Research On Multi-Level Graph Embedding Methods And Applications For Large-Scale Graphs

With the development of the Internet and big data,graphs have emerged as an important data representation structure in various real-world scenarios.Graph embedding,as a method of graph analysis,encodes nodes and edges of a graph into low-dimensional vectors while preserving the graph structure information.However,when it comes to largescale graphs,most graph embedding methods suffer from issues such as low embedding quality,high computational complexity,excessive memory consumption,over-smoothed training models,and ignoring time attributes.Additionally,the dynamic nature of graphs has also brought new challenges to graph embedding.The large-scale graph embedding method based on the graph embedding framework greatly shortens running time while improving the quality of the embedding.In addition,adding a time attribute to graph embedding provides a new research direction for embedding in large-scale dynamic graphs.To address the issues that arise during the process of large-scale graph embedding,this paper conducts research on large-scale graph embedding methods and their applications,which are elaborated as follows:(1)Research on Large-Scale Static Graph Embedding Method Based on Multi-level Graph Embedding Refinement Framework.To address the problem of low efficiency and long running time in large-scale graph embedding,this paper proposes a large-scale static graph embedding method based on the multi-level graph embedding refinement framework.This method uses a spectral distanceconstrained coarsening algorithm to repeatedly coarsen the graph into smaller-scale graphs,which preserves the main structure of the graph and computes a basic embedding on the coarsest graph to improve the efficiency of graph embedding.This article improves the graph convolutional neural network model by increasing the initial value and identity mapping,refining the embedding from the coarsest graph to the original graph through iteration.Experimental results and analysis show that this method improves over-smoothing and improves the quality of graph embedding.(2)Research on Large-Scale Dynamic Graph Embedding Method Based on Continuous-Time Graph Embedding Framework.To address the dynamic nature of large-scale graphs,this paper proposes a large-scale dynamic graph embedding method based on the continuous-time graph embedding framework.This method uses graph partitioning algorithms to partition large-scale graphs with time attributes into multiple subgraphs,and performs continuous-time subgraph embeddings on each subgraph based on spatiotemporal walks.Effective walks are represented by continuous node sequence lists,where the nodes themselves have time attributes and the edges are connected by non-decreasing time information between nodes.The experimental results and analysis show that this method uses global aggregation to find a global vector space in time,maps multiple local subgraph embedding spaces,and improves the quality of embedding large-scale continuous-time dynamic graphs.(3)Application Research of Ethereum Transaction Data Based on Continuous-Time Weighted Graph Embedding.In response to the characteristics of the Ethereum transaction network,this paper models the Ethereum transaction network as a time-weighted graph,where each node represents a unique Ethereum account,and each edge represents a transaction weighted by the amount and assigned with a timestamp.This paper proposes a method based on continuous-time weighted graph embedding,which captures more comprehensive properties in the transaction network by utilizing the time information of edges,analyzes the legality of Ethereum transaction data,and improves the security of transaction data.Experimental results and analysis show that the proposed method not only improves the accuracy of transaction network analysis but also shortens the training time to meet the application requirements of Ethereum transaction data.Finally,the paper summarizes the main research content and prospects for future research directions based on the multi-level graph embedding method and its application on large-scale graphs.