Research map / Probabilistic and scientific modeling / Variational and generative sampling
Feature Shift Detection: Localizing Which Features Have Shifted via Conditional Distribution Tests
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the score-based Fisher-divergence statistic instead of KS with the Gaussian density model | Feature shift detection and localization · Unknown single sensor, MI 0.2 | recall: 0.87 -> 0.96 | Helped |
| Use the score-based Fisher-divergence statistic instead of KS with the Gaussian density model | Feature shift detection and localization · Unknown single sensor, MI 0.2 | run time per test: 1.408 -> 0.044 | Helped |
| Use conditional feature tests rather than marginal KS tests | Unknown single sensor detection · Unknown single sensor, MI 0.2 | recall: 0.00 -> 0.96 | Helped |
| Use the score-based statistic instead of MB-KS | Three-sensor localization · Complete graph, MI 0.2 | precision: 0.62 -> 0.71 | Helped |
| Use the score-based statistic instead of MB-KS | Three-sensor localization · Complete graph, MI 0.2 | recall: 0.69 -> 0.78 | Helped |
| Replace simple bootstrapping with time-dependent bootstrapping (ablation) | MB-SM on real-world unshuffled Energy data · Energy, feature shift detection | recall: 100% -> 8.86% | Hurt |
| Replace the Gaussian density model with an iterative Gaussianization normalizing flow while using Time-Boot (ablation) | Feature shift detection · Energy, unshuffled | precision: 16.67% -> 55.00% | Helped |
| Replace the Gaussian density model with an iterative Gaussianization normalizing flow while using Time-Boot (ablation) | Feature shift detection · Energy, unshuffled | recall: 8.86% -> 62.62% | Helped |
| Use a larger sliding-window size (ablation) | Score-based feature shift detection · Time-series localization across graph types | recall: See Table: compromised sensor localization with window sizes 200, 400, 600, 800, 1000. | Mixed |
| Use a larger number of compromised sensors (ablation) | MB-SM localization · Simulated multiple-sensor attacks | recall | Hurt |
Closest papers by meaning
- Sequential Covariate Shift Detection Using Classifier Two-Sample Tests (ICML 2022 · similarity 0.86)
- Tracking the risk of a deployed model and detecting harmful distribution shifts (ICLR 2022 · similarity 0.86)
- Sequential Harmful Shift Detection Without Labels (NeurIPS 2024 · similarity 0.85)
- Feature Shift Localization Network (ICML 2025 · similarity 0.85)
- Multiply Robust Estimation for Local Distribution Shifts with Multiple Domains (ICML 2024 · similarity 0.83)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.83)
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.