Semi-supervised Learning Based On Signed Graphs

Semi-supervised learning is an important branch of machine learning,which relies on a small number of labeled samples to complete the learning task and ensures better learning performance while reducing resource consumption,and is a research area that has received widespread attention.As the research object of graph theory,a branch of mathematics,graphs have a mature theoretical foundation.There are many studies on machine learning models based on graphs.Compared with unsigned graphs,Signed graphs can better express the widespread negative dyadic relationships in the real world with the help of negative edges.Although Signed graphs have some advantages,most current studies are based on unsigned graphs.Convolutional neural networks have achieved impressive results,but they are difficult to migrate to non-Euclidean structured data.There exists a large amount of non-Euclidean structured graph data in real life,such as social networks,etc.Graph Convolutional Networks(GCN)implement the definition of graph convolution and concise and effective hierarchical propagation rules on non-Euclidean structured graph data,and work has been done to introduce them to semi-supervised learning.Signed Graph Convolutional Networks(SGCN),on the other hand,aggregate and propagate information among the layers of Signed graph models with the help of balance theory,and design hierarchical propagation rules for node information incorporating negative edge information.This paper focuses on semi-supervised learning based on signed graphs,and the main work includes the following two aspects.(1)In recent years,some Signed graph Laplacian clustering models have emerged,but there is no comprehensive summary analysis on them at present.This paper summarizes a variety of Signed graph Laplacian matrices and their variants,constructs Signed graphs based on semi-supervised information through pairwise constraints,converts the semi-supervised learning problem into a Signed graph model solution problem,and solves it on unstructured data and image data,and performs an analytical comparison.And the effectiveness and negative side benefits of semi-supervised learning based on Signed graph Laplacian are verified on unstructured data and image data respectively.(2)SGCN implements deep learning on the structure data of Signed graphs.At present,no relevant literature has been found to introduce SGCN into semi-supervised learning.In this paper,we introduce SGCN into semi-supervised learning,Propose a semi-supervised learning method based on SGCN.This paper verifies the improvement of the performance of SGCN by the additional effect of semi-supervised learning and negative edges,and the effectiveness of semi supervised learning based on SGCN is proved.