Local And Global Structure Information Is Preserved For Graph Represents Learning Research

As the data generated in the industrial era increasingly exceeds the storage capacity of traditional databases and devices,how to store data more efficiently has become a hot topic in academic research.Graph representation learning is an effective approach to data information mining,which can represent hundreds of millions of data as simple vertex embedding vectors,and can be used for downstream tasks.The key to graph representation is to learn the local and global structural information of a given graph,and integrating structural information with different hop counts is an effective strategy.Existing research has shown that transition probability can reflect the relationship between vertices with different hops to a certain extent,and can obtain local and global structure information of a graph through the calculation of probability matrices.However,many current efforts simply connect representations of different hops into a graph representation.In other words,existing research is unclear about the contribution of each k-step/hop(k≥1)structural information in the graph representation.Obviously,as the distance between vertices increases,their mutual influence will weaken.Focusing on this issue,the main work of this study is as follows:(1)The existing shallow model methods are difficult to capture the nonlinear relationships of graph structures,and the graph convolution model in graph neural network technology can cause over-smooth problems.This paper proposes a deep learning model based on T(T>1)feedforward neural networks,which extracts nonlinear structures from graph structures through a deep model.T sub-models effectively capture local and global(higher order)relational information of the graph,and these models are given different contributions in the final vector representation,thereby leveraging the advantages of different local and global relational information.In experiments,experimental results in two tasks,link prediction and vertex classification,have shown that the framework is more competitive than the baseline method,and the comparison benchmark algorithm can achieve an improvement of about 20%.(2)A unified framework based on matrix decomposition is proposed,which flexibly considers the contributions of different local information in graph representation.This method reconsiders the contribution of different hop counts in the embedded representation from two perspectives:(i)flexibly assigning different weights to different step loss functions,and(ii)integrating the vector representations of all K steps into one output graph representation based on different scales according to different distances(k values)of vertex pairs.Based on this,this method will more effectively obtain local and global structure information of the graph.The framework has proven to achieve competitive performance in vertex classification,link prediction,and visualization tasks on multiple datasets.