Feature Selection on Epistatic Problems Using Genetic Algorithms with Nested Classifiers

Feature selection is becoming an essential part of machine learning pipelines, including the ones generated by recent AutoML tools. In case of datasets with epistatic interactions between the features, like many datasets from the bioinformatics domain, feature selection may even become crucial. A recent method called SLUG has outperformed the state-of-the-art algorithms for feature selection on a large set of epistatic noisy datasets. SLUG uses genetic programming (GP) as a classifier (learner), nested inside a genetic algorithm (GA) that performs feature selection (wrapper). In this work, we pair GA with different learners, in an attempt to match the results of SLUG with less computational effort. We also propose a new feedback mechanism between the learner and the wrapper to improve the convergence towards the key features. Although we do not match the results of SLUG, we demonstrate the positive effect of the feedback mechanism, motivating additional research in this area to further improve SLUG and other existing feature selection methods.