Positive-Unlabeled Learning for Control Group Construction in Observational Causal Inference