A Novel Clustering Techniques Evaluation using Monte Carlo Simulation

Abstract

In Machine Learning clustering is one the most significant method. Today, we have data in rich from many sources but in order to get meaningful information from it is very boring task. Machine learning clustering algorithms to create cluster to decode the meaningful information from the data, this analysis approach has gained much popularity in recent years. This paper explores evaluation performance of frequently used existing clustering techniques such as single linkage, complete linkage, average linkage, centroid, and Ward’s method based on the proximity measures like Euclidean distance, Minkowski distance, Manhattan distance, Maximum distance, Correlation based distance and Canberra distance. Together with other commonly used clustering techniques such as SOM, Fuzzy C-means, Partitioning Around Medoids (PAM), model-based clustering, K-means, Kernel K-means, Robust K-means as well as our newly proposed technique K-HMs (K-Harmonic Means) are applied to decide the most suitable method for the identification of homogeneous items. Stability of the cluster is also tested based on the measures of performance evaluation such as Recall/Sensitivity, Precision, Accuracy and F-Score. We have also checked the performance followed by ROC curve. We have simulated the two types of data sets. It is to be noted that evaluations using Monte Carlo simulations show that our proposed method K-HMs is almost always reliable for clustering homogeneous items based on both types of datasets. Alternatively other methods correctly identify homogeneous items when equal variance but not unequal variances.