Cleaning and querying large uncertain databases

Abstract

The management of uncertain databases has recently attracted tremendous interest from both industry and academy communities. In particular, there is a need to handle uncertain data in many emerging applications, such as the wireless sensor network, biometric and biological databases, location-based services, and data stream applications. To obtain meaningful results over these uncertain data, probabilistic queries are proposed, which augment query results with confidence. Although probabilistic queries are useful, evaluating them is costly, in terms of both I/O and computation. Moreover, the calculation of answer probabilities involves expensive numerical integrations. Therefore the efficient evaluation of probabilistic queries is a challenge for uncertain database management. In this thesis, we report our works for speeding up the evaluation performance of three kinds of important probabilistic queries - nearest-neighbor queries, fc-nearest-neighbor queries, and imprecise location-dependent queries. New approaches are proposed to improve the efficiency in both I/O and computation, and they are evaluated by extensive simulations over real and synthetic data sets. Another important issue that we consider in this thesis is the cleaning of uncertain data with the goal of achieving higher quality. Since the applications handling imprecise data have resource limitation, the cleaning process must optimize the use of resources. We study theoretically and experimentally on how the result quality could be maximized with constrained resources, with the use of entropy-based metrics. We also outline the future directions of our work.