A novel approach to use Multi-Armed Bandit for feature selection

Abstract

This work explores the application of Multi-Armed Bandit (MAB) algorithms for feature selection (FS) in machine learning, aiming to address the challenges posed by high-dimensional data, such as computational complexity and overfitting. While traditional FS methods are widely used, the integration of MAB in this context remains unexplored. This research proposes novel MAB-based algorithms, specifically adapting the Epsilon-greedy (MAB-EgreedyFS) and Upper Confidence Bound (MAB-UCBFS) algorithms, to dynamically manage feature inclusion and exclusion. In this way, the feature set is formed with the aim of providing the best accuracy for the classifier in the classification task. For this, each feature’s status is treated as an "arm" in the bandit problem approach that abstracts the search for the best features as the exploration-exploitation dilemma. During the process a Support Vector Machine (SVM) is applied as the classifier to evaluate the methods. A experimental set was performed, the proposed methods were evaluated in seven dataset and compared against established FS methods: SVM-RFE, extra-trees-based method, genetic algorithm-based method a ANOVA filter method. The results indicate that MAB-UCBFS consistently achieved strong performance, notably ranking as the "Best Method" for most of the evaluated data sets. While not universally superior, MAB-UCBFS demonstrated robust and competitive performance across most scenarios. Statistical analysis using Conover test heatmaps further corroborated these findings, highlighting significant differences between MAB-UCBFS and other techniques on several datasets. This study successfully validates the viability and strong performance of MAB-based algorithms, particularly MAB-UCBFS, as innovative and effective solutions for feature selection.