Please register for the event here: https://forms.cloud.microsoft/r/LpaRRcTyux

Abstract:

In high-dimensional Multivariate Time Series (MTS) forecasting, identifying a parsimonious subset of predictors is crucial to avoid overfitting and the curse of dimensionality. While traditional Time Series Selection methods focus on finding a single Markov Boundary (MB), real-world data often contain informational equivalences that result in the existence of multiple valid MBs. Selecting only one solution can mislead domain experts and distort causal explanations. In this talk we will present ChronoEpilogi, a sequential greedy algorithm designed to solve the multiple MB selection problem. ChronoEpilogi simultaneously addresses four core challenges: Scalability to high-dimensional MTS via fast forward selection; Optimality with provable convergence to a true MB under relaxed assumptions; Multiplicity via heuristics that uncover all equivalent MBs; and Generality across model assumptions, mixed-type data, and tasks. Evaluated within an AutoML framework using new metrics for multiple solutions, ChronoEpilogi outperforms existing baselines on MB recovery and computation time. Using a series of experiments over synthetic and real MTS, we demonstrate that ChronoEpilogi solves the TSS single problem with on-par predictive performance with GroupLasso while being an order of magnitude faster, and statistically significantly improves over NoSelection. More importantly, ChronoEpilogi is the only algorithm capable of recovering an MB regardless of multiplicity. Finally, ChronoEpilogi can effectively handle count, numerical, and binary TS, while keeping causal f1 metrics above 0.9.  

Biography:

Vassilis Christophides received his engineering diploma from NTUA (1988), his DEA from Université Paris VI (1992), and his Ph.D. from CNAM Paris (1996). He holds a Full Professorship at ENSEA, Cergy, and is currently delegated by the CNRS to the IPAL International Research Laboratory in Singapore, where he works in partnership with NUS as an Adjunct Professor. His background includes 16 years of academic faculty service at the University of Crete, including 11 years as Full Professor, alongside the coordination of numerous competitive research projects. His research portfolio focuses on Generative AI, Machine Learning Systems, Big Data Analytics, and Web Databases, supported by over 190 publications and the 2004 ACM SIGMOD Test of Time Award. His professional service includes serving as General Chair of EDBT/ICDT 2014 as well as several international conferences’ executive and steering boards.