Research map / Learning theory and algorithms / Prediction and decision losses
Feature Importance Ranking for Deep Learning
Maksymilian Wojtas, Ke Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 dual-net feature selection with a single classifier instead of best prior method with five classifiers | dual-net · RNA-seq | accuracy: 0.9881 -> 0.9938 | Improved |
| keep ID feature in Glass data | dual-net · Glass | accuracy: 75%-80% -> 90%+ | Improved |
| concatenate the mask with the masked features instead of using a placeholder value for removed features (ablation) | operator net · feature selection tasks | performance | Improved |
| use weighted selector loss with weights 10, 5, 1 | selector net · feature selection tasks | search effectiveness | Improved |
| compare with DFS for subset size above 15 | dual-net · Enhancer-Promoter | accuracy | Improved |
| compare with LIME and SAN instance aggregation | dual-net · Enhancer-Promoter, s 15 to 55 | accuracy | Improved |
| compare with RFE and random forest | dual-net · Enhancer-Promoter | accuracy | Improved |
| use dual-net with alternate learning instead of RFE | dual-net · Enhancer-Promoter | training time (s): 1,700 -> 1,100 | Improved |
| use dual-net instead of random forest | dual-net · Enhancer-Promoter | training time (s): 2.5 -> 1,100 | Hurt |
| compare with CCM on TOX-171 | dual-net · TOX-171 | accuracy | Hurt |
| increase mask size s beyond 30 | dual-net · Yale | feature importance map quality: 30 best among 10,30,50,70,90 | Hurt |
| use operator validation loss with optimal mask only for early stopping | dual-net · synthetic and benchmark sets | early stopping signal | Improved |
Closest papers by meaning
- Balancing Feature Similarity and Label Variability for Optimal Size-Aware One-shot Subset Selection (ICML 2024 · similarity 0.83)
- Optimizing importance weighting in the presence of sub-population shifts (ICLR 2025 · similarity 0.83)
- Contextual Feature Selection with Conditional Stochastic Gates (ICML 2024 · similarity 0.83)
- Testing Semantic Importance via Betting (NeurIPS 2024 · similarity 0.83)
- SAND: One-Shot Feature Selection with Additive Noise Distortion (ICML 2025 · similarity 0.83)
- Active feature acquisition via explainability-driven ranking (ICML 2025 · 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.
