Search

Analysis of gene-expression data often requires that a gene (feature) subset is selected and many feature selection (FS) methods have been devised. However, FS methods often generate different lists of features for the same dataset and users then have to choose which list to use. One approach to support this choice is to apply stability metrics on the generated lists and selecting lists on that base. The aim of this study is to investigate the behavior of stability metrics applied to feature subsets generated by FS methods. The experiments in this work explore a plethora of gene expression datasets, FS methods, and expected number of features to compare several stability metrics. The stability metrics have been used to compare five feature selection methods (SVM, SAM, ReliefF, RFE + RF and LIMMA) on gene expression datasets from the EBI repository. Results show that the studied stability metrics display a high amount of variability. The reason behind this is not clear yet and is being further investigated. The final objective of the research, that is to define how to select a FS method, is an ongoing work whose partial findings are reported herein.

Microarray technologies produce very large amounts of data that need to be classified for interpretation. Large data coupled with small sample sizes make it challenging for researchers to get useful information and therefore a lot of effort goes into the design and testing of feature selection tools; literature abounds with description of numerous methods. In this paper we select five representative review papers in the field of feature selection for microarray data in order to understand their underlying classification of methods. Finally, on this base, we propose an extended taxonomy for categorizing feature selection techniques and use it to classify the main methods presented in the selected reviews.