Reproduce “Change Point Detection in Time Series Data using Autoencoders with a Time-Invariant Representation” 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
Tim De Ryck, Maarten De Vos, Alexander Bertrand
The TIRE method introduces an autoencoder-based approach for change point detection that learns partially time-invariant features from combined time and frequency domain data. This method, enhanced by a novel topographic prominence-based postprocessing step, consistently achieves higher detection accuracy and fewer false alarms on various simulated and real-world time series datasets.
Key insights
- Explicitly enforcing time-invariance in learned features via a custom autoencoder loss enables the model to extract representations highly relevant for change point detection, ignoring irrelevant variations.
- Combining time-domain and frequency-domain analysis allows for comprehensive detection of various change types, including shifts in mean, variance, autocorrelation, and frequency spectrum, particularly in non-i.i.d. time series.
- Sophisticated postprocessing, incorporating matched filtering and a novel application of topographic prominence, is crucial for dramatically reducing false alarms and precisely localizing change points across diverse detection methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.