Major Issues in Spectral Clustering Algorithm to Improve the Quality of Output Clusters

Abstract

Spectral clustering is playing a vital role in day-to-day technology in different forms since recent past. Spectral Clustering techniques are developed by using concepts of graph, algebra theories and conventional clustering methods. The theories and experiments conducted in various areas like, pattern recognition and image segmentation prove that it is having global reputation. This paper is to provide a framework for spectral clustering in different phases and its advantages when compared with traditional data clustering algorithms. In lieu of this paper, it is upgraded by intervening the inputs such as to making the bounds of clustering i.e., providing the issues through its relevance factors-produced by spectral clustering. On the other hand, some fundamental issues related to improving the quality of the output clusters are also covered. Constructing a sound similarity matrix is the most direct and efficient way to improve the quality of clusters. For building the similarity matrix sometimes we may choose the distance measures for measuring the similarity. This paper provides the issues related to improving the quality of the clusters produced by spectral clustering. We implemented spectral clustering in WEKA tool and produced the results for different similarity measures and observed how the quality of the clusters improved. For practical purpose we have used the datasets available in UCI machine library.