A New Perspective on Pool-Based Active Classification and False-Discovery Control

摘要: In many scientific settings there is a need for adaptive experimental design to guide the process of identifying regions search space that contain as true positives possible subject low rate false discoveries (i.e. alarms). Such could differ drastically from predicted set minimizes 0/1 error and accurate identification require very different sampling strategies. Like active learning binary classification, this cannot be optimally chosen priori, but rather data must taken sequentially adaptively in closed loop. However, unlike classification with error, collecting find high positive discovery (FDR) not well understood. paper, we provide first provably sample efficient algorithm problem. Along way, highlight connections between combinatorial bandits, FDR control making contributions each.