive ability of the model; A type-specific GAT layer, which is a ModuleList composed of multiple GraphAttentionLayer, and is used to perform multi-layer attention mechanism processing on node features; A GCN layer, which uses the GCNConv layer as a feature enhancement module and is used to fuse graph structure information to update node features; An output layer, including a linear layer, which maps high-dimensional features to a single predicted value, and the predicted value is the predicted value of the correct rate of students answering questions; It also includes a graph convolutional layer for aggregating the information of all students. Its input and output dimensions are both the student embedding dimension student_embed_dim. By constructing the graph structure relationship between students and performing graph convolution operations, the feature representation of each student can fuse the relevant information of other students.

2. The knowledge tracing and test question answering prediction method based on the heterogeneous graph attention network according to claim 1, wherein It also includes parameters for constructing the association relationship between students, which are used to represent the knowledge similarity weight and the question similarity weight. When constructing the adjacency matrix between students, the adjacency matrices constructed according to knowledge point association and question association are fused.

3. The knowledge tracing and test question answering prediction method based on the heterogeneous graph attention network according to claim 2, wherein The forward propagation process of the model includes: S1, Data preprocessing and embedding layer operation. According to the mode, the graph data of a single student or all students is processed, the features of each type of node and the information of the edges are extracted, and the discrete numbers are converted into vector representations through the corresponding embedding layers; S2, Graph attention layer processing. The embedding representations are sequentially passed through multiple graph attention layers for feature extraction and update; S3, Aggregate the information of all students. By constructing the association adjacency matrix of students on knowledge points and questions and performing graph convolution operations, the aggregated student feature representation is obtained; S4, GCN layer and output layer operation. The embedding representation after aggregating the information of all students is further enhanced by the GCNConv layer, the node pair embedding representation for prediction is constructed, and the predicted correct rate value is obtained by inputting it into the output layer.

4. The knowledge tracing and test question answering prediction method based on the heterogeneous graph attention network according to claim 3, wherein The process of constructing the association relationship between students includes: S11, build_knowledge_adjacency_matrix method, which is used to calculate the similarity based on the knowledge embeddings in the student embeddings, and determine the adjacency relationship according to the threshold setting, so as to obtain the adjacency matrix based on the knowledge embedding similarity; S12, build_question_adjacency_matrix method, which is used to calculate the similarity based on the question embeddings in the student embeddings, and determine the adjacency relationship according to the threshold setting, so as to obtain the adjacency matrix based on the question embedding similarity; S13, compute_similarity method, which is used to calculate the cosine similarity between the input embedding vectors and return the similarity matrix.

Images