Subspace Discovery through Evolutionary Multi-objective Optimization in Overlapping Clustering

Abstract

This paper employs a multi-objective optimization technique to concurrently partition data into multiple overlapping subspace clusters. Simultaneously, the data grouping and the identification of relevant subspace feature sets corresponding to these groups are performed. The study utilizes validity indices, including the ICC-index, PSM-index, and a novel MNR-index, where the latter optimizes the overlapping of objects into distinct clusters. Furthermore, existing mutation operators such as large deletions, large duplications, and large translocations are adapted to enhance the exploration of the search space effectively. The proposed method is tested on ten standard real-life datasets and sixteen synthetic datasets to identify diverse overlapping subspace clusters. Comparative analyses with existing methods highlight the advantages of incorporating multiple objectives and the newly defined objective function, demonstrating superior performance in the majority of cases. Additionally, the paper illustrates the application of this method in the bi-clustering of gene expression profile data, showcasing its versatility and efficacy across different domains.