Performance Analysis of Improved Clustering Algorithm on Real and Synthetic Data

Clustering is an important technique in data mining to partition the data objects into clusters. It is a way to generate groups from the data objects. Different data clustering methods or algorithms are discussed in the various literature. Some of these are efficient while some are inefficient for large data. The k-means, Partition Around Method (PAM) or k-medoids, hierarchical and DBSCAN are various clustering algorithms. The k-means algorithm is more popular than the other algorithms used to partition data into k clusters. For this algorithm, k should be provided explicitly. Also, initial means are taken randomly but this may generate clusters with poor quality. This paper is a study and implementation of an improved clustering algorithm which automatically predicts the value of k and uses a new technique to take initial means. The performance analysis of the improved algorithm and other algorithms by using real and dummy datasets is presented in this paper. To measure the performance of algorithms, this paper uses running time of algorithms and various cluster validity measures. Cluster validity measures include sum squared error, silhouette score, compactness, separation, Dunn index and DB index. Also, the k predicted by the improved algorithm is compared with optimal k suggested by elbow method. It is found that both values of k are almost similar. Most of the values of validity measures for the improved algorithm are found to be optimal.