Research On Layered Multiplex Heterogeneous Graph Embedding Method Based On Meta-Path Random Walk

Graph embedding mainly refers to using low-dimensional,continuous,dense and correlated vectors to represent all nodes in target graph,which enables nodes to do follow-up semantic calculation and inferences.Graph embedding algorithms are based on algebra or random walk methods.With the development of deep learning,methods combining deep learning accelerate embedding process.Through the deep learning graph embedding method,the researchers studied homogeneous graphs where nodes and edges belong to the same type and heterogeneous graphs where nodes and edges belong to different types.However,in the real world,there are still a large number of multiplexed heterogeneous graphs with many different types of relationships between two nodes.Multiplex heterogeneous graphs have great importance in studying relationships of different edge types between two nodes.Many existing methods only focus on the heterogeneity of edge types between different nodes,and ignore the difference of multiplex edges between the same pair of nodes;other methods only consider the impact of different edge types on the graph structure,ignoring relationship among edges.These problems all lead to the inaccuracy of the embedding vector of the graph.Based on the problems of existing methods mentioned above,this paper proposes a multiplex heterogeneous graph embedding method based on intra-layer inter-layer integration.At the same time,a meta-path walking mechanism considering edge types is proposed.The work of the thesis is divided into three parts:(1)A multiplex heterogeneous graph embedding method based on intra-layer inter-layer integration is proposed.This paper introduces the method of multiplex heterogeneous graph layering.Different relationship types construct different relationship subgraphs.A intra-layer aggregation is used to capture the physical structure of the graph,and then the inter-layer random walk is conducted across the relationship sub-graphs,mining logical semantic information to make this method suitable for multiplexed heterogeneity.The graph realizes complete information conservation.(2)A meta-path walking mechanism considering edge types is proposed.In addition to solving the problem of incomplete graph structure information capture in multiplex heterogeneous graphs,this paper also designs a meta-path walking mechanism that considers edge types for multi-layer subgraphs.This paper uses semantic information with edge types to artificially guide the random walking process and to capture the interaction between edges and edges through semantics,extending the two-dimensional walking mechanism to three-dimensional space,ensuring the accuracy of the generated vector.(3)The feasibility and effectiveness of the above work are verified by experiments.Based on more than 2 million pieces of data on 4 data sets,the paper conducts experiments on the proposed scheme and existing algorithms,and analyzes the results.It is verified by experiments that embedding method and random walk mechanism proposed in the paper promote embedding ROC-AUC?PR?F1 accuracy by 1-27%?1-35%?1-27%respectively.