Reproduce “Conformal Changepoint Localization and Root Cause Analysis with Corrupted Observations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seunghun Yu, Meiyi Zhu, Petar Popovski, Joonhyuk Kang, Osvaldo Simeone
The method narrows changepoint and root-cause confidence sets by downweighting observations whose classifier signals uncertainty, while using within-segment permutations for calibration. On DomainNet at contamination probability 0.7, weighted sets average 6.76 candidate times with 0.970 coverage, versus 137.43 and 0.965 for the unweighted baseline. Coverage relies on exchangeability and independent Huber-type corruption; smaller sets are experimental, not guaranteed.
Key insights
- Conformal changepoint and root cause localization methods inherently maintain their finite-sample coverage guarantees even when observations are subject to Huber-type contamination.
- By appropriately downweighting observations based on classifier uncertainty, the informativeness of confidence sets can be substantially improved (reduced in size) while preserving the crucial coverage guarantees.
- Meta-learning allows for adaptive weighting schemes that are robust to unknown contamination levels and can even improve the underlying classifier's robustness, leading to sharper confidence sets than fixed-parameter approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.