A prerequisite learning method based on hypergraph neural network multi-angle feature fusion

Abstract

The application relates to a prerequisite learning method based on a hypergraph neural network multi-angle feature fusion, and belongs to the field of data mining. Based on the ability of hypergraph coding high-order topological structure, a hypergraph structure is modeled according to the relevancy of concepts in document resources, the semantic similarity between concepts, and the membership relationship between document resources and concepts, so as to sufficiently model and express the association relationship between concepts. Two-stage feature fusion is performed on the generated concept latent representation under different semantic hypergraph structures, and the model is jointly trained and optimized in terms of concept prerequisite relationship and document prerequisite relationship. Synonymous concept expansion is used to expand the positive samples of the concept prerequisite relationship to solve the data imbalance problem. The above method improves the prediction accuracy of the model.

Description

Technical Field

[0001] The present invention belongs to the field of data mining and relates to a prior learning method based on multi-angle feature fusion of hypergraph neural network. Background Art

[0002] With the rapid development of online education platforms, a vast amount of resources has been accumulated for learners. Planning precise, reasonable, and personalized learning paths for each learner within this vast learning resource is a key research topic in adaptive learning, and identifying concept prerequisite relationships plays a key role in this process. Concept prerequisite relationships (Concept Prerequisite Learning) are the dependencies between concepts. Once established, they can be widely used in downstream tasks such as course recommendations, learning path planning, learning resource sorting, and knowledge tracking.

[0003] Existing research on concept prerequisite relationships is primarily based on feature extraction and binary graph structures. Feature extraction-based methods rely on handcrafted features and perform well in textbook document resources with standardized document structures, but lack generalization and have high computational time costs. With the development of graph neural networks, more work is being done to model binary graph structures for concepts and the document resources to which they belong. However, this approach struggles to represent the complex relationships between concepts and document resource objects. When identifying concept prerequisite relationships in courses, concepts and document resources exhibit complex relationships, such as many-to-many relationships between concepts and documents, forward-backward associations between documents, and clustering of concepts with similar concepts. However, methods based on binary graph structures cannot fully express these complex relationships between concepts and document resources, making it difficult to fully model complex, high-order relationships between objects. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a prerequisite learning method based on multi-angle feature fusion of hypergraph neural network. In view of the fact that single hypergraph structure modeling is insufficient in expressing the association relationship of modeling objects, it is proposed to construct a hypergraph from three perspectives, namely the structural relationship of concepts in document resources, the semantic similarity relationship between concepts, and the affiliation between document resources and concepts, to deeply explore the high-order relationship of modeling objects from multiple perspectives, and learn the potential representations of concepts under different semantic structures. In addition, a two-stage fusion mechanism is proposed to fuse multiple potential representations of concepts obtained by multi-angle modeling. Since the sequence of document resources contains the concept prerequisite relationship, the cross-entropy loss function is used to jointly optimize the concept prerequisite relationship learning and the document resource prerequisite relationship learning, thereby improving the prediction accuracy of the model.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] Obtain a real, publicly available dataset for learning concept prerequisite relationships, preprocess the dataset, and annotate the dataset with synonyms for the concept set. Delete concepts that do not appear in the document resources and delete documents that do not contain the concepts.

[0007] Use concept synonyms to replace the concept prerequisite relationship pairs annotated in the dataset to enhance the data of positive samples;

[0008] Multi-angle hypergraph modeling uses concept-concept relationships and concept-document relationships to build concept structure hypergraphs, concept semantic distance hypergraphs, and document concept hypergraphs. The concept structure hypergraph calculates the correlation between concepts in document resources using the co-occurrence window of point-by-point mutual information (PMI) to obtain the concept structure matrix. (N represents the number of concepts). Concept semantic distance hypergraph modeling uses isotropic feature mapping (ISOMAP) to calculate the manifold geodesic distance between concepts. And convert it to get the semantic similarity matrix The document concept hypergraph is constructed by the relationship between concepts and documents. Compared with binary graph structures, the ability of hypergraphs to represent complex, high-order relationships is more conducive to modeling many-to-many and clustering relationships between documents and concepts.

[0009] Hypergraph convolution uses a hypergraph neural network to perform convolution operations on the concept structure hypergraph, concept semantic distance hypergraph, and document concept hypergraph, learning the latent representation of concepts under different semantic hypergraph structures. In addition, the document concept hypergraph also provides the latent representation of document resources.

[0010] Concept feature fusion uses a two-stage feature fusion process to combine the latent representations of concepts under different semantic structures, derived from a multi-angle modeling hypergraph structure, to further explore concept prerequisite relationships. In the first stage, multiple latent features of concepts are vertically concatenated, and then the features are fused using the maximum, mean, or summation. In the second stage, a self-attention mechanism is used to further fuse features across the entire concept domain.

