Research map / Label-efficient and robust learning / Recognition and long-tail
Active feature acquisition via explainability-driven ranking
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 an active feature acquisition method that ranks features per instance with local explanations (SHAP/LIME etc.) and trains a decision-transformer policy to predict the next feature in that ranking via a two-stage training scheme.
Problem
Active feature acquisition: selecting a small, instance-specific subset of features sequentially under a feature budget, when per-instance feature importance is unavailable at training time.
Approach
Uses local explanation methods (TreeSHAP on CatBoost for tabular, FastSHAP for images) to produce instance-wise feature importance rankings on training data, then reframes AFA as supervised feature prediction: a causal decision-transformer policy learns to predict the next-highest-ranked unacquired feature, with a second stage training on the policy's own imperfect acquisition trajectories.
Key results
- On ImageNette with 2 and 5 patches of 196, the method reaches 63.64% and 74.95% average accuracy.
- On CKD with 10 features, average AUROC is 0.8465.
- Second training stage improves average accuracy by 0.94% (CIFAR-100) to 7.35% (Metabric) over first stage alone; first-stage 250-epoch training gives nearly identical results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add second-stage training (ablation) | Explainability-driven AFA · CIFAR-10 | accuracy: 75.76±0.19% -> 78.44±0.15% | Helped |
| add second-stage training (ablation) | Explainability-driven AFA · Metabric | accuracy: 62.48±1.39% -> 69.83±0.41% | Helped |
| add second-stage training (ablation) | Explainability-driven AFA · CPS | accuracy: 67.21±0.15% -> 67.45±0.13% | No clear effect |
| add second-stage training (ablation) | Explainability-driven AFA · CTGS | AUROC: 0.916±.0004 -> 0.916±.0001 | No clear effect |
| use 250 first-stage epochs instead of second-stage training (ablation) | Explainability-driven AFA · BloodMNIST | accuracy: 79.83±0.19% -> 83.87±1.05% | Hurt |
| replace TreeSHAP rankings with LIME rankings (ablation) | Explainability-driven AFA · Metabric | accuracy: 69.8±0.41% -> 69.2±0.18% | Hurt |
| replace TreeSHAP rankings with INVASE rankings (ablation) | Explainability-driven AFA · CPS | accuracy: 67.5±0.13% -> 68.4±0.23% | Helped |
| use ResNet-10 instead of ResNet-18 to generate feature rankings (ablation) | Explainability-driven AFA on CIFAR-10 · CIFAR-10 test set | accuracy: 78.44% -> 78.22% | Hurt |
| replace the decision transformer policy with a ResNet block (ablation) | Explainability-driven AFA · CIFAR-10 | accuracy: 78.44% -> 76.83% | Hurt |
| replace the decision transformer policy with a ResNet block (ablation) | Explainability-driven AFA · CIFAR-100 | accuracy: 46.99% -> 46.70% | Hurt |
| replace the decision transformer policy with a CNN block (ablation) | Explainability-driven AFA · BloodMNIST | accuracy: 83.87% -> 78.23% | Hurt |
| use nearest-neighbor ranking to choose acquired features instead of the learned policy | Explainability-driven AFA · Metabric | accuracy: 69.8±0.41% -> 68.1±0.75% | Hurt |
Surprising findings
- Replacing the decision transformer with a ResNet block or CNN block lowers accuracy, but the method remains effective with alternative architectures.
- Extending first-stage training to 250 epochs does not match second-stage gains on most datasets (BloodMNIST is an exception).
Limitations
- Assumes uniform feature acquisition costs.
- Performance depends on quality of explanation-derived rankings; the pre-trained model capacity affects ranking quality.
- Higher training time than baselines.
Experimental setup
Datasets: Spambase Metabric CPS CTGS CKD CIFAR-10 CIFAR-100 BloodMNIST ImageNette
Benchmarks: Spambase Metabric CPS CTGS CKD CIFAR-10 CIFAR-100 BloodMNIST ImageNette
Metrics: accuracy AUROC
Baselines: DIME GDFS CAE OPL center-cropping random selection empirical oracle NN-based AACO-inspired baseline
Builds on: Decision Transformer GDFS DIME FastSHAP TreeSHAP
Closest papers by meaning
- Information Templates: A New Paradigm for Intelligent Active Feature Acquisition (NeurIPS 2026 · similarity 0.88)
- Stochastic Encodings for Active Feature Acquisition (ICML 2025 · similarity 0.87)
- Active Learning with LLMs for Partially Observed and Cost-Aware Scenarios (NeurIPS 2024 · similarity 0.84)
- Generator Assisted Mixture of Experts for Feature Acquisition in Batch (AAAI 2024 · similarity 0.83)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.83)
- Estimating Conditional Mutual Information for Dynamic Feature Selection (ICLR 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.