Differential Evolution with Local Search Algorithms for Data Clustering: A Comparative Study

Abstract

Clustering is an unsupervised data mining task which groups objects in the unlabeled dataset based on some proximity measure. Many nature-inspired population-based optimization algorithms have been employed to solve clustering problems. However, few of them lack in balancing exploration and exploitation in global search space in their original form. Differential Evolution (DE) is a nature-inspired population-based global search optimization method which is suitable to explore the solution in global search space. However, it lacks in exploiting the solution. To overcome this deficiency, few literatures incorporate local search algorithms in DE to achieve a good solution in the search space. In this work, we have performed a comparative study to show effectiveness of local search algorithms, such as chaotic local search, Levy flight, and Golden Section Search with DE to balance exploration and exploitation in the search space for clustering problem. We employ an internal validity measure, Sum of Squared Error (SSE), to evaluate the quality of cluster which is based on the compactness of the cluster. We select F-measure and rand index as external validity measures. Extensive results are compared based on six real datasets from UCI machine learning repository.