Multiclass Bayesian feature selection

Abstract

Many bioinformatics studies aim to find biomarkers that discriminate between two or more groups. While in many applications the groups are inherently multiclass, most feature selection algorithms focus on the binary case, and many are not straightforward to generalize. In this work, we demonstrate that a recently proposed family of binary Bayesian models for feature selection can be extended to the multiclass case. We propose multiclass generalizations of two algorithms: (1) a feature filtering algorithm that is optimal under independent features, and (2) a fast suboptimal feature selection algorithm that accounts for correlations. The proposed algorithms perform well on several synthetic datasets, and produce many genes and pathways with interesting annotations when applied to a real four-class colon cancer microarray dataset.