Improving the performance of fuzzy clustering algorithms through outlier identification

Abstract

Major clustering algorithms consider all data objects as good objects while dividing data-set into clusters, except some, that consider noise/outliers to some extent. As a result those algorithms are not capable to produce efficient clusters as there is some effect of noise on location of cluster centroids. The task of outlier identification is to find small groups of data objects that are exceptional when compared with rest large amount of data. They are not required or acceptable while dividing a data–set into clusters, as clusters refer to the similar group of data and these outliers don't belong to any of the similar group. Yet they can be important in other applications. Through this paper we are trying to prove that efficient clusters can only be produced by identifying outliers and separating them from the data-set into one cluster before applying any clustering algorithm. In this paper a density based algorithm for outlier identification is proposed. Before applying any of the clustering algorithms; proposed algorithm is applied on the data-set to identify outliers and separate them from original data-set. Proposed algorithm is applied on fuzzy clustering algorithms (FCM, PCM and PFCM). Numerical examples and tests show that fuzzy algorithms after applying proposed algorithm gives better results when compared with the performance of fuzzy clustering algorithms without applying proposed technique.