On the Suitability of Combining Feature Selection and Resampling to Manage Data Complexity

Abstract

The effectiveness of a learning task depends on data complexity (class overlap, class imbalance, irrelevant features, etc.). When more than one complexity factor appears, two or more preprocessing techniques should be applied. Nevertheless, no much effort has been devoted to investigate the importance of the order in which they can be used. This paper focuses on the joint use of feature reduction and balancing techniques, and studies which could be the application order that leads to the best classification results. This analysis was made on a specific problem whose aim was to identify the melodic track given a MIDI file. Several experiments were performed from different imbalanced 38- dimensional training sets with many more accompaniment tracks than melodic tracks, and where features were aggregated without any correlation study. Results showed that the most effective combination was the ordered use of resampling and feature reduction techniques.