Variable selection for robust model-based learning from contaminated data

Several contributions to the recent literature have shown that supervised learning is greatly enhanced when only the most relevant features are selected for building the discrimination rule. Unfortunately, outliers and wrongly labelled units may undermine the determination of relevant predictors, and almost no dedicated methodologies have been developed to face this issue. In the present paper, we introduce a new robust variable selection approach, that embeds a classifier within a greedy-forward procedure. An experiment on synthetic data is provided, to under- line the benefits of the proposed method in comparison with non-robust solutions.