Cost-effective identification of on-topic search queries using multi-armed bandits

ABSTRACTIdentifying the topic of a search query is a challenging problem, for which a solution would be valuable in diverse situations. In this work, we formulate the problem as a ranking task where various rankers order queries in terms of likelihood of being related to a specific topic of interest. In doing so, an explore-exploit trade-off is established whereby exploiting effective rankers may result in more on-topic queries being discovered, but exploring weaker rankers might also offer value for the overall judgement process. We show empirically that multi-armed bandit algorithms can utilise signals from divergent query rankers, resulting in improved performance in extracting on-topic queries. In particular we find Bayesian non-stationary approaches to offer high utility. We explain why the results offer promise for several use-cases both within the field of information retrieval and for data-driven science, generally.