Research map / Label-efficient and robust learning / Recognition and long-tail
Feature Shift Localization Network
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.
Introduces FSL-Net, a convolutional neural network that localizes shifted features between two datasets in a single forward pass, trained on simulated shifts over 1,350 datasets and generalizing without retraining.
Problem
Localizing which features cause a distribution shift between two datasets is accurate only with slow or non-scalable methods, which struggle on large, high-dimensional data.
Approach
A permutation-aware neural network that compares learned statistical functional maps of reference and query datasets (via normalized squared difference) and outputs per-feature corruption probabilities, trained end-to-end on many simulated shifts so it transfers to unseen datasets without retraining.
Key results
- FSL-Net achieves the highest average F-1 across methods, comparable to DataFix, while being about 36x faster on average than DataFix and up to 136x faster on Phenotypes.
- On the Phenotypes dataset FSL-Net completes in 2 minutes versus 13 hours for DataFix; on Canine, 12 minutes versus 9 hours.
- Ablation: the SM-only baseline improves from F-1 0.307 to 0.710 when adding the Prediction Network; full model reaches 0.894 after 7-day training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FSL-Net instead of DataFix | Feature shift localization · CIFAR10 | F-1 score: 0.8911 -> 0.9565 | Helped |
| Use FSL-Net instead of DataFix | Feature shift localization · COIL-100 | F-1 score: 0.9805 -> 0.9720 | Hurt |
| Use FSL-Net instead of DataFix | Feature shift localization · CIFAR10 | Mean runtime (hours): 0.2948 -> 0.0319 | Helped |
| Use FSL-Net instead of DataFix | Feature shift localization · COIL-100 | Mean runtime (hours): 2.9228 -> 0.4906 | Helped |
| Add the Prediction Network to statistical measures (ablation) | FSL-Net · Benchmark validation | F-1 score: 0.307 -> 0.710 | Helped |
| Add the Moment Extraction Network to statistical measures and the Prediction Network (ablation) | FSL-Net · Benchmark validation | F-1 score: 0.710 -> 0.855 | Helped |
| Add the Neural Embedding Network to statistical measures and the Prediction Network (ablation) | FSL-Net · Benchmark validation | F-1 score: 0.710 -> 0.878 | Helped |
| Add statistical measures to Moment Extraction Network, Neural Embedding Network, and Prediction Network (ablation) | FSL-Net · Benchmark validation | F-1 score: 0.770 -> 0.889 | Helped |
| Add Neural Embedding Network to statistical measures, Moment Extraction Network, and Prediction Network (ablation) | FSL-Net · Benchmark validation | F-1 score: 0.855 -> 0.889 | Helped |
| Add Moment Extraction Network to statistical measures, Neural Embedding Network, and Prediction Network (ablation) | FSL-Net · Benchmark validation | F-1 score: 0.878 -> 0.889 | Helped |
| Extend full-model training from 3 days to 7 days (ablation) | FSL-Net · Benchmark validation | F-1 score: 0.889 -> 0.894 | Helped |
| Add attention only within the Prediction Network (ablation) | FSL-Net · Feature shift localization | F-1 score | Hurt |
Surprising findings
- Thresholding the statistical measures directly gives F-1 of only 0.307; the learned Prediction Network is essential.
Limitations
- Fails to localize very minimal perturbations (E4, E9) where DataFix performs better.
- Attention layers caused CUDA out-of-memory issues and reduced performance when used in the Prediction Network.
Experimental setup
Datasets: OpenML tabular datasets (1,032) simulated algebraic datasets (184) simulated probabilistic datasets (184) Gas Covid Energy Musk2 Scene MNIST Polynomial Cosine Dilbert
Benchmarks: Gas Covid Energy Musk2 Scene MNIST Polynomial Cosine Dilbert Phenotypes Founders Canine
Metrics: F-1 score wall-clock runtime
Baselines: DataFix MB-SM MB-KS KNN-KS Deep-SM MI SelectKBest MRMR FAST-CMIM
Builds on: DataFix Deep Sets Maximum Mean Discrepancy EVA attention
Closest papers by meaning
- Adversarial Learning for Feature Shift Detection and Correction (NeurIPS 2023 · similarity 0.87)
- Feature Shift Detection: Localizing Which Features Have Shifted via Conditional Distribution Tests (NeurIPS 2020 · similarity 0.85)
- Explaining Concept Shift with Interpretable Feature Attribution (ICML 2026 · similarity 0.82)
- Detecting and Correcting for Label Shift with Black Box Predictors (ICML 2018 · similarity 0.81)
- TabFSBench: Tabular Benchmark for Feature Shifts in Open Environments (ICML 2025 · similarity 0.79)
- Learning Neural Networks with Distribution Shift: Efficiently Certifiable Guarantees (ICLR 2025 · similarity 0.79)
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.