Research On Heterogeneous Graph Representation Learning Algorithm Based On Meta-path

Heterogeneous graphs contain different types of nodes and different types of relationships,which are more complex than homogeneous graphs,but more in line with real-life data.Most of the existing research is carried out on isomorphic graphs.The main purpose of heterogeneous graph representation learning is to learn the representation of each node in the graph in low-dimensional space,and retain the structural and semantic information of the whole graph for downstream tasks,such as node classification,node clustering,link prediction,etc.In view of the complexity of heterogeneous graph structure,meta-path has become an important method to capture heterogeneous graph structure.The existing research mainly utilizes meta-paths to capture high-order neighbors.However,the traditional definition of meta-paths only distinguishes whether there is a connection between nodes in the network pattern,ignoring the type of edges.This leads to inaccurate node representations and suboptimal downstream task performance.In addition,existing methods mainly perform feature aggregation on instantiated paths,without enhancing the node features on instantiated paths.If facing a sparse heterogeneous graph structure with noisy node features,there will be certain interference to the learning of heterogeneous graph representation,which will result in a decrease in the accuracy of node representation.The main work of this paper aims to address the existing problems and challenges in the field,and focuses on developing a heterogeneous graph representation learning algorithm based on meta-path.The specific tasks include:(1)A heterogeneous graph embedding learning algorithm based on meta-path aggregation isproposed and applied to link prediction tasks.In a heterogeneous graph,a node can be connected by multiple types of edges.In fact,each type of edge represents a scene.Intuitively,if the embedding representations of training nodes in different scenarios are organically combined,a more accurate representation of nodes can be obtained.Firstly,a novel meta-path definition is proposed,which incorporates the edge type,that is,the relationship between nodes.From the perspective of this novel meta-path,a heterogeneous graph can be regarded as a composite of multiple relational subgraphs.In different subgraphs,the embedding of nodes is trained separately by encoding and aggregating the neighbors in the path,which is the instance layer of the proposed novel meta-path.Then,the final embedding of nodes is obtained by using the attention mechanism,which aggregates nodes from paths,which is the semantic layer of the novel meta-path.Link prediction is a downstream task to evaluate the effectiveness of learning embedding.A large number of link prediction experiments were performed on four real heterogeneous graph datasets.(2)A meta-path graph enhanced heterogeneous graph neural network method using contrastlearning is proposed.For the sparse problem of heterogeneous graph structures,the existing meta-path method is to perform first-order and high-order neighbor sampling on the target node to enrich the representation of the target node,resulting in a high degree of similarity in node characteristics.Aiming to address this problem,this paper uses the information enhancement characteristics of contrastive learning to narrow the similarities between nodes and push dissimilar nodes further apart.Specifically,the characteristics of the same node on different meta-paths have strong similarity,and the characteristics of the same type of nodes on the same meta-path are quite different from those of other nodes.In addition to contrasting the target nodes within the meta-path with different meta-paths,the graph-based attention node representation learning(structural view)and the meta-path-based representation learning(semantic view)are further contrasted across views to improve the accuracy of node representation.Finally,the trained features are aggregated to obtain the final representation of the node.Node classification is used as a downstream task to evaluate the effectiveness of the proposed method.Experiments show that the proposed method significantly improves upon existing methods.In summary,this article proposes solutions for the problems of traditional meta-path-based methods and the sparsity issue of heterogeneous graph structures in heterogeneous networks,and evaluates their performance on multiple datasets.Experimental results demonstrate that the proposed methods achieve better performance on node embedding representations in heterogeneous graphs.