Graph convolutions that can finally model local structure

30 Nov 2020  ·  Rémy Brossard, Oriel Frigo, David Dehaene ·

Despite quick progress in the last few years, recent studies have shown that modern graph neural networks can still fail at very simple tasks, like detecting small cycles. This hints at the fact that current networks fail to catch information about the local structure, which is problematic if the downstream task heavily relies on graph substructure analysis, as in the context of chemistry. We propose a very simple correction to the now standard GIN convolution that enables the network to detect small cycles with nearly no cost in terms of computation time and number of parameters. Tested on real life molecule property datasets, our model consistently improves performance on large multi-tasked datasets over all baselines, both globally and on a per-task setting.

PDF Abstract


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Graph Property Prediction ogbg-molpcba GINE+ w/ APPNP Test AP 0.2979 ± 0.0030 # 10
Validation AP 0.3126 ± 0.0023 # 5
Number of params 6147029 # 9
Ext. data No # 1
Graph Property Prediction ogbg-molpcba GINE+ w/ virtual nodes Test AP 0.2917 ± 0.0015 # 14
Validation AP 0.3065 ± 0.0030 # 10
Number of params 6147029 # 9
Ext. data No # 1