Robuuste methoden voor het opsporen van anomalieën met toepassingen voor empirisch onderbouwd beleid

The project - in cooperation with the Joint Research Centre (JRC) of the European Commission - develops and applies robust statistical techniques for outlier/anomaly detection. The main methodological contributions relate to robust co-clustering and robust regression techniques, addressing both casewise and cellwise outliers. Such methods will are applied to high-dimensional and heterogeneous data provided by the JRC coming from different fields of application, including fraud detection and international trade monitoring. In more detail, the work consists of four contributions spanning clustering and regression. First, two robust co-clustering methods are introduced under the Latent Block Model. The RoCo-TRICC approach handles row-wise and column-wise contamination via impartial trimming, while Cell-TRICC is the first method to address cellwise outliers in this framework and incorporates missing data, model selection, and new diagnostic tools. Both methods are validated through simulations and real applications. The second part of the PhD thesis concerns robust regression. A fraud-detection framework combining MM-regression with spatial autocorrelation measures is proposed for the EU tobacco track-and-trace system. Finally, a robust and regularised Generalised Linear Model for high-dimensional mixed-type data is developed, featuring protection against leverage and vertical outliers, tailored penalties, robust initialisation, and model selection.