Research map / Label-efficient and robust learning / Recognition and long-tail
CIDR: A Cooperative Integrated Dynamic Refining Method for Minimal Feature Removal Problem
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 CIDR, which models pairwise word interactions with Cooperative Integrated Gradients and recasts the minimal feature removal problem as randomized knapsack problems whose candidate sets are filtered by frequency to find minimal feature sets for NLP classifiers.
Problem
Identifying the minimal feature set (MFS) that explains a text classifier's prediction, where prior greedy methods rely on a monotonicity assumption and ignore feature interactions.
Approach
Cooperative Integrated Gradients score word pairs for interaction; the minimal feature removal problem is cast as a 0-1 knapsack with perturbed item values to generate many candidate sets, and features shared across candidate sets (frequency above a threshold) are retained.
Key results
- On SST2 with BERT, CIDR reaches FMS 0.592 versus 0.401 for the best baseline SIG (about +19.1 points).
- CIDR achieves the best LO and Comp across all nine model-dataset settings, e.g. Comp 0.344 on SST2/BERT versus 0.331 for SIG.
- Removing CIG (CIDR w/o CIG) lowers Comp by 2.9% on BERT/SST2 relative to full CIDR, per the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CIDR instead of the strongest baseline, SIG | BERT · SST2 | FMS: 0.401 -> 0.592 | Helped |
| Use CIDR instead of the strongest baseline, SIG | DistilBERT · IMDB | FMS: 0.421 -> 0.582 | Helped |
| Add CIG to CIDR (ablation) | CIDR · SST2, IMDB, and Rotten Tomatoes | FMS: 0.459 -> 0.592 | Helped |
| Add Minimal Feature Refinement instead of the greedy variant (ablation) | CIDR · SST2, IMDB, and Rotten Tomatoes | FMS: 0.373 -> 0.592 | Helped |
| Add CIG to CIDR (ablation) | CIDR · IMDB | Comp: 0.289 -> 0.357 | Helped |
| Add Minimal Feature Refinement instead of the greedy variant (ablation) | CIDR · IMDB | Comp: 0.306 -> 0.349 | Helped |
Surprising findings
- On the IMDB and Rotten Tomatoes settings, the CIDR variant without CIG achieves higher FMS than full CIDR in several cases (e.g. BERT/IMDB FMS 0.459 vs 0.566 for full CIDR is lower; but DistilBERT/IMDB w/o CIG 0.480 vs full 0.582 is also lower), so the CIG contribution is mainly to LO/Comp rather th
Limitations
- Only binary sentiment classification datasets evaluated with fine-tuned encoder models.
- CIG approximates individual Shapley contributions with only two subset cases rather than full enumeration.
- Feature minimality is checked with a threshold t=0.5 and frequency threshold hyperparameters that must be tuned.
Experimental setup
Datasets: SST2 IMDB Rotten Tomatoes
Benchmarks: SST2 IMDB Rotten Tomatoes
Metrics: Log-odds (LO) Comprehensiveness (Comp) Feature Minimality Score (FMS)
Baselines: Grad*Input DeepLIFT GradientSHAP IG DIG SIG
Builds on: integrated gradients minimal feature removal (greedy, monotonic assumption) SIG DIG
Closest papers by meaning
- Spurious Feature Eraser: Stabilizing Test-Time Adaptation for Vision-Language Foundation Model (AAAI 2025 · similarity 0.81)
- Domain2Vec: Vectorizing Datasets to Find the Optimal Data Mixture without Training (ICML 2025 · similarity 0.81)
- Learning Multi-Level Features with Matryoshka Sparse Autoencoders (ICML 2025 · similarity 0.80)
- Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization (NeurIPS 2025 · similarity 0.80)
- What to Forget in Unlearning? Forget Set Curation for Language Models (NeurIPS 2026 · similarity 0.80)
- Learning to Reject with a Fixed Predictor: Application to Decontextualization (ICLR 2024 · similarity 0.80)
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.