Fast derivation of Shapley based feature importances through feature extraction methods for nanoinformatics

Abstract

This work presents an alternative model-agnostic attribution method to compute feature importance rankings for high dimensional data requiring dimension reduction. We make use of Shapley values within the Shapley additive explanation framework to determine the importance values of each of the feature in the data set. We then demonstrate that it is possible to significantly reduce the computational complexity of ranking features in high dimensional spaces by first applying principal component analysis. This transformation into lower dimensional spaces in conjunction with our normalisation approach does not yield a significant loss of information when performing feature selection tasks beyond a threshold. The efficacy of our approach is demonstrated on several examples of nanomaterial data, in particular graphene oxide. Our approach is ideal for the applied physical science communities where datasets are of high dimensionality and computational complexity is a matter for concern.