Evaluation of Different Heuristics for Accommodating Asymmetric Loss Functions in Regression

Abstract

Most machine learning methods used for regression explicitly or implicitly assume a symmetric loss function. However, recently an increasing number of problem domains require loss functions that are asymmetric in the sense that the costs for over- or under-predicting the target value may differ. This paper discusses theoretical foundations of handling asymmetric loss functions, and describes and evaluates simple methods which might be used to offset the effects of asymmetric losses. While these methods are applicable to any problem where an asymmetric loss is used, our work derives its motivation from the area of predictive maintenance, which is often characterized by a small number of training samples (in case of failure prediction) or monetary cost-based, mostly non-convex, loss functions.