Efficient processing of top-k queries: selective NRA algorithms

Efficient processing of top-k queries has drawn increasing attention from both industry and academia due to its varied applications. Lower access cost is a crucial concern for a top-k query processing. Typically, when answering a top-k query, there exist two types of accesses: sorted access and random access. In some scenarios, the latter is not supported by the data source. Fagin et al. proposed the No Random Access (NRA) algorithm (Fagin et al, J Comput Syst Sci 66:614–656, 2003) for this situation. In this paper, we motivate our work by a key observation of the NRA algorithm: the number of accesses could be further reduced by selectively (instead of in parallel) performing sorted accesses to different lists of the dataset. Based on this insight, we propose a Selective NRA (SNRA) algorithm aiming to cut down the unnecessary access cost. Later, we optimize the SNRA algorithm in terms of runtime cost and present the SNRA-opt algorithm. Furthermore, we address the problem of instance optimality theoretically and turn SNRA (and SNRA-opt) into instance optimal algorithms, termed as Hybrid-SNRA (HSNRA) and HSNRA-opt. Extensive experimental results show that our algorithms perform significantly fewer sorted accesses than NRA (and its state-of-the-art variations). In terms of runtime cost, the proposed SNRA-opt and HSNRA-opt algorithms are two orders of magnitude faster than the NRA algorithm. In addition, we discuss the parameter selection problem of the SNRA algorithms, both theoretically and experimentally.