Informatics Publishing Limited

R. Tamilselvan ¹, V. Hariharaprabu ¹, R. Bhaskaran ¹, C. Palanisamy ²

Affiliations
1 Department of IT, Vivekananda Institute of Engineering and Technology for Women, Elayampalayam, Namakkal-637205, Tamilnadu, IN
2 Department of IT, Bannari Amman Institute of Technology, Sathyamangalam, Erode-637205, Tamilnadu, IN

Abstract

This paper proposes a new innovative algorithm is called Visual assessment of cluster tendency, uses a visual approach to find the number of clusters in data. The Visual Assessment of (cluster) Tendency (VAT) method readily displays cluster tendency for small data sets as grayscale images, but is too computationally costly for larger data sets. We first study an important visual methods have been widely studied and used in data cluster analysis. The basis of the method is to regard D as a subset of known values that is part of a larger, unknown N×N dissimilarity matrix, and then impute the missing values from D. The VAT algorithm generally represent D as an N×N Image I(̅D) where the objects are reordered to reveal hidden cluster structure along the diagonal of the image. This paper addresses the limitation by proposing a VAT algorithm, where D is mapped ̅D in a graph embedding space and then reordered to ̅D using VAT algorithm. Two important points: i) because VAT is scalable by sVAT to data sets of arbitrary size, and because coVAT depends explicitly on VAT, this new approach is immediately scalable to say, the sVAT model, which works for even very large(unloadable) data sets without alteration; and ii) VAT, sVAT and coVAT are autonomous, parameter free models-no “hidden values” are needed to make them work. A sampling-based extended scheme is also proposed to enable visual cluster analysis for large data sets.Extensive experimental results on several synthetic and real-world data sets validate our VAT algorithms.

Keywords

Clustering, Cluster Analysis, Cluster Tendency, Hidden Values, VAT, sVAT, coVAT.

Full Text

     

C. Palanisamy ¹, S. Selvan ²

Affiliations
1 Department of Information Technology, Bannari Amman Institute of Technology, Tamil Nadu, IN
2 Alpha Engineering College, Chennai, Tamil Nadu, IN

Abstract

Most of the subspace clustering algorithms uses monotonicity property to generate higher dimensional subspaces. But this property is not applicable here since different subspace cardinalities have varying densities ie., if a k-dimensional unit is dense, any (k-1) dimensional projection of this unit may not be dense. So in DENCOS a mechanism to compute upper bounds of region densities to constrain the search of dense regions is devised, where the regions whose density upper bounds are lower than the density thresholds will be pruned away in identifying the dense regions. They compute the region density upper bounds by utilizing a data structure, DFP-tree to store the summarized information of the dense regions. DFP-Tree employs FP-Growth algorithm and builds an FP-Tree based on the prefix tree concept and uses it during the entire subspace identification process. This method performs repeated horizontal traversals of the data to generate relevant subspaces which is time consuming. To reduce the time complexity, we employ ITL data structure to build Density Conscious ITL (DITL) tree to be used in the entire subspace identification process. ITL reduces the cost by scanning the database only once, by significantly reducing the horizontal traversals of the database. The algorithm is evaluated through experiments on a collection of benchmark data sets datasets. Experimental results have shown favourable performance compared with other popular clustering algorithms.

Keywords

Subspace Clustering, ITL Tree, Recall, Precision.

Full Text