Research And Application Of Graph Neural Networks With Edge Attributes

In recent years,with the continuous development of deep learning technology,astonishing results have been achieved in many research fields.At the same time,more researchers began to pay attention to a type of special structure data,that is,non-Euclidean data,as in graph structure data.This type of data can be seen everywhere in life,such as knowledge graphs,social networks,chemical molecules,etc.However,due to its own constraints,it is difficult to solve related problems by traditional deep learning models,thus Graph neural networks have been proposed.However,most of the current graph neural network models focus on modeling node representations,while ignoring attributes on edges such as edge label and semantic information,resulting in information loss.Therefore,this work mainly studies how to make better use of edge attributes to improve the feature extraction and expression capabilities of graph neural networks.This research is based on Graph Neural Network models with edge enhanced,The main work and innovations are as follows:(1)A graph neural network with edge enhanced under the message passing framework is proposed.For graphs with labels or dense attributes on the edges,the(source node,edge,target node)triples are regarded as short sequences,and a sequence representation function is used to extract the sequence representation as the information passed from the source node to the target node.The passed information helps the target node to update the representation of itself.Our method is called graph neural network with edge enhanced,which can be applied to most graph neural network models based on the message passing framework,and we gives the realization of graph convolutional neural network model and graph attention neural network model.We performed both node classification and graph classification experiments on multiple datasets and achieved comparable results.(2)An aspect-based sentiment analysis model based on sgraph neural network with edge enhanced is proposed.For aspect-based sentiment analysis tasks,by introducing dependency syntax trees and sequence-enhanced graph neural network models,we propose an aspect-level sentiment analysis model based on sequence representation enhancement,which implements aspect-based sentiment analysis.The aspect based feature extraction and sentiment analysis have achieved good results on the public data set.(3)A random walk algorithm with edge attributes is proposed.Unlike the classic random walk algorithm that only focuses on the representation of nodes,we propose to add edges to the path of random walks,to obtain the initial representations of nodes and edges at the same time through the improved Deep Walk or Node2vec model,as an effective graph Pre-training methods on graphs.