Graph convolutions that can finally model local structure
arXiv:2011.15069
Abstract
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.
References in corpus (13)
- Neural Message Passing for Quantum Chemistry
- Graph Neural Networks: A Review of Methods and Applications
- Junction Tree Variational Autoencoder for Molecular Graph Generation
- Open Graph Benchmark: Datasets for Machine Learning on Graphs
- MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing
- DeeperGCN: All You Need to Train Deeper GCNs
- Strategies for Pre-training Graph Neural Networks
- Invariant and Equivariant Graph Networks
- Learning Graph-Level Representation for Drug Discovery
- GraphNorm: A Principled Approach to Accelerating Graph Neural Network Training
- Relational Pooling for Graph Representations
- Hierarchical Inter-Message Passing for Learning on Molecular Graphs
- Can Graph Neural Networks Count Substructures?