Research map / Probabilistic and scientific modeling / Causal discovery
Task-Driven Causal Feature Distillation: Towards Trustworthy Risk Prediction
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 TDCFD, which distills each feature into a causal feature attribution via propensity-score estimation with adaptive group Lasso over a DNN relational graph, then trains a DNN on the distilled data for trustworthy, high precision/recall risk prediction.
Problem
Risk prediction models are black boxes that rely on correlations rather than causal features, and they perform poorly on precision and recall under severe class imbalance.
Approach
Replace each raw feature value with a causal feature attribution (the interventional expectation E[Y(x_j)] minus a baseline), estimated via potential outcome framework with a propensity score learned by adaptive group Lasso that keeps confounders/adjustment variables and removes instrumental/spurious ones, then feed the distilled data to a standard DNN.
Key results
- On the real Alipay corporate risk data, TDCFD reports precision 0.86 and recall 0.80 versus 0.71 and 0.71 for the best baseline (Transformer).
- On the synthetic risk data, TDCFD reaches precision 0.82 and recall 0.90, above the best baseline (Transformer at 0.77 and 0.85).
- On News-4/8/16 causal effect benchmarks, TDCFD has the lowest epsilon_mATE (8.77, 5.93, 5.04), though on News-2 PM is slightly better (3.99 vs 4.25).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TDCFD instead of PM | News-2 ATE estimation · News-2 | ε_ATE: 3.99 -> 4.25 | Hurt |
| Use TDCFD instead of PD | News-4 mATE estimation · News-4 | ε_mATE: 8.47 -> 8.77 | Hurt |
| Use TDCFD instead of PM | News-8 mATE estimation · News-8 | ε_mATE: 6.51 -> 5.93 | Helped |
| Use TDCFD instead of PM | News-16 mATE estimation · News-16 | ε_mATE: 5.76 -> 5.04 | Helped |
| Use TDCFD instead of Transformer | risk prediction · Synthetic risk prediction | precision: 0.77 -> 0.82 | Helped |
| Use TDCFD instead of Transformer | risk prediction · Synthetic risk prediction | recall: 0.85 -> 0.90 | Helped |
| Use TDCFD instead of Transformer | risk prediction · Real corporate risk | precision: 0.71 -> 0.86 | Helped |
| Use TDCFD instead of Transformer | risk prediction · Real corporate risk | recall: 0.71 -> 0.80 | Helped |
| Use TDCFD instead of XGBoost | risk prediction · Synthetic risk, no hidden variable | precision: 0.91 -> 0.90 | Hurt |
| Use TDCFD instead of XGBoost | risk prediction · Synthetic risk, no hidden variable | recall: 0.96 -> 0.95 | Hurt |
| Use TDCFD instead of XGBoost | risk prediction · Synthetic risk, hidden variable | precision: 0.78 -> 0.82 | Helped |
| Use TDCFD instead of XGBoost | risk prediction · Synthetic risk, hidden variable | recall: 0.70 -> 0.89 | Helped |
Surprising findings
- With hidden variables present, DNN and XGBoost precision and recall drop sharply (e.g., DNN recall 0.92 to 0.65), while TDCFD keeps recall at 0.89.
Limitations
- Real corporate dataset is proprietary Alipay data and not released, limiting reproducibility.
- Relational graph is not a true causal graph, so causal claims depend on assumptions such as weak unconfoundedness.
- Causal effect benchmark results on News-2 do not show TDCFD as best.
Experimental setup
Datasets: News (causal effect benchmark, 2/4/8/16 interventions) synthetic Bayesian network risk data (20 features) real Alipay corporate risk data (114 features)
Benchmarks: News synthetic risk prediction data real corporate risk data (Alipay)
Metrics: epsilon_ATE epsilon_mATE accuracy precision recall
Baselines: kNN PSM RF CF BART GANITE PD TARNET CFRNET SITE PM LR
Builds on: Potential Outcome Framework adaptive group Lasso mixture density network
Closest papers by meaning
- Causal Concept Graph Models: Beyond Causal Opacity in Deep Learning (ICLR 2025 · similarity 0.83)
- Do causal predictors generalize better to new domains? (NeurIPS 2024 · similarity 0.83)
- Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation (ICML 2024 · similarity 0.82)
- CausalPC: Improving the Robustness of Point Cloud Classification by Causal Effect Identification (CVPR 2024 · similarity 0.82)
- Towards Causal Foundation Model: on Duality between Optimal Balancing and Attention (ICML 2024 · similarity 0.82)
- quanda: An Interpretability Toolkit for Training Data Attribution Evaluation (NeurIPS 2026 · similarity 0.81)
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.