Computational aspects around preference queries
Résumé
Preference queries present two main challenges: di culty for users to express their preferences and the computational complexity. For skyline queries, the preferences can be on attributes, e.g., some user may look for the best flights regarding price and number of stops, and others may look for the best flights regarding number of stops and duration. In addition, preference can be expressed as a (par-tial) order on attributes domains, e.g., some user may prefer flight company A over B while another one may have the opposite preference. For top-k queries, users define a score function to rank objects, e.g., users who give more importance to price could define the following score function: price ⇤ 2 + duration of the flight. In general, several rounds are required before converging towards a satisfying answer where at each round, more precise preferences are given by the user. This is due to the di culty to figure out the precise formulation of the user's preferences. Therefore, a more or less high number of queries need to be evaluated. Our research work aims to make these queries answering faster through dedicated index structures and precomputed views. The main challenges when adopting this strategy are (i) lightweight memory consumption and (ii) fast maintenance process. Our first step was NSC, an index structure that optimizes skyline queries. However, the structure was designed for a static context making it unsuitable when data can be inserted/deleted. We redesigned NSC to cope with dynamic data and in some cases, we proposed further approaches when the structure is not suitable. In this paper, we summarize our previous contributions and present some perspective research regarding the link between regret minimization queries and what we did so far.