An Era Of Enhanced Subspace Clustering In High-Dimensional Data
RAMA DEVI J · i-manager s Journal on Computer Science · 2016
Clustering is a descriptive task in the data mining process which aims at grouping similar data sets into clusters (A.K. Jain, et al., 1988; Berkhin P, 2006) . Traditional clustering techniques were designed to deliver the clusters in fulldimensional original space, by estimating the proximity between the data objects using all of the attributes of a full data set (Zhang T, et al., 1996; Han J, et al., 2011) . Now-adays, as the dimensionality of data increases in advances in data collection management. But as the attributes of data increases, some of the attributes become irrelevant to some of the cluster groups. These traditional clustering algorithms are inefficient for high-dimensional data. The high dimensional data suffers several problems from the curse of dimensionality (Houle ME, et al., 2010) . A two fold problem determines the summarization of the curse of dimensionality with respect to clusters. (i) The dimensionality of data grows, several dimensions may not be relevant to define cluster and may, the distance computations usually performed in full dimensional space. The cluster may form in subset of attributes, but it may not be possible to detect the clusters perfectly with the full dimensional space. (ii) The subset of dimensions related to some other cluster may be different from the subset of dimensions relate to another cluster. These two clusters may differ from the relevant subspace for a third cluster. So it may be defined as overlapping clusters in high-dimensional data, i.e. one data point belongs to cluster C in a certain subset, but the d1 cluster C in another subset. This is not accounted in d2 traditional clustering methods such as partitioning and density based algorithms. Figure 1 represents Existence of redundant and irrelevant attributes (Aggarwal R, et al., 1998).