Report of Reviewer #1 #2
Description
Associate Editor: MP Etienne
Reviewer : eaEU chose to remain anonymous
Reviewer: Reviewing history
- Paper submitted June 2023
- Reviewer invited June, 27th 2023
- Review 1 received August, 8th 2023
- Paper revised February, 26th 2024
- Reviewer invited March, 14th 2024
- Review 2 received May, 21st 2024
- Paper conditionally accepted June, 3rd 2024
First Round
This article presents a novel pruning strategy for multivariate change point detection. By relaxing the functional pruning thanks to geometrical considerations, the methods allows to discard a large number of change point candidates and thus to improve computational costs. The article is well written and very clear. Here are my remarks (that are mostly linked to the Results section):
Section 4.2 raises an important issue, which is that if the pruning strategy is too complex and computationally demanding, the computational cost improvement caused by the pruning is somehow ineffective since the algorithm spends a lot of time on pruning. Some discussions and additional results would help to better understand this point. In a standard “real life” configuration (e.g. multivariate signal with some breakpoints), what is the complexity of the S-type and R-type pruning ? What is the actual time spent on these steps rather on the baseline optimization ?
In Section 4.3 you present a novel version of the algorithm which is faster that the standard GeomFPOP algorithm. However, it would be interesting to see the number of change point candidates stored over time as in 4.1, as intuitively, using a randomized approach probably degrades the pruning strategy ? A study of the compromize between the two variables (empirical complexity vs. number of pruned change points) could be interesting.
In the introduction of section 4 you state that you first compared the change points provided by PELT and your method and “made sure” they were identifical. A brief report on these simulations would be useful.
An open question : what would happen is some changes are only present in a subset of dimensions ? Would your pruning strategy be robust to that ? Would both pruning strategy behave similarly ?
Second round
The discussion is available on
https://openreview.net/forum?id=YIes47zsCE