Feature selection in generalized additive models with metaheuristics

Abstract

In supervised machine learning our aim is to predict a well-defined target variable as accurately as possible by utilizing the known values of several feature variables. Nowadays many complex algorithms are available to solve this task. On the other hand, algorithms that provide the most accurate estimates of the target variable are usually poor at determining marginal effects of the feature variables to the target. However, in certain practical applications, the most important result of supervised learning is not necessarily the accurate estimation of the target, but the discovery of each feature's marginal effect. For example, a bank has to offer a clear reasoning when declining a credit application. In our current big data environment, when the number of possible features is large, determining marginal effects can be challenging even for a linear regression model. One tool that can be utilized to make supervised learning models more interpretable is feature selection.