Graph Neural Network Expressivity and Meta-Learning for Molecular Property Regression
This work addresses the challenge of efficient learning in molecular property prediction for researchers, but it is incremental as it applies existing meta-learning methods to GNNs.
The paper tackled the problem of learning new molecular property regression tasks with few updates by applying Reptile meta-learning to Graph Neural Networks (GNNs), resulting in improved performance and rapid convergence compared to randomly initialized GNNs, with GNN ensembles yielding the best results.
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.