Research map / Probabilistic and scientific modeling / Treatment effects
Continuous Treatment Effect Estimation Using Gradient Interpolation and Kernel Smoothing
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 GIKS, which augments training individuals with randomly sampled treatments and pseudo-labels them via gradient interpolation (for nearby treatments) and GP kernel smoothing with variance-based down-weighting (for distant treatments), to estimate individualized continuous treatment effects from observational data.
Problem
Individualized continuous treatment effect (ICTE) estimation from observational data, where treatment assignment is confounded with covariates in training but test-time treatments are independent of covariates.
Approach
Directly minimize per-instance counterfactual loss by augmenting training points with independently sampled treatments and pseudo-outcomes: first-order Taylor (gradient interpolation) for nearby treatments, and a GP over learned embeddings with neighbors in treatment space (kernel smoothing) for distant ones, down-weighting high-variance pseudo-labels.
Key results
- On CF error GIKS achieves 0.152, 0.080, 0.127, 1.891 and 1.079 on TCGA-0, TCGA-1, TCGA-2, IHDP and NEWS, beating all six baselines on every dataset.
- Adding GIKS losses to TARNet reduces CF error on TCGA-0/1/2 from about 1.68/1.42/3.40 to about 1.08/0.56/0.75, showing the approach is model-agnostic.
- On skin-lesion recourse, GIKS reaches 63% recourse accuracy versus 59-62% for baselines (averaged over 5 seeds).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the GI and GP counterfactual losses to factual training (ablation) | VCNet · TCGA-0, TCGA-1, TCGA-2, IHDP, NEWS | CF error: 0.18 (0.10) -> 0.15; 0.09 (0.08) -> 0.09; 0.17 (0.03) -> 0.12; 2.05 (0.00) -> 1.89; 1.09 (0.00) -> 1.07 | Helped |
| Add the GI loss without the GP loss (ablation) | VCNet · IHDP | CF error: 2.05 (0.00) -> 1.91 | Helped |
| Add the GP loss without the GI loss (ablation) | VCNet · TCGA-2 | CF error: 0.17 (0.03) -> 0.16 | Helped |
| Add the GI and GP losses (ablation) | VCNet · TCGA-1 | CF error: 0.09 (0.08) -> 0.09 | No clear effect |
| Replace the cosine kernel with the NTK kernel (ablation) | GIKS · IHDP | CF error: 1.842 -> 1.925 (0.35) | Hurt |
| Replace the cosine kernel with the dot product kernel (ablation) | GIKS · IHDP | CF error: 1.842 -> 1.882 (0.43) | Hurt |
| Use marginal treatment sampling instead of uniform sampling (ablation) | GIKS · IHDP and NEWS | CF error: 1.891 -> 1.902 (0.42); 1.079 -> 1.086 (0.09) | No clear effect |
| Use inverse propensity treatment sampling instead of uniform sampling (ablation) | GIKS · IHDP and NEWS | CF error: 1.891 -> 1.903 (0.41); 1.079 -> 1.080 (0.36) | No clear effect |
| Integrate GIKS losses with TARNet instead of factual-only TARNet | TARNet · TCGA-0, TCGA-1, TCGA-2, IHDP, NEWS | CF error: 1.678 -> 1.077; 1.465 -> 0.555; 3.396 -> 0.746; 3.068 -> 3.037; 1.163 -> 1.159 | Helped |
| Integrate GIKS losses with DRNet instead of factual-only DRNet | DRNet · TCGA-0, TCGA-1, TCGA-2, IHDP, NEWS | CF error: 1.673 -> 1.073; 1.417 -> 0.556; 3.365 -> 0.747; 2.731 -> 2.532; 1.126 -> 1.129 | Helped |
| Integrate GIKS losses with TransTEE instead of factual-only TransTEE | TransTEE · TCGA-0, TCGA-1, TCGA-2, IHDP, NEWS | CF error: 0.164 -> 0.165; 0.146 -> 0.132; 0.201 -> 0.172; 2.266 -> 2.023; 1.239 -> 1.176 | Mixed |
Surprising findings
- GIKS slightly worsens factual error on some datasets (e.g., NEWS 1.197 vs 1.107 for VCNet+HSIC), a trade-off the authors attribute to counterfactual supervision.
- The SciGAN GAN-based augmentation baseline often failed to converge, showing unstable training.
Limitations
- Relies on differentiability and overlap assumptions (A1, A3) that are not verified in practice.
- GP-based inference requires nearest-neighbor search over the entire training dataset, adding running-time cost.
- Evaluation relies largely on semi-synthetic benchmarks with simulated responses.
Experimental setup
Datasets: IHDP NEWS TCGA(0) TCGA(1) TCGA(2) SimGlucose
Benchmarks: IHDP NEWS TCGA(0) TCGA(1) TCGA(2)
Metrics: CF error factual error dosage policy error AMSE HSIC recourse accuracy
Baselines: TARNet DRNet SciGAN TransTEE VCNet+TR VCNet+HSIC
Builds on: VCNet DRNet TARNet TransTEE SciGAN
Closest papers by meaning
- Adversarially Balanced Representation for Continuous Treatment Effect Estimation (AAAI 2024 · similarity 0.91)
- PairNet: Training with Observed Pairs to Estimate Individual Treatment Effect (ICML 2024 · similarity 0.88)
- A Non-parametric Direct Learning Approach to Heterogeneous Treatment Effect Estimation under Unmeasured Confounding (NeurIPS 2024 · similarity 0.88)
- Counterfactual Contrastive Learning with Normalizing Flows for Robust Treatment Effect Estimation (ICML 2025 · similarity 0.88)
- Learning Representations of Instruments for Partial Identification of Treatment Effects (ICML 2025 · similarity 0.87)
- Enhancing Treatment Effect Estimation via Active Learning: A Counterfactual Covering Perspective (ICML 2025 · similarity 0.87)
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.