Identifying Interesting Instances for Probabilistic Skylines

Uncertain data arises from various applications such as sensor networks, scientific data management, data integration, and location based applications. While significant research efforts have been dedicated to modeling, managing and querying uncertain data, advanced analysis of uncertain data is still in its early stages. In this paper, we focus on skyline analysis of uncertain data, modeled as uncertain objects with probability distributions over a set of possible values called instances. Computing the exact skyline probabilities of instances is expensive, and unnecessary when the user is only interested in instances with skyline probabilities over a certain threshold. We propose two filtering schemes for this case: a preliminary scheme that bounds an instance’s skyline probability for filtering, and an elaborate scheme that uses an instance’s bounds to filter other instances based on the dominance relationship. We identify applications where instance-level filtering is useful and desirable. Our algorithms can be easily adapted to filter at the object level if the application domain requires it. Moreover, the uncertain model we adopt in this paper allows missing probabilities of uncertain objects as well as arbitrary probability distributions over instances. We experimentally demonstrate the effectiveness of our filtering schemes on both the real NBA data set and the synthetic data set.