BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:DUST: A Duality-Based Pruning Method For Exact Multiple Change-Poi
 nt Detection
DTSTART:20260210T130000Z
DTEND:20260210T140000Z
DTSTAMP:20260425T064400Z
UID:indico-event-15720@indico.math.cnrs.fr
DESCRIPTION:Speakers: Vincent Runge\n\nWe address the computational challe
 nge of detecting multiple change points in large time series. In particula
 r\, we focus on minimising penalised likelihoods for exponential-family mo
 dels. While dynamic programming yields exact solutions with at most quadra
 tic complexity\, existing pruning methods have important limitations: (i) 
 PELT prunes inefficiently when change points are sparse\, and (ii) FPOP is
  not well suited to multi-parameter models.\nWe introduce DUST (DUal Simpl
 e Test)\, a duality-based pruning framework that discards candidate change
  points by comparing a dual function with a threshold. Experiments across 
 regimes and models show that DUST combines PELT’s simplicity with FPOP
 ’s efficiency\, with particularly strong gains for non-Gaussian data. A 
 detailed introduction will situate this contribution within the state of t
 he art in dynamic programming for time-series analysis.\n\nhttps://indico.
 math.cnrs.fr/event/15720/
URL:https://indico.math.cnrs.fr/event/15720/
END:VEVENT
END:VCALENDAR