Identifying best interventions through online importance sampling

摘要: Motivated by applications in computational advertising and systems biology, we consider the problem of identifying best out several possible soft interventions at a source node V an acyclic causal directed graph, to maximize expected value target Y (located downstream V). Our setting imposes fixed total budget for sampling under various interventions, along with cost constraints on different types interventions. We pose this as arm identification bandit K arms where each is intervention V, leverage information leakage among provide first gap dependent error simple regret bounds problem. results are significant improvement over traditional results. empirically show that our algorithms outperform state art Flow Cytometry data-set, also apply algorithm model interpretation Inception-v3 deep net classifies images.