Using Explainable Machine Learning to Interpret the Effects of Policies on Air Pollution: COVID-19 Lockdown in London

Abstract

When estimating discrete choice models, the prospect of using ever-larger datasets is limited by the poor scalability of maximum likelihood estimation. This paper proposes a simple and fast dataset reduction method that is specifically designed to preserve the richness of observations originally present in a dataset, while reducing its size. Our approach leverages locality-sensitive hashing to create clusters of similar observations, from which representative observations are then sampled and weighted. We demonstrate the efficacy of our approach by applying it on a real-world mode choice dataset; the obtained results confirm that a carefully selected and weighted subsample of observations is capable of providing close-to-identical estimation results while being, by definition, less computationally demanding.