GaCoVi: a Correlation Visualization to Support Interpretability-Aware Feature Selection for Regression Models

Abstract

The recent growth of interest in explainable artificial intelligence (XAI) has resulted in a large number of research efforts to provide accountable and transparent machine learning systems. Although a large volume of research has focused on algorithm transparency, there are other factors that influence the interpretability of a system, such as end-users' understanding of individual features and the total number of features. Thus, involving end-users in the feature selection process may be key to achieving interpretability. In addition, previous work has suggested that to obtain satisfactory interpretability and predictive performance, the feature selection process should look for a subset of features that are highly correlated with the response variable yet uncorrelated to each other. Taking this into account, in this paper, we present a work-in-progress design study of a novel system for correlation visualization, GaCoVi. GaCoVi is designed to put domain experts in the loop of feature selection for regression models in scenarios where transparency of the machine learning systems is crucial.