A0835
Title: Advancing statistical jump models: A novel method for robust temporal clustering with state-dependent feature selection
Authors: Federico Cortese - University of Milan (Italy) [presenting]
Antonio Pievatolo - National Research Council of Italy (Italy)
Abstract: Statistical sparse jump models offer a flexible alternative to traditional hidden Markov models by capturing complex dynamics via smooth transitions between regimes while simultaneously performing feature selection. However, they typically select features globally, assuming their relevance across all states, and might be sensitive to outliers. The aim is to introduce a robust sparse jump model that estimates a latent state sequence over time, along with a matrix of state-specific feature weights. This allows for modeling regime-dependent feature relevance, which might be crucial in applications where different subsets of features possibly drive different regimes. Robustness is ensured by: (1) integration of an outlyingness factor that down-weights atypical observations during estimation, and (2) the use of medoids, rather than prototypes, as state representatives.
