Research map / Time series / Causal and generative time series
Hubness Change Point Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Hub-CPD, a model-free change point detector that uses hubness suppression (centering plus length normalization of residuals between sliding-window sub-series) and a Rayleigh uniformity test to flag changes.
Problem
Detecting changes in multivariate time-series from machine sensors quickly and accurately, without heavy computation suitable for edge deployment.
Approach
Uses hubness, normally a nuisance for nearest-neighbor search, as a signal: after iteratively suppressing hubness between two adjacent sub-series sets via residual-to-local-centroid normalization, the data stay uniform on the hypersphere only if no change occurred, measured by the Rayleigh score.
Key results
- On hydraulic system sensor data, Hub-CPD (r=2) achieved the highest average AUC of 0.933 (r=1: 0.927) versus 0.919 for BnB and 0.891 for RuLSIF.
- On gas sensor data, Hub-CPD (r=1) achieved the highest average AUC of 0.952 versus 0.940 for RuLSIF.
- Hub-CPD ran in 0.5 s on hydraulic data versus 11.8 s for RuLSIF and 212.8 s for KL-CPD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use r = 2 instead of r = 1 (ablation) | Hub-CPD · Hydraulic system data | AUC: 0.927 -> 0.933 | Mixed |
| Use r = 2 instead of r = 1 (ablation) | Hub-CPD · Gas sensor data | AUC: 0.952 -> 0.920 | Hurt |
| Use r = 2 instead of r = 1 (ablation) | Hub-CPD · PS3 | AUC: 0.858 -> 0.832 | Hurt |
| Use r = 2 instead of r = 1 (ablation) | Hub-CPD · EPS1 | AUC: 0.802 -> 0.838 | Helped |
| Use r = 2 instead of r = 1 (ablation) | Hub-CPD · Figaro | AUC: 0.905 -> 0.842 | Hurt |
| Use r = 2 instead of r = 1 (ablation) | Hub-CPD · FIS | AUC: 0.999 -> 0.999 | No clear effect |
Surprising findings
- On the synthetic AR-model scaling-variance setting, the model-based mSSA-MW outperformed the model-free methods, including Hub-CPD.
Limitations
- Number of residual-normalization iterations r must be chosen; too many iterations can detect differences between identical distributions.
- Accuracy degrades for small window size w and decreases with small n, similar to other methods.
- Ran worse than mSSA-MW on AR scaling-variance synthetic data and on the EPS1 sensor.
Experimental setup
Datasets: synthetic time-series (Gaussian, t-distribution, AR(1) with jumping mean/scaling variance) UCI hydraulic system dataset UCI gas sensor dataset (CO concentration)
Benchmarks: synthetic time-series change detection UCI hydraulic system data UCI gas sensor data
Metrics: AUC runtime (seconds)
Baselines: RuLSIF KL-CPD BnB mSSA-MW
Builds on: hubness reduction centering RuLSIF KL-CPD BnB mSSA
Closest papers by meaning
- On Finding Hubs in High Dimensions with Sampling (AAAI 2025 · similarity 0.83)
- Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.81)
- Non-parametric Online Change Point Detection on Riemannian Manifolds (ICML 2024 · similarity 0.81)
- Optimal Online Change Detection via Random Fourier Features (NeurIPS 2025 · similarity 0.78)
- The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.78)
- Multiresolution Analysis and Statistical Thresholding on Dynamic Networks (NeurIPS 2025 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.