Impact of new seed and performance criteria in proposed rough k-means clustering

Rough k-means algorithm is one of the widely used soft clustering methods for clustering. However, the rough k-means clustering algorithm has certain issues like it is very sensitive to random initial cluster centroid and subjectivity involved in fixing the constant value of zeta parameter to decide the fuzzy elements. Also, there are no appropriate performance measures for the rough k-means algorithm. This study proposes a new initialization algorithm to address the issue of random allocation of data elements to improve the performance of the clustering. A new method to determine the best zeta values for the rough k-means algorithm is evolved. Also, new performance criteria are introduced, such as distance between the centroid and the farthest element in the cluster, number of elements in the intersection area and distance between the centroid of clusters. The seven experiments were carried out by using several benchmark datasets such as Microarray, Synthetic and Forest cover datasets. The performance criteria were compared across the Proposed algorithms, Peters (k-means++), Peters PI, Ioannis' algorithm, Vijay algorithm Mano algorithm, and Peters (random) initialization methods. The Root mean square standard deviation (RMSSTD) and S/T index were used to validate the performance of the rough k-means clustering algorithm. Also, the frequency table was constructed to gauge the degree of fuzziness in each algorithm. The program completion time of each initialization algorithm was used to measure the variability among the proposed and existing rough k-means clustering algorithms. It was found that the initialization algorithms excelled in all criteria and performed better than the existing rough k-means clustering algorithms.