A multi-objective evolutionary feature selection approach for the classification of multi-label data

Abstract

There has been a growing need to resolve the Curse Of Dimensionality (COD) in multi-label data in recent years, which has attracted much attention for Feature Selection (FS). The multi-label FS method has gotten a lot of interest since it can considerably increase classification accuracy by selecting important features. Here, we have designed a binary version of multi-objective FS approach for Multi-Label Classification (MLC) based upon Whale Optimization Algorithm (WOA). Instead of a random search in WOA, we have applied the tournament search for the selection of a new whale. Tournament Selection (TS) entails holding a series of "tournaments" among a select group of people randomly from the general population. Here the multi-objective criteria consist of two objectives; where the first objective is to maximize the Jaccard similarity, and the another is to reduce the selected features. To check the robustness of the proposed method, we have used multi-label datasets from different areas. We added a comparative analysis of the proposed methodology with various traditional machine learning and multi-label classifiers. Empirical results on widely used multi-label datasets show that proposed FS achieves competitive performance, especially when labels are limited.