Abstract

Graph Neural Networks have been extensively used to learn non-Euclidian structures like graphs. There have been several attempts to improve the training efficiency and to reduce the learning complexity in modelling of large graph datasets. In this paper we have reviewed the approaches which perform convolutions to model large graphs for classification and prediction. We have critically analysed each of these approaches and veracity of their claims of reduced complexity and have reported their shortcomings. We have further analysed the approaches from graph-dataset perspective.