Quantcast

Information-theoretic feature selection with discrete k-median clustering

Research paper by Onur Şeref, Ya-Ju Fan, Elan Borenstein, Wanpracha A. Chaovalitwongse

Indexed on: 09 Apr '14Published on: 09 Apr '14Published in: Annals of Operations Research



Abstract

We propose a novel computational framework that integrates information-theoretic feature selection with discrete \(k\)-median clustering (DKM). DKM is a domain-independent clustering algorithm which requires a pairwise distance matrix between samples that can be defined arbitrarily as input. In the proposed DKM clustering, the center of each cluster is represented by a set of samples, which induce a separate set of clusters for each feature dimension. We evaluate the relevance of each feature by the normalized mutual information (NMI) scores between the base clusters using all features and the induced clusters for that feature dimension. We propose a spectral cluster analysis (SCA) method to determine the number of clusters using the average of the relevance NMI scores. We introduce filter- and wrapper-based feature selection algorithms that produce a ranked list of features using the relevance NMI scores. We create an information gain curve and calculate the normalized area under this curve to quantify information gain and identify the contributing features. We study the properties of our information-theoretic framework for clustering, SCA and feature selection on simulated data. We demonstrate that SCA can accurately identify the number of clusters in simulated data and public benchmark datasets. We also compare the clustering and feature selection performance of our framework to other domain-dependent and domain-independent algorithms on public benchmark datasets and a real-life neural time series dataset. We show that DKM runs comparably fast with better performance.