Graph Neural Network Expressivity and Meta-Learning for Molecular Property Regression

TL;DR

This paper explores how meta-learning algorithms like Reptile enhance GNNs for molecular property regression, demonstrating improved efficiency and performance with expressive layers and ensembles in few-shot learning scenarios.

Contribution

It introduces the application of model-agnostic meta-learning to GNNs in molecular regression and shows how layer expressivity and ensembles improve rapid task adaptation.

Findings

Meta-learning enables quick adaptation to new chemical tasks with few updates.

GNN layer expressivity correlates with better meta-learning performance.

Ensembles of GNNs achieve superior accuracy and faster convergence in k-shot learning.

Abstract

We demonstrate the applicability of model-agnostic algorithms for meta-learning, specifically Reptile, to GNN models in molecular regression tasks. Using meta-learning we are able to learn new chemical prediction tasks with only a few model updates, as compared to using randomly initialized GNNs which require learning each regression task from scratch. We experimentally show that GNN layer expressivity is correlated to improved meta-learning. Additionally, we also experiment with GNN emsembles which yield best performance and rapid convergence for k-shot learning.