A diversity maximizing active learning strategy for graph neural network models of chemical properties

Abstract

This paper presents a diversity-maximizing strategy for actively constructing a compact molecule set for training graph neural network molecular property models. In particular, we consider the core-set selection problem, viz., finding a training set S that is (1) representative and (2) a subset of a pre-defined space U of interest. The strategy iteratively adds new molecules into S so that the its diversity is maximized (in a greedy way) with respect to U; the diversity itself is determined from a Euclidean distance metric of a feature vector that is extracted from the graph neural network model at that iteration. We apply this strategy to retrospectively construct compact training sets for a number of experimental and computed molecular properties and show that it outperforms random sampling of U in almost all cases. Random sampling and the proposed active learning strategy, however, perform similarly for the QM7 (computed heat of atomization) dataset; further inspection using data visualization and analysis indicates that this is attributable to the manner in which the molecule set was created to maximize functional group diversity. Our method, in general, is property agnostic and does not require the calculation of prediction uncertainty at each iteration.