Justin Bedo

Supervised feature selection via dependence estimation

We introduce a framework for filtering features that employs the Hilbert-Schmidt Independence Criterion (HSIC) as a measure of dependence between the features and the labels. The key idea is that good features should maximise such dependence. Feature selection for various supervised learning problems (including classification and regression) is unified under this framework, and the solutions can be approximated using a backward-elimination algorithm. We demonstrate the usefulness of our method on both artificial and real world datasets.

An efficient alternative to SVM based recursive feature elimination with applications in natural language processing and bioinformatics

The SVM based Recursive Feature Elimination (RFE-SVM) algorithm is a popular technique for feature selection, used in natural language processing and bioinformatics. Recently it was demonstrated that a small regularisation constant C can considerably improve the performance of RFE-SVM on microarray datasets. In this paper we show that further improvements are possible if the explicitly computable limit C →0 is used. We prove that in this limit most forms of SVM and ridge regression classifiers scaled by the factor

Feature selection via dependence maximization

\frac{1}{C}

Gene selection via the BAHSIC family of algorithms

converge to a centroid classifier. As this classifier can be used directly for feature ranking, in the limit we can avoid the computationally demanding recursion and convex optimisation in RFE-SVM. Comparisons on two text based author verification tasks and on three genomic microarray classification tasks indicate that this straightforward method can surprisingly obtain comparable (at times superior) performance and is about an order of magnitude faster.