We consider the problem of Active Search, where a maximum of relevant objects- ideally all relevant objects - should be retrieved with the minimum effort orminimum time. Typically, there are two main challenges to face when tacklingthis problem: first, the class of relevant objects has often low prevalenceand, secondly, this class can be multi-faceted or multi-modal: objects could berelevant for completely different reasons. To solve this problem and itsassociated issues, we propose an approach based on a non-stationary (akarestless) extension of Thompson Sampling, a well-known strategy for Multi-ArmedBandits problems. The collection is first soft-clustered into a finite set ofcomponents and a posterior distribution of getting a relevant object insideeach cluster is updated after receiving the user feedback about the proposedinstances. The "next instance" selection strategy is a mixed, two-leveldecision process, where both the soft clusters and their instances areconsidered. This method can be considered as an insurance, where the cost ofthe insurance is an extra exploration effort in the short run, for achieving anearly "total" recall with less efforts in the long run.
展开▼
