Sorted Outlier Detection Approach Based on Silhouette Coefficient

Abstract

In this era when data is generated continuously in various domains of machine learning, different algorithms are budding to improve and enhance the learning process. Clustering is one of such machine learning techniques. It is considered to be most important tool of unsupervised learning but it is sensitive to outlier. Thus it is essential to remove the outlier before clustering the data. Most of the outlier detection techniques require some user-defined parameters, which make their accuracy user-dependent. Thus an algorithm which is least dependent on user-defined values is proposed here. The algorithm takes number of cluster in which user want to cluster its data and detect outlier within those clusters using Silhouette Coefficient. The algorithm was compared with some of the existing algorithm in domain of outlier detection. And the experimental analysis is performed on some relevant benchmark dataset presented in UCI repository. Through the experimental results it can be seen that the algorithm we have proposed has performed better than the existing algorithms.