Query Reranking As A Service

Abstract: The ranked retrieval model has rapidly become the de facto way for search query processing in client-server databases, especially those on the web. Despite of the extensive efforts in the database community on designing better ranking functions/mechanisms, many such databases in practice still fail to address the diverse and sometimes contradicting preferences of users on tuple ranking, perhaps (at least partially) due to the lack of expertise and/or motivation for the database owner to design truly effective ranking functions. This paper takes a different route on addressing the issue by defining a novel {\em query reranking problem}, i.e., we aim to design a third-party service that uses nothing but the public search interface of a client-server database to enable the on-the-fly processing of queries with any user-specified ranking functions (with or without selection conditions), no matter if the ranking function is supported by the database or not. We analyze the worst-case complexity of the problem and introduce a number of ideas, e.g., on-the-fly indexing, domination detection and virtual tuple pruning, to reduce the average-case cost of the query reranking algorithm. We also present extensive experimental results on real-world datasets, in both offline and live online systems, that demonstrate the effectiveness of our proposed techniques.