Design of Single and Multi objective Metaheuristic Algorithms for Effective Data Clustering

Abstract

Clustering is an important data analysis technique to find similar data objects in a given dataset. It is unsupervised learning and has proven its capability in diverse research fields such as medical diagnosis, market segmentation, image segmentation, customer behaviour analysis, outlier detection, and feature selection. Clustering aims to determine the set of identical data objects and put these data objects into a single cluster. The data objects within the clusters have more similar characteristics than other clusters. The research community presents several clustering techniques- partitional, hierarchal, model-based, grid-based, density-based, etc. But, the popular one is partitional clustering. This thesis work focuses on partitional clustering. In partitional clustering, a dataset divides into k number of partitions known as clusters. A distance function is utilized for allocating the data objects to clusters based on minimum distance. However, the number of clusters (k) should be known in advance. The performances of partitional clustering algorithms depend on the selection of initial cluster centroids. Several traditional algorithms, like K-Means, K-Mediods, K-Harmonic Mean etc., are successfully implemented for solving partitional clustering problems. But, these algorithms have several drawbacks, such as being sensitive to initial cluster selection, local optima, convergence rate and predefined method for updating cluster centroids. Several researchers explore metaheuristic algorithms capabilities to overcome the issues of traditional clustering algorithms. These are GA, PSO, ACO, ABC, TS, SA etc., and provide state-of-the-art clustering results for partitional clustering problems. However, some issues are also associated with metaheuristic algorithms, such as an imbalance in local search and global search mechanisms, population diversity, sometimes stuck in local optima, and population generation.