Granularity Conscious Modeling for Probabilistic Databases

The convergence of embedded sensor systems and stream query processing suggests an important role for database techniques, in managing data that only partially ­ and of- ten inaccurately ­ capture the state of the world. Reasoning about uncertainty as a first class citizen, inside a database system, becomes an increasingly important operation for processing non deterministic data. An essential step for such an approach lies in the choice of the appropriate un- certainty model, that captures the probabilistic information in the data, both accurately and at the right semantic de- tail level. This paper introduces Hierarchical First-Order Graphical Models (HFGMs), an intuitive and economical representation of the data correlations stored in a Proba- bilistic Data Management system, in a hierarchical setting. HFGM semantics allow for an efficient summarization of the probabilistic model that can be induced from a dataset at various levels of granularity, effectively controlling the trade-off of the model´s complexity vs its accuracy.