Risk-Optimal Prediction under Unseen Causal Perturbations

ICLR 2026 Conference Submission18084 Authors

Keywords: causal inference, risk-optimal prediction, out-of-distribution, zero-shot

TL;DR: We propose a framework for predicting causal effects under unseen interventions. Our method achieves environment-specific risk-optimality without test covariate data and outperforms baselines on real-world gene and compound datasets.

Abstract:

Predicting intervention effects is important in various scientific fields, including biomedicine. Classical methods depend on fully specified causal graphs and extensive observational data, while recent invariance-based approaches typically assume access to the state of perturbed features.
These assumptions may not hold in practical settings with unknown causal relationships, partial and limited interventional data, and the need to consider novel, untested interventions/perturbations.
We propose a novel framework for causal effect estimation under such conditions that uses interventional embeddings to capture perturbation-specific information. Leveraging ideas from causality and robust learning, we propose a predictor that
targets a form of interventional regime-specific risk-optimality but that does so using transformations of available data and hence does not require access to interventional data from the target regime.
We put forward an end-to-end attention-based model that jointly learns embedding transformations and similarity-based weighting, enabling scalable prediction of causal effects even when no features are observed under intervention. Experiments on synthetic and real-world datasets show that our framework generalizes effectively to unseen interventions, hence addressing a critical challenge in prediction of causal effects in complex, real-world settings.

Primary Area: causal reasoning

Submission Number: 18084