Lighting the black box: explaining individual predictions of machine learning algorithms

Abstract

Many machine learning techniques remain ''black boxes'' because, despite their high predictive performance, it is difficult to understand the role of each variable involved in the prediction task. In this thesis, we will study three methods that explain individual predictions of any model and determine what explanatory variables are most influential for each particular observation, which brings transparency to machine learning algorithms. Additionally, we will test these methods on a simple dataset for which we can assess the quality of the explanations.