Research map / Probabilistic and scientific modeling / Treatment effects
Contrastive Balancing Representation Learning for Heterogeneous Dose-Response Curves 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 CRNet, which learns double balancing representations (independent of both continuous treatments and outcomes given covariate representation) via a contrastive regularizer based on a partial distance measure plus an MSE outcome loss, for heterogeneous dose-response curve estimation.
Problem
Estimating heterogeneous dose-response curves with continuous treatments from observational data, where treatment-independent representations can discard outcome-relevant covariate information and propensity-based constraints can lose information.
Approach
Defines a double balancing representation (conditional independence of covariates from both treatment and potential outcome given the representation) and enforces it with a contrastive regularizer that uses partial distance correlation between covariate representations and treatments, contrasting real covariates (positives) against row-shuffled covariates (negatives), combined with an MSE outcome
Key results
- CRNet achieves MISE 1.69 on Data-1 vs 1.83 for the best baseline (Causal Forest), and 7.55 on Data-10 vs 14.9 for Causal Forest.
- On IHDP-1 CRNet reports MISE 0.22 vs 0.65 for SCIGAN and 0.79 for Causal Forest.
- Removing the prognostic (MSE) loss (w/o PR) degrades MISE to roughly 52-55 on synthetic Data-1/2/5/10, while removing the balancing loss (w/o BR) gives 2.04 on Data-1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CRNet instead of the strongest listed baseline | HDRC estimation · Data-1, Data-2, Data-5, Data-10, IHDP-1, News-2, News-4, News-8, News-16 | MISE: Data-1: 1.83 ± 0.6 -> 1.69 ± 0.5; Data-2: 2.50 ± 0.7 -> 2.07 ± 0.8; Data-5: 5.16 ± 0.9 -> 3.05 ± 0.7; Data-10: 14.9 ± 2. | Helped |
| add balancing representation (BR) to CRNet (ablation) | CRNet · Data-1, Data-2, Data-5, Data-10, IHDP-1, News-2, News-4, News-8, News-16 | MISE: Data-1: 2.04 ± 0.5 -> 1.69 ± 0.5; Data-2: 2.56 ± 1.0 -> 2.07 ± 0.8; Data-5: 4.76 ± 1.2 -> 3.05 ± 0.7; Data-10: 9.69 ± 5. | Helped |
| add prognostic representation (PR) to CRNet (ablation) | CRNet · Data-1, Data-2, Data-5, Data-10, IHDP-1, News-2, News-4, News-8, News-16 | MISE: Data-1: 52.9 ± 16 -> 1.69 ± 0.5; Data-2: 53.9 ± 14 -> 2.07 ± 0.8; Data-5: 51.0 ± 13 -> 3.05 ± 0.7; Data-10: 55.2 ± 16 -> | Helped |
| increase negative-sample shuffle count from m=0 to m=1 (ablation) | CRNet · Data-1, Data-10, IHDP-1, News-16 | MISE: Data-1: 1.99 ± 0.6 -> 1.69 ± 0.5; Data-10: 11.0 ± 3.1 -> 7.55 ± 2.6; IHDP-1: 0.58 ± 0.3 -> 0.22 ± 0.1; News-16: 10.2 ± 2 | Helped |
| increase negative-sample shuffle count beyond m=1 (ablation) | CRNet · IHDP-1 and News-16 | MISE: IHDP-1: 0.22 ± 0.1 (m=1) -> 0.24 ± 0.1 (m=2); News-16: 9.18 ± 2.9 (m=1) -> 10.1 ± 4.0 (m=2) | Mixed |
Surprising findings
- Naively minimizing the partial distance measure without negative samples (m=0) degrades performance, which the authors attribute to mode collapse.
- Increasing the number of negative shuffles beyond m=1 does not consistently improve results, and the default m=1 is used for efficiency.
Limitations
- Relies on the unconfoundedness assumption; unmeasured confounders may bias results.
- Conclusions may depend on the choice of conditional dependence measure.
- Partial distance of zero does not strictly guarantee conditional independence in non-Gaussian cases.
Experimental setup
Datasets: IHDP News Synthetic Data-q (simulated)
Benchmarks: Synthetic Data-1/2/5/10 IHDP-1 News-2/4/8/16
Metrics: MISE std over 30 replications
Baselines: GPS CBGPS Causal Forest SCIGAN DRNet VCNet
Builds on: DRNet VCNet SCIGAN covariate balancing propensity score
Closest papers by meaning
- Adversarially Balanced Representation for Continuous Treatment Effect Estimation (AAAI 2024 · similarity 0.88)
- Counterfactual Contrastive Learning with Normalizing Flows for Robust Treatment Effect Estimation (ICML 2025 · similarity 0.86)
- Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.85)
- Continuous Treatment Effect Estimation Using Gradient Interpolation and Kernel Smoothing (AAAI 2024 · similarity 0.85)
- Conditional Counterfactual Mean Embeddings: Doubly Robust Estimation and Learning Rates (NeurIPS 2026 · similarity 0.84)
- Enhancing Treatment Effect Estimation via Active Learning: A Counterfactual Covering Perspective (ICML 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.