Research map / Label-efficient and robust learning / Recognition and long-tail
Flow-Disentangled Feature Importance
Xingshu Chen, Yifeng Guo, Jin-Hong Du
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 |
|---|---|---|---|
| strong correlation (rho=0.8) between X1 and X2 in a linear model | LOCO and nLOCO · Simulation M1 | importance of X1 | Hurt |
| use DFI (disentangled attribution) instead of LOCO/nLOCO under strong correlation | feature importance · Simulation M1 (rho=0.8) | importance assigned to X1: DFI >= 16 for X1 | Improved |
| use DFI instead of LOCO variants and Shapley in additive nonlinear model | feature importance · Simulation M2 | total importance vs signal variance | Improved |
| use DFI instead of LOCO-based methods (except dLOCO) with interactions | feature importance · Simulation M3 | importance ranking of correlated features | Improved |
| random-forest regression estimate instead of oracle regression function (ablation) | DFI · Simulation M3 | magnitude of DFI estimates | Hurt |
| nonlinear but linearly uncorrelated X1, X2 (misspecified for DFI) | DFI · Simulation M4 | attribution to X2 | Mixed |
| vary entropic regularization epsilon from 0 to 1 (ablation) | DFI · Simulation M1 null-feature coverage | coverage: 0.942 (eps=0) vs 0.904-0.910 (eps>0) at rho=0, n=100 | No clear effect |
| avoid refitting submodels and conditional distribution estimation (DFI vs dLOCO and sampled Shapley) | DFI · Simulation computation time | computational time | Improved |
| group-level DFI attribution on HIV-1 neutralization data | DFI · HIV-1 neutralization resistance | significant features: 19 of 832 -> 30 of 832 (disentangled vs original space) | Improved |
Closest papers by meaning
- A Differentiable Rank-Based Objective for Better Feature Learning (ICLR 2025 · similarity 0.85)
- Conditional Feature Importance with Generative Modeling Using Adversarial Random Forests (AAAI 2025 · similarity 0.84)
- Feature Importance Disparities for Data Bias Investigations (ICML 2024 · similarity 0.84)
- Variable Importance in High-Dimensional Settings Requires Grouping (AAAI 2024 · similarity 0.84)
- Feature Importance Metrics in the Presence of Missing Data (ICML 2025 · similarity 0.84)
- Regional Explanations: Bridging Local and Global Variable Importance (NeurIPS 2025 · similarity 0.84)
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.