[0011] Prerequisite relationship prediction: the concept latent representation obtained after concept feature fusion and the document resource latent representation obtained by the document concept hypergraph are respectively sent to the twin network to predict the concept prerequisite relationship and document resource prerequisite relationship;

[0012] Use the cross entropy loss function to measure the concept prerequisite loss and document resource prerequisite loss The final loss is adjusted by the hyperparameter λ.

[0013] The beneficial effects of the present invention are:

[0014] Based on the clustering characteristics of the concepts and their related concepts in the document resources, and the clustering characteristics of the concepts and their similar concepts, and the many-to-many relationship between the concepts and the document resources, the complex and high-order correlation relationship is more accurately and fully modeled by using the hypergraph structure, three hypergraph structures with different expression capabilities are constructed from three angles, which is beneficial to deeply mine the correlation between concepts, more accurately learn the concept representation, and improve the prediction ability of the model to the concept prerequisite relationship.

[0015] Based on the manifold geodesic distance metric distance between concepts, the similarity between concepts is further calculated, and compared with the traditional cosine similarity calculation method, the similarity between concepts is more accurately measured under the high-dimensional features of concept word embedding, which is beneficial to the concept semantic hypergraph modeling based on concept similarity.

[0016] The same concept in the document resource has different expressions of the synonym concept, the synonym concept is used to replace the positive sample of the concept pair with the prerequisite relationship for data enhancement, and the influence of the classification imbalance problem on the error of the positive sample classification is reduced.

[0017] Other advantages, objects and features of the present application will be set forth in part in the following specification, and in part will be apparent from the study thereof, or can be learned from practice of the present application. The objectives and other advantages of the present application can be realized and obtained by the following description. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to make the objects, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be combined with the drawings as follows, wherein:

[0019] Figure 1 The schematic diagram of the present application. DETAILED DESCRIPTION

[0020] The embodiments of the present application are described below by specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the disclosure of the present specification. The present application can also be implemented or applied by different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the diagrams provided in the following examples only illustrate the basic concept of the present application in a schematic manner, and the following examples and features in the examples can be combined with each other without conflict.

[0021] Among them, the drawings are only used for illustrative description, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation on the present application; in order to better illustrate the embodiments of the present application, some components of the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings can be omitted.

[0022] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "front", "back" and the like indicate the orientation or positional relationship shown in the drawings, only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the positional relationship described in the drawings is only used for illustrative description, and cannot be understood as a limitation on the present application, for those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0023] Referring to Figure 1 , Figure 1 is a structural schematic diagram of the concept prerequisite learning method based on hypergraph neural network and multi-angle feature fusion of the present application. The concept prerequisite learning method based on hypergraph neural network and multi-angle feature fusion model is composed of a hypergraph convolution module, a concept fusion module and a twin network module, specifically as follows:

[0024] The concept set C and the document resource set D provided by the preprocessed data set construct a concept structure hypergraph Concept semantic distance hypergraph And document concept hypergraph Modeling the high-order relationship between document resources and concept objects, wherein the number of concept sets is N, and the number of document resource sets is M.

[0025] Hypergraph (Hypergraph) is a general graph structure, whose edges can be connected with any number of nodes, called hyperedge (Hyperedge). Hypergraph is a hypergraph node set, and ε is a hyperedge set, is a diagonal matrix, representing the hyperedge weight, and the hypergraph The topological structure of the hypergraph is represented by the incidence matrix .

[0026] Concept structure hypergraph The concept structure matrix STM[i, j] = PMI(i, j) = logP(i, j) / P(i)P(j), P(i, j) = #W(i, j) / #W, P(i) = #W(i) / #W, where #W(i, j) is the number of the concept c i and c j The number of sliding windows, #W(i) is the number of sliding windows containing concept c i The number of sliding windows of D, #W is the total number of sliding windows of D. When the concept c i Corresponding to {STM i，l >0|l=1,2,..,N,i≠l} is not an empty set, c i As hyperedges, the incidence matrix H of the concept structure hypergraph struct =STM.

[0027] Conceptual semantic distance hypergraph Using ISOMAP to calculate the geodesic distance of concept manifold and concept word embedding (d is the word embedding dimension) obtained by SimCSE, and the concept pairwise Euclidean distance D is calculated from the word embedding X X , select the n nearest neighbors of a concept to form an undirected weighted graph, and use the Floyd algorithm to update the geodesic distance between concepts to obtain Using the formula Calculate the concept semantic similarity matrix When the concept i Corresponding to {SEM i，l >0.5|l=1,2,..,N,i≠l} is not an empty set, c i As hyperedges, the incidence matrix H of the concept semantic distance hypergraph semantic =SEM.

