Robust twin boosting for feature selection from high-dimensional omics data with label noise

Omics data such as microarray transcriptomic and mass spectrometry proteomic data are typically characterized by high dimensionality and relatively small sample sizes. In order to discover biomarkers for diagnosis and prognosis from omics data, feature selection has become an indispensable step to find a parsimonious set of informative features. However, many previous studies report considerable label noise in omics data, which will lead to unreliable inferences to select uninformative features. Yet, to the best of our knowledge, very few feature selection methods are proposed to address this problem. This paper proposes a novel ensemble feature selection algorithm, robust twin boosting feature selection (RTBFS), which is robust to label noise in omics data. The algorithm has been validated on an omics feature selection test bed and seven real-world heterogeneous omics datasets, of which some are known to have label noise. Compared with several state-of-the-art ensemble feature selection methods, RTBFS can select more informative features despite label noise and obtain better classification results. RTBFS is a general feature selection method and can be applied to other data with label noise. MATLAB implementation of RTBFS and sample datasets are available at: http://www.cs.bham.ac.uk/∼szh/TReBFSMatlab.zip