Title: Counterfactual Mean Embedding for Nonparametric Causal Inference
We cannot form any conception about the world around us without the idea of causal connection between things. Counterfactual is one of the most important concepts in causality. In this talk, I will discuss our causal inference framework based on counterfactuals. This work is motivated by the potential outcome framework of Rubin who characterizes the causal effect by the discrepancy between (possibly counterfactual) distributions over the outcome variable of treatment and control populations. To each an inference on these counterfactual distributions, we introduce a Hilbert space representation of a counterfactual distribution called counterfactual mean embedding (CME). The CME relies on the renowned RKHS embedding of a probability distribution. Such a representation allows us to leverage the distributional effect in a counterfactual analysis and offers elegant tools for the nonparametric causal inference. It has applications ranging from econometrics and social science to recommendation systems and ad placement.
joint work with Motonobu Kanagawa (ISM)
|Date||March 9, 2017 (Thu) 10:00 - 12:00|