Robust mean-shift clustering based on impartial trimming

A robust version of the mean-shift algorithm is developed to cope with the presence of contaminated data when targeting a clustering problem by means of a modal approach. The goal is to protect nonparametric density-based clustering from the deleterious effect of outliers. Their occurrence affects the analysis mainly because outliers lead to the detection of spurious modes and groups. Therefore, the proposed methodology aims to recover the underlying clustered data configuration, while detecting and discarding outliers. A strategy to select the level of trimming is discussed. The finite sample behaviour of the proposed method is investigated by Monte Carlo numerical studies and empirical applications.