Indian Journal of Science and Technology

Abstract

Grouping of high dimensional information is an imperative exploration subject in the information mining, in light of the fact that the genuine datasets frequently have high dimensional components. The objective of the clustering is to group the features which should be similar to each other. Many text mining approaches are optimized to mine the sparse data which incurs high computation cost. In this paper, we process a novel technique named as affine subspace clustering which incorporates the Hubness property to handle the local feature relevance value and Curse of dimensionality. The Hubness property reduces the discrimination problem in the cluster formation and used as clustering method with effects relevant to cluster structures. Rather than endeavoring to keep away from the scourge of dimensionality by watching a lower dimensional component subspace, we use substantial dimensionality by exploiting downward closure property and outlier detection in the k nearest neighbor list. Additionally we combine Feature weighting method to minimize the average inside cluster scattering and augment the average between cluster scatterings along all the element spaces. The experimental results prove that proposed system yields the good performance in numerous settings, especially within the sight of huge amounts of commotion. The proposed techniques are optimized for the most part to detect the cluster center accuracy and extended properly to handle clusters of random sizes. Average inside cluster scattering is minimized and average between-cluster scattering is expanded along all the element spaces.

Keywords: Clustering, Curse of Dimensionality, Dynamic Centroid