A New Approach to Cluster Datasets without Prior Knowledge of Number of Clusters

Abstract

The paper discusses yet another approach of clustering datasets whose cluster numbers are not known beforehand. The suggested approach effectively determines the number of clusters or partitions while running the algorithm. The proposed method is only limited to partitional clustering inspired from the K-means algorithm. In this work a Modified Teaching-Learning-Based Optimization (MTLBO) is used to form the clusters and determine the number of clusters on the run. The comparison of the results obtained by MTLBO is done with the classical TLBO and Classical Differential Evolution (DE) technique. The results show that MTLBO gives better accuracy than the other two with respect to the number of function evaluations and cluster validity measures. Several benchmark datasets are simulated from the UCI machine repository and results are tabulated in the paper.