Automatic Data Clustering Framework Using Nature-Inspired Binary Optimization Algorithms

Abstract

Cluster analysis using metaheuristic algorithms has earned increasing popularity and acceptance over recent years due to the great success of these algorithms in finding high-quality clusters in complex real-world problems. This paper proposes a novel framework for automatic data clustering with the capability of generating clusters with approximately the same maximum distortion using nature-inspired binary optimization algorithms. The inherent problem with clustering using such algorithms is having a huge search space. Therefore, we propose a binary encoding scheme for the particle representation to alleviate this problem as well. The proposed clustering solution requires no prior knowledge of the number of clusters and proceed with the process based on re-clustering, merging, and modifying the small clusters to compensate for the distortion gap between groups with different sizes. The proposed framework's performance has been evaluated over a wide range of synthetic, real-life, and higher dimensional datasets first by considering four different binary optimization algorithms for the optimizer module. Then, it has been compared with multiple classical and new clustering solutions in addition to two automatic clustering techniques using continuous search space in terms of separation and compactness of the clusters by utilizing internal validity measures. The experimental results show the proposed solution is highly efficient in creating well-separated and compact clusters with approximately the same distortion. Moreover, the application of the proposed framework to the correlated binary data is also reported as a case study. The presence of correlation in the binary dataset results from the similarity between data in the same category, such as repeated measurements in remote sensing, crowdsourced multi-view video uploading, and augmented reality. The simplicity, customizability, and flexibility to add additional conditions to the proposed solution in addition to having a dynamic number of clusters are the advantages of the proposed framework.