Speaker
Description
Changepoint detection seeks to identify structural breaks in sequential data, often arising as noisy observations of underlying stochastic dynamical systems. However, exact likelihood-based methods can be computationally expensive, particularly in large-scale settings. We study maximum likelihood estimation for multiple changepoint models under possible overfitting and show that, even when the number of segments is misspecified, the resulting estimator converges to the true signal at a fast parametric $N^{-1/2}$ rate. This observation motivates a bottom–up correction strategy for over-segmented solutions.
We propose Dendrogram Pruning and Merging (DPM), an agglomerative algorithm that starts from an overfitted segmentation and iteratively merges adjacent segments using likelihood-based distances, producing a hierarchical dendrogram of candidate changepoints. We further introduce DsSIC, a dendrogram-based model selection criterion combining DPM with a strengthened Schwarz Information Criterion, and establish consistency of the resulting changepoint estimates.
Simulation studies and a single-molecule tracking application demonstrate that DsSIC achieves accuracy comparable to exact methods while substantially reducing computational cost.
