Top-k vectorial aggregation queries in a distributed environment

Given a large set of objects in a distributed database, the goal of a top-k query is to determine the top-k scoring objects and return them to the user. Efficient top-k ranking over distributed databases has been the focus of recent research, with most current algorithms operating on the assumption that each node holds a single or small subset of each object's numerical attributes. However, in many important setups each node might hold instead a full d-dimensional vector of numerical attributes for each object. Examples include website activity in distributed servers, sales statistics for a retail chain, or share price information in different stock markets. For these setups, we define a novel ranking problem, top-kvectorial aggregation queries, where each object's score is determined by first aggregating the attribute vectors held for it and then applying the scoring function over the aggregated vector. Our communication-efficient algorithm uses a blend of geometric and skyline related machinery, some of which is newly developed, as well as an algorithmic framework for defining generic local constraints. Whereas previous algorithms have reduced data sharing by defining local thresholds for each attribute, such tailored solutions might perform poorly. Experimental results on real-world data demonstrate that our algorithm maintains low latency, with a communication cost up to four orders of magnitude lower than that of existing solutions.