Megadose The AI curve, compressed.

Joint Learning of Factual and Counterfactual Worlds for Causal Prior-data Fitted Networks

· ArXiv · AI/CL/LG ·
Learning both possible outcomes together improved treatment-effect estimates and uncertainty measures in most tested settings, the authors report.

Existing causal prior-data fitted networks train on synthetic data but learn only one outcome per unit, discarding dependence between factual and counterfactual outcomes. The proposed setup learns their joint distribution to derive response and individual treatment-effect densities. It then uses optimal transport across units to approximate the average treatment-effect density. Tests on synthetic, semi-synthetic and real benchmarks found significant improvements over the baseline models in most settings.

ArXiv · AI/CL/LG's note

score 4

Categories: Research