[0028] Document Concept Hypergraph The document concept matrix is ​​obtained by constructing the relationship between concepts and documents and calculating the term frequency inverse document frequency (tfidf) of the concept from the document resource D. Documentation Resources i As a hyperedge, d i The concepts contained are the nodes connected to it, and the association matrix H of the document concept hypergraph doc =DCM.

[0029] Using Hypergraph Neural Network (HGNN) to learn node representation, H struct 、H semantic and H doc and concept embedding X are fed into Using the hypergraph convolution formula Generate concept latent representation X c struct , X c semantic , Xc doc and document resource latent representation X doc .

[0030] The concept fusion module adopts two-stage feature fusion. The first stage: X F =F(Concat(X c_struct , X c_semantic , X c_doc )), fusion operation F∈{Max,Avg,Sum}. In the second stage, the multi-head self-attention mechanism is used to capture the dependency between concepts in the concept domain, and X F Make residual connection. X concept =LayerNorm(X F +Concat(head1,...,head h )W o ),in

[0031] The Siamese Network module uses a weight-shared Siamese Network to predict the prerequisite relationship between document resources and concepts, and uses cross entropy loss to calculate the loss value of concept pairs and document pairs. and The final loss value

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A prior learning method based on multi-angle feature fusion of hypergraph neural network, characterized by: The method comprises the following steps: Step 1: Annotate the concept vocabulary provided by the concept prerequisite relationship dataset and annotate synonyms; Step 2: Multi-angle hypergraph modeling, using concept-concept relationships and concept-document relationships to construct concept structure hypergraph, concept semantic distance hypergraph, and document concept hypergraph; Conceptual Structure Hypergraph The conceptual structure matrix is ​​calculated by point-by-point mutual information PMI STM[i,j]=PMI(i,j)=logP(i,j) / P(i)P(j),P(i,j)=#W(i,j) / #W,P(i)=#W(i) / #W, where #W(i,j) is the number of the concept c i and c j The number of sliding windows, #W(i) is the number of sliding windows containing concept c i The number of sliding windows of D, #W is the total number of sliding windows of D; when the concept c i Corresponding to {STM i，l >0|l=1,2,..,N,i≠l} is not an empty set, c i As hyperedges, the incidence matrix H of the concept structure hypergraph struct =STM; Conceptual semantic distance hypergraph Using isometric feature mapping to calculate concept manifold geodesic distance and concept word embedding Obtained by SimCSE, d is the word embedding dimension; the concept pairwise Euclidean distance D is calculated from the word embedding X X , select the n nearest neighbors of a concept to form an undirected weighted graph, and use the Floyd algorithm to update the geodesic distance between concepts to obtain Using the formula Calculate the concept semantic similarity matrix When the concept i Corresponding to {SEM i，l >0.5|l=1,2,..,N,i≠l} is not an empty set, c i As hyperedges, the incidence matrix H of the concept semantic distance hypergraph semantic =SEM; Document Concept Hypergraph The document concept matrix is ​​obtained by constructing the relationship between concepts and documents and calculating the inverse document frequency of the concept term frequency from the document resource D. Documentation Resources i As a hyperedge, d i The concepts contained are the nodes connected to it, and the association matrix H of the document concept hypergraph doc =DCM; Step 3: Hypergraph convolution: Perform convolution operations on each of the hypergraphs constructed in step 2 to obtain the concept latent representation and document resource latent representation; Step 4: Concept feature fusion: The concept latent representations obtained from the three hypergraphs are subjected to two-stage feature fusion to obtain the final concept latent representation; Step 5: Prerequisite relationship prediction, using the twin network to predict concept prerequisite relationships and document resource prerequisite relationships; Step 6: Use the cross entropy loss function to measure the concept prerequisite loss and document resource prerequisite loss The final loss is adjusted by the hyperparameter λ.

2. The prior learning method based on multi-angle feature fusion of hypergraph neural network according to claim 1 is characterized by: The synonym tagging is to use synonym replacement for positive sample data with concept prerequisite relationship, realize data enhancement for minority class samples, and solve the problem of data imbalance classification.

3. The prior learning method based on multi-angle feature fusion of hypergraph neural network according to claim 2 is characterized by: The concept feature fusion is as follows: in the first stage, the concept potential representations obtained by hypergraph convolution are spliced, and the features are fused using the maximum value, mean value or summation; in the second stage, the concept features are further fused based on the self-attention mechanism.

Images