- Rating
- 2
- Confidence
- 1
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper propose In-Context Change-Point Localization (ICCP), which uses previously solved tasks to adapt to new localization problems without updating model parameters. The authors provide theoretical analysis of localization error and consistency. Experiments are conducted on synthetic Gaussian datasets with mean, variance, and covariance changes.
However, I find the paper relatively difficult to follow, mainly due to the limited visual illustrations and experimental results in the main paper. More importantly, the experiments are only conducted on simplified synthetic datasets, without evaluation on real-world data or more complex setting. This makes it difficult to assess the practical effectiveness and generalizability of the proposed method.
Critical strengths
-
Clear Problem Formulation and Motivation. The paper identifies the limitation that different change mechanisms require different localization strategies. To address this challenge, the authors propose an in-context learning framework that adapts to different change mechanisms without parameter updates.
-
Careful Statistical Analysis. The paper pays attention to statistical considerations. The theoretical analysis provides finite-sample error bounds and localization consistency guarantees under specified assumptions.
Critical weaknesses
-
Limited Baseline Comparisons. The paper only compares the proposed method with single baseline, lacking comparisons with existing other change-point localization methods. This makes it difficult to assess the effectiveness and competitiveness of the proposed approach.
-
Limited Experimental Settings. All experiments are conducted on synthetic data generated from a Gaussian distribution. The lack of evaluation on real-world datasets limits the evidence for the generalization and practical applicability of the proposed method.
-
Presentation and Organization Issues. The paper contains excessive whitespace and uneven spacing. Despite using only eight pages for the main text, it includes just two figures and limited experimental results, while substantial material is placed in the appendix. The available space could be better utilized to provide more experimental evidence and analysis.
Questions for authors
See weaknesses.
3 comments
Thank you for the rebuttal. I have informed the AC and SAC that my expertise in this domain is limited. Please give less weight to my review when making the final decision.
Thank you for your rigorous attitude towards academia. We hope our work will make it easier for researchers and practitioners without specialized training in change-point analysis to apply it to their own data.
Response to Reviewer V3AK
Thank you for raising these questions. We noted your confidence rating of 1. Your questions bring us to the practical problem that motivated this work.
Identifying structural changes helps users avoid pooling observations from different regimes, which can improve downstream analyses. ICCP uses labeled historical examples to localize changes in a new sequence without requiring users to design a separate statistic for each change mechanism. We provide theoretical support for this approach by establishing finite-sample localization error bounds and relative localization consistency under stated assumptions. Users without a background in change-point analysis can thus use ICCP to locate change points during preprocessing. For sensitive data, users can train and run a domain-specific model locally so that raw data stay on their own systems.
Real-world data and baselines. We added experiments on two public benchmarks used in human activity recognition and time-series segmentation, UCI HAPT [1] and HAS2023 [2]. We included recent multivariate nonparametric baselines such as changeforest [3] and changeAUC [4]. ICCP achieves lower localization errors than the four evaluated offline baselines on both benchmarks (Table R5 and Section 5.2, Table 2 of the revised manuscript). In the simulations, we added comparisons with classical localizers matched to each of the three change mechanisms (Section 5.1, Table 1).
Presentation and organization. We reorganized the Experiments section and added figures and tables for the simulations and real-data evaluations so readers can see how ICCP performs across settings and how it compares with the baselines. The simulation results are in Section 5.1 (Figs. 2–4 and Table 1), and the real-data results are in Section 5.2 (Table 2).
Table R5. Comparison of localization methods on HAPT and HAS2023. MAE is in samples. ICCP entries are means ± sample SD over five seeds.
References
[1] Y. Zhang, L. Wang, Q. Gao, C. Xiang, M. Bennamoun, and Y. Guo. Triple Spectral Fusion for Sensor-Based Human Activity Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 48(9), pp. 10856–10873, 2026.
[2] A. Ermshaus, P. Schäfer, and U. Leser. CLaP – State Detection from Time Series. Proceedings of the VLDB Endowment, 19(1), pp. 70–83, 2025.
[3] M. Londschien, P. Bühlmann, and S. Kovács. Random Forests for Change Point Detection. Journal of Machine Learning Research, 24(216), pp. 1–45, 2023.
[4] R. Kanrar, F. Jiang, and Z. Cai. Model-free Change-Point Detection Using AUC of a Classifier. Journal of Machine Learning Research, 26(190), pp. 1–50, 2025.