Research map / Label-efficient and robust learning / Recognition and long-tail
AFS: An Attention-Based Mechanism for Supervised Feature Selection
Ning Gui, Danni Ge, Ziyin Hu
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 |
|---|---|---|---|
| Reuse a pretrained learning model and globally fine-tune AFS for 40 steps instead of training AFS for 3000 steps (ablation) | AFS on MNIST · MNIST | feature weight generation time: 218s -> 25s | Improved |
| Reuse a pretrained learning model and globally fine-tune for 40 steps rather than train AFS for 3000 steps (ablation) | AFS on MNIST · MNIST, top 25 features | accuracy: 87.05 -> 82.09 | Hurt |
| Freeze the reused learning module and tune only the attention module for 40 steps (ablation) | AFS on MNIST · MNIST, top 25 features | accuracy: 87.05 -> 76.59 | Hurt |
| Use global rather than local tuning of the reused learning module, at 40 steps (ablation) | AFS-R on MNIST · MNIST, top 25 features | accuracy: 76.59 -> 82.09 | Improved |
| Use global rather than local tuning of the reused learning module, at 40 steps (ablation) | AFS-R on MNIST · MNIST, top 35 features | accuracy: 85.21 -> 88.02 | Improved |
| Use global rather than local tuning of the reused learning module, at 40 steps (ablation) | AFS-R on MNIST · MNIST, top 65 features | accuracy: 91.17 -> 94.50 | Improved |
| Use global rather than local tuning of the reused learning module, at 40 steps (ablation) | AFS-R on MNIST · MNIST, top 95 features | accuracy: 95.40 -> 95.76 | Improved |
| Increase global fine-tuning from 10 to 40 steps (ablation) | AFS-R-GlobalTune on MNIST · MNIST, top 25 features | accuracy: 63.68 -> 82.09 | Improved |
| Increase local fine-tuning from 10 to 40 steps (ablation) | AFS-R-LocalTune on MNIST · MNIST, top 25 features | accuracy: 72.33 -> 76.59 | Improved |
| Initialize AFS with Fisher Score feature weights (ablation) | AFS on Lung_discrete · Lung_discrete | average accuracy: +1.09% | Improved |
| Initialize AFS with ReliefF feature weights (ablation) | AFS on Lung_discrete · Lung_discrete | average accuracy: +1.73% | Improved |
| Initialize AFS with ReliefF feature weights (ablation) | AFS on Isolet · Isolet | average accuracy: about 1.77% | Improved |
Closest papers by meaning
- Sequential Attention for Feature Selection (ICLR 2023 · similarity 0.87)
- SAND: One-Shot Feature Selection with Additive Noise Distortion (ICML 2025 · similarity 0.87)
- Algorithmic stability and generalization of an unsupervised feature selection algorithm (NeurIPS 2021 · similarity 0.85)
- E2E-FS: An End-to-End Feature Selection Method for Neural Networks (TPAMI 2022 · similarity 0.85)
- Active feature acquisition via explainability-driven ranking (ICML 2025 · similarity 0.83)
- Hypernetworks for Dynamic Feature Selection (NeurIPS 2026 · similarity 0.82)
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.
