Geometric Graph Convolutional Neural Networks

  • 2019-09-11 18:59:41
  • Przemysław Spurek, Tomasz Danel, Jacek Tabor, Marek Śmieja, Łukasz Struski, Agnieszka Słowik, Łukasz Maziarka
  • 46

Abstract

Graph Convolutional Networks (GCNs) have recently become the primary choicefor learning from graph-structured data, superseding hash fingerprints inrepresenting chemical compounds. However, GCNs lack the ability to take intoaccount the ordering of node neighbors, even when there is a geometricinterpretation of the graph vertices that provides an order based on theirspatial positions. To remedy this issue, we propose Geometric GraphConvolutional Network (geo-GCN) which uses spatial features to efficientlylearn from graphs that can be naturally located in space. Our contribution isthreefold: we propose a GCN-inspired architecture which (i) leverages nodepositions, (ii) is a proper generalisation of both GCNs and ConvolutionalNeural Networks (CNNs), (iii) benefits from augmentation which further improvesthe performance and assures invariance with respect to the desired properties.Empirically, geo-GCN outperforms state-of-the-art graph-based methods on imageclassification and chemical tasks.

 

Quick Read (beta)

loading the full paper ...