Ensemble Based Graph Convolutional Network for Semi Supervised Learning

Abstract

AbstractIn this paper, an Ensemble based Graph Convolutional Network (EGCN) is proposed for relational datasets. EGCN encodes the pairwise similarity between data points for semi supervised learning and feature representation. The GCN model with Laplacian filter is not able to completely explore the local geometry of the data distribution as the null space of Laplacian remains constant along the manifold. This leads to poor extrapolating ability of the GCN model. In order to improve this GCN model, Hessian filter is introduced. The Hessian filter is more efficient and has richer null space. It can encode the relationship among the data points better by fusing structural properties with original features. In our EGCN, two parallel GCNs are executed, one with Laplacian filter and other with Hessian filter. In EGCN model Laplacian, filter accumulate the information from a data point to all its neighbors and Hessian filter gathers the information between adjacent data points that lie in the neighborhood of a data point. Extensive experiment conducted on various relational datasets indicates that EGCN out performs both Laplacian and Hessian filter based GCN models by a significant margin on various relational datasets.KeywordsGraph Convolutional NetworkSemi supervised learningLaplacianHessianRelational dataset