survivor bias in observational data leads the optimization of recommender
systems towards local optima. Currently most solutions re-mines existing
human-system collaboration patterns to maximize longer-term satisfaction by
reinforcement learning. However, from the causal perspective, m