Sequential community mode estimation

Abstract

Several applications in online learning involve sequential sampling/polling of an underlying population. A classical learning task in this space is online cardinality estimation, where the goal is to estimate the size of a set by sequential sampling of elements from the set (see, for example, [2,4,7]). The key idea here is to use 'collisions,' i.e., instances where the same element is sampled more than once, to estimate the size of the set. Another recent application is community exploration, where the goal of the learning agent is to sample as many distinct elements as possible, given a family of sampling distributions/domains to poll from (see [3, 6]).