We present ABC3, a Bayesian active learning algorithm for learning conditional average treatment effect in causal inference setting. Theoretical analysis links ABC3 and maximum mean discrepancy (MMD), showing the importance of balanced sampling in accurate estimation.
