Model-X knockoffs is a new statistical tool that reliably selects which of the many potentially explanatory variables of interest (e.g. the absence or not of a mutation) are indeed truly associated with the response under study (e.g. the risk of getting a specific form of cancer). This framework can deal with very complex statistical models; in fact, they may be so complex that they can be treated as black boxes. The idea is to construct fake variables - knockoffs - which obey some crucial exchangeability properties with the real variables we wish to assay so that they can be used as negative controls. To leverage the full power of this framework, however, we need flexible tools to construct knockoffs from sampled data. This talk presents a machine that can produce knockoffs for arbitrary and unspecified data distributions, using deep generative models. The main idea is to iteratively refine a knockoff sampling mechanism until a criterion measuring the validity of the knockoffs we produce is minimized. Extensive numerical experiments and quantitative tests confirm the generality, effectiveness, and power of our approach. This results in a model-free variable selection method, and we present an application to the study of mutations linked to changes in drug resistance in the human immunodeficiency virus.
Speaker: Yaniv Romano, Stanford
Contact:Website: Click to Visit
Save this Event:iCalendar
Windows Live Calendar