Learning Graph Convolution Filters from Data Manifold

Guokun Lai, Hanxiao Liu, Yiming Yang

Abstract: Convolution Neural Network (CNN) has gained tremendous success in computer vision tasks with its outstanding ability to capture the local latent features. Recently, there has been an increasing interest in extending CNNs to the general spatial domain. Although various types of graph convolution and geometric convolution methods have been proposed, their connections to traditional 2D-convolution are not well-understood. In this paper, we show that depthwise separable convolution is a path to unify the two kinds of convolution methods in one mathematical view, based on which we derive a novel Depthwise Separable Graph Convolution that subsumes existing graph convolution methods as special cases of our formulation. Experiments show that the proposed approach consistently outperforms other graph convolution and geometric convolution baselines on benchmark datasets in multiple domains.

TL;DR: We devise a novel Depthwise Separable Graph Convolution (DSGC) for the generic spatial domain data, which is highly compatible with depthwise separable convolution.

Keywords: Label Propagation, Depthwise separable convolution, Graph and geometric convolution

Data: [CIFAR-10](https://paperswithcode.com/dataset/cifar-10), [CIFAR-100](https://paperswithcode.com/dataset/cifar-100)