TabTalk #3: Unifying Time Series Segmentation: Methods, Metrics, and Large-SCale Evaluation

Time Series Segmentation unifies Change Point Detection (CPD) and State Detection (SD) into a single problem: partitioning temporal data into meaningful, homogeneous segments. Despite its central role across domains, the field remains fragmented in methods, evaluation, and large-scale validation.

In this talk, we first present a unified view of existing approaches through a typology of segmentation methods, together with tsseg, an open-source library that standardizes 30+ algorithms within a common API, together with an interactive demo. We then introduce new evaluation measures, notably the State Matching Score (SMS), providing more interpretable and fine-grained assessment of segmentation quality. Building on this, we present the first large-scale joint CPD and SD benchmark, enabling systematic comparison and links across both tasks. Finally, we discuss emerging perspectives, including the role of foundation models for time series segmentation.

‍

‍

Félix Chavelli is a rising researcher at Inria Paris / ENS-PSL (VALDA team), pushing the boundaries of interpretability in multivariate time series. He has a world-class academic background spanning ENSTA Paris and the National University of Singapore, and has published at NeurIPS 2025.

His website: https://felix.chavelli.fr

Felix Chavelli est chercheur à l'INRIA Paris / ENS-PSL (équipe VALDA); ses recherches portent sur l'interprétabilité des séries temporelles multivariées. Il a suivi une formation académique d'envergure internationale, entre l'ENSTA Paris et la National University of Singapore, et a publié à NeurIPS 2025.‍

Son site: https://felix.chavelli.fr

‍