Simplifying Graph Convolutional Networks

[zarina-aniraz]

Venue: ICML 2019
Summary: Proposes a simplified linear graph neural network architecture (GCN with non-linearity layers removed). New architecture is significantly faster than the state of the art models (i.e FastGCN) and scales to large datasets (Reddit).

Observation: The paper presents baseline results (speed and accuracy) of the contemporary graph neural networks and of the application of the model on different domains (text classification, semi-supervised user geolocation, relation extraction, zero-shot image classification, graph classification)

Links
Web: https://arxiv.org/pdf/1902.07153.pdf
GitHub: https://github.com/Tiiiger/SGC

[liqimai]

I think you would also be interested in our paper "Label Efficient Semi-Supervised Learning via Graph Filtering"(https://arxiv.org/abs/1901.09993), which share same ideas as the paper you mentioned. Hope it can provide some insight of GCN for you.

Our paper was accepted for publication in CVPR 2019 and an early version of this work submitted to ICLR 2019 (https://openreview.net/forum?id=SygjB3AcYX), in which we analyzed the “low-pass” effect of GCN and label propagation methods for semi-supervised learning. Our paper was online at the OpenReview website on Sept 28, 2018, several months before the SGC work (Simplifying GCN). SGC may be also inspired by ours ☺.