Research map / Probabilistic and scientific modeling / Treatment effects
DISCRET: Synthesizing Faithful Explanations For Treatment Effect Estimation
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 DISCRET, a self-interpretable ITE framework that synthesizes rule-based explanations via deep Q-learning, which serve as database queries for consistent treatment effect estimation; regularizing black-box TransTEE with it sets new accuracy results.
Problem
Individual treatment effect (ITE) estimation models that are deployed in critical settings such as healthcare need to be both accurate and faithfully explainable, but existing black-box models lack explanations, post-hoc explainers lack faithfulness guarantees, and self-interpretable models lose acc
Approach
Explanations are logic rules that double as database queries: for each sample a conjunctive rule is synthesized (features, thresholds, operators) and evaluated on the training database to select a similar subgroup whose ATE gives the prediction. Because the rule is non-differentiable, the rule generator is trained with deep Q-learning using a reward based on the squared outcome error of the retrie
Key results
- DISCRET achieves near 100% consistency on IHDP, TCGA and IHDP-C versus at most about 0.3 for post-hoc explainers (e.g., Anchor 0.293 on IHDP-C).
- DISCRET outperforms self-interpretable baselines on IHDP out-of-sample ATE error (0.150 vs 0.240 for Causal Forest), and reaches 0.001 vs 0.011 on EEEC text.
- TransTEE + DISCRET reduces ITE error by up to 18% on TCGA relative to the neural baselines and reaches the lowest error across datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Regularize TransTEE with DISCRET predictions | TransTEE · TCGA | εATE (out-of-sample): 0.067$\pm$0.008 -> 0.055$\pm$0.009 | Helped |
| Regularize TransTEE with DISCRET predictions | TransTEE · IHDP (in-sample) | εATE: 0.081$\pm$0.009 -> 0.082$\pm$0.009 | Hurt |
| Regularize TransTEE with DISCRET predictions | TransTEE · IHDP (out-of-sample) | εATE: 0.138$\pm$0.014 -> 0.120$\pm$0.014 | Helped |
| Use DISCRET predictions to regularize TransTEE | TransTEE · IHDP-C | AMSE: 0.112$\pm$0.008 -> 0.102$\pm$0.007 | Helped |
| Use DISCRET predictions to regularize TransTEE | TransTEE · News | AMSE: 0.063$\pm$0.005 -> 0.043$\pm$0.005 | Helped |
| Add propensity-score regularization to the reward (ablation) | DISCRET · Uganda | outcome error: 1.701$\pm$0.161 -> 1.662$\pm$0.136 | Helped |
| Add automatic hyperparameter fine-tuning to the reward (ablation) | DISCRET · Uganda | outcome error: 1.742$\pm$0.151 -> 1.701$\pm$0.161 | Helped |
| Increase the IHDP training database from 100 to 200 samples (ablation) | DISCRET · IHDP | ITE test error: 0.55$\pm$0.13 -> 0.47$\pm$0.10 | Helped |
| Increase the IHDP training database from 200 to 400 samples (ablation) | DISCRET · IHDP | ITE test error: 0.47$\pm$0.10 -> 0.32$\pm$0.15 | Helped |
| Increase the IHDP training database from 400 to 747 samples (ablation) | DISCRET · IHDP | ITE test error: 0.32$\pm$0.15 -> 0.21$\pm$0.05 | Helped |
| Increase decision-tree depth from 6 to 100 (ablation) | Decision Tree · IHDP (out-of-sample) | εATE: 0.613$\pm$0.045 -> 0.549$\pm$0.052 | Helped |
| Increase decision-tree depth from 6 to 100 (ablation) | Decision Tree · TCGA (out-of-sample) | εATE: 0.202$\pm$0.012 -> 0.445$\pm$0.004 | Hurt |
Surprising findings
- DISCRET outperforms the black-box models on the Uganda image dataset, attributed to the highly concentrated outcome distribution.
- Post-hoc explanations (Lore, Lime, Anchor) have near-zero consistency on most datasets despite being standard practice.
Limitations
- Requires users to fix the rule grammar and hyperparameters such as number of literals before training.
- Relies on interpretable symbols or concepts extracted from unstructured data (images, text), which may not be available.
- Slightly underperforms state-of-the-art black-box models, notably TransTEE on IHDP-C.
Experimental setup
Datasets: IHDP TCGA IHDP-C News EEEC Uganda
Benchmarks: IHDP TCGA IHDP-C News EEEC Uganda
Metrics: epsilon_ATE in-sample epsilon_ATE out-of-sample AMSE outcome error consistency sufficiency
Baselines: Decision Tree Random Forest NAM ENRL Causal Forest BART Dragonnet TVAE TARNet Ganite DRNet VCNet
Builds on: TransTEE Deep Q-learning Anchor Causal Forest
Closest papers by meaning
- PairNet: Training with Observed Pairs to Estimate Individual Treatment Effect (ICML 2024 · similarity 0.86)
- AutoCATE: End-to-End, Automated Treatment Effect Estimation (ICML 2025 · similarity 0.85)
- Conformal Inference of Individual Treatment Effects Using Conditional Density Estimates (AAAI 2025 · similarity 0.85)
- An effective framework for estimating individualized treatment rules (NeurIPS 2024 · similarity 0.85)
- Impossibility of Distribution-Free Predictive Inference for Individual Treatment Effects (NeurIPS 2026 · similarity 0.85)
- Coupling Generative Modeling and an Autoencoder with the Causal Bridge (NeurIPS 2025 · similarity 0.85)
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.