Decomposition and Clustering-Based Many-Objective Optimization for Multi-Label Feature Selection

Somayeh Kanani, Parham Moradi, Mahdi Jalili

Abstract: In this paper a novel many-objective optimization algorithm called MaEA/DC is proposed. The proposed method utilizes the random objective division (ROD) strategy to decompose solutions across various objectives as a pivotal step, thereby enhancing search capabilities through simultaneous application of multiple decomposition methods. Moreover, to address the abundance of non-dominant solutions in many-objective problems, a density-based clustering method is employed to cluster the solutions along the Pareto front. The cluster centers are then selected, while eliminating the others to reduce the density of the Pareto front. This approach ensures the improvement of population convergence while maintaining its diversity. Experimental validation demonstrates our method effectively balances multiple objectives, eliminates irrelevant and redundant features, and achieving satisfactory classification results in comparison to other methods.