Improving the performance of visualized clustering method

Abstract

In data domains, the process of clustering is expressed as exploratory data analysis in which similar objects can be grouped as subsets according to the properties of a cluster. Discovering the number of clusters is an important issue in clustering. It is noted that k-means gives poor clustering results when the user attempts an incorrect ‘k’ value. The visual access tendency (VAT) is a widely used technique for discovering the number of clusters. Recently, Bezdek et al. introduced extended ideas of VAT such as SpecVAT, and iVAT. The SpecVAT uses spectral approach and produces accurate clustering results than VAT. The limitation of SpecVAT is that it unables to solve the clustering tendency problem for path-based clustered data. The iVAT technique solves this issue. These techniques use an Euclidean space for dissimilarity matrix computation. In this paper, we use a multi-view point based similarity (MVS) cosine metric for achieving robust results. We present two proposed methods, namely, cSpecVAT and GMMMVS-VAT. The cSpecVAT is developed by cosine metric and spectral concepts and it extracts efficient clustering results over the comprehensive datasets such as synthetic, real, genetic and image. For audio datasets, there is another method proposed called as GMMMVS-VAT, which includes the following steps: modelling the speech data by Gaussian mixture model (GMM), and MVS for extracting the similarity features as reference to multi-view points; hence, it works more effectively on speech datasets. In MVS, we use a number of view-points as reference making it more robust than a single view-point approach.