Research On Some Node Classification Methods Based On Graph Learning

With the enrichment and expansion of current social life,the structure presented by data also becomes more and more complex.How to effectively manage and apply these complex structured data has become a serious challenge for researchers.As a common data structure,Graph can describe complex connections between things,so more and more emerging technology fields begin to use graph structure to represent complex data.For example,in the field of biotechnology,graph structures are used to describe the internal structure of proteins.Inspired by Convolutional neural networks,the Graph Convolutional Network(GCN)model has been widely used in irregular Graph structure data and achieved good results in graph related tasks.However,a large number of studies have shown that when GCN uses original fixed graph structure data to perform graph convolution and other operations,the quality of graph structure directly affects the performance of GCN model to complete downstream tasks.For example,when the data itself has no graph topology,the previous work is usually based on simple distance measurement,which is difficult to mine valuable deep data information.For the data with graph structure,there will be inevitable errors in data collection,resulting in the collected data usually with noise and missing.Aiming at the problems existing in graph structure of graph convolution network model,this paper proposes two improvement strategies based on graph learning theory,Graph Laplacian rank constraint and low rank learning.(1)A graph convolutional network classification model based on graph Laplacian rank constraints is proposed.For data without its own graph structure,adaptive k nearest neighbor graph learning and graph Laplacian rank constraint are used to learn a high quality graph structure as the learning objectives of the GCN model to complete downstream tasks.Specifically,firstly,the data relationship is established through adaptive k nearest neighbor graph learning,and then a high quality graph structure with connected components consistent with the number of data categories is learned through the graph Laplacian rank constraint,and finally,the high quality graph structure is input into GCN Complete the node classification task.(2)A dynamic graph convolutional network classification model based on fusion graph learning is proposed.For itself with a diagram of the structure of the data,in order to reduce the noise generated by the data acquisition,stray adverse factors,such as at the same time,fully exert the characteristic information of the data mapping is proposed a fusion image convolution network classification model,respectively to the topology optimization and feature mapping data,the topology of the depth of mining data information and node characteristics,Furthermore,the two graphs are fused and the graph convolutional network is used to complete the downstream node classification task.In addition,the proposed model is an end to end structure that combines fusion graph learning and graph convolution in one framework and can dynamically learn the optimal fusion graph structure based on downstream tasks.In this thesis,we propose two methods to optimize the graph structure by focusing on the unusual sensitivity of the GCN model to the quality of the graph structure when completing downstream tasks,and the experimental results show that the proposed methods are effective.