A Feature Clustering Approach for Dimensionality Reduction and Classification

Abstract

Dimensionality reduction is one of the primary challenges when handling high dimensional data. Feature clustering is a powerful approach for reducing the dimensionality of the global feature vector when performing classification. In this paper, we discuss the current research issues in handling data streams and high dimensional data and introduce an approach to perform dimensionality reduction by computing the standard deviation of each feature with every transaction or document of the entire dataset. We then rank and cluster the features of the global feature vector to obtain feature-cluster matrix. The feature-cluster matrix so formed is used to perform dimensionality reduction. Then we show how the reduced dimensionality can be used to perform classification after elimination of noise. In this work, we classify the new test document or transaction after reducing dimensionality. In future, the idea is to cluster the features using a kernel measure and perform clustering and classification of text streams dynamically.