Learning Decomposed Representation for Counterfactual Inference