Much of the real-world data generated these days is inherently uncertain or
probabilistic in nature. For instance, sensor data typically has some notion of
quality attached to it -- it could be the confidence of existence, trust,
accuracy, a probability distribution over possible values, or a mix of these.
Similarly, when attempting to integrate heterogenous data sources (data
integration) or extracting structured information from text (information
extraction), the results are approximate and uncertain at best.

In this talk, I will present an overview of our approach to integrate
statistical and probabilistic models into relational database systems so as to
make it easy to manage and reason about uncertain data. I will first discuss our
work on the MauveDB project that supports an abstraction called "model-based
views" using which users can specify statistical models to be applied to
streaming sensor data. The output of the modeling process is presented to the
users as a relational table that can be queried using a declarative language. I
will then present our ongoing work on query processing over uncertain relational
databases, which occur naturally in many settings such as information extraction
or may be the result of a probabilistic model-based view. We have developed a
uniform, closed framework for representing and querying uncertain data based on
concepts from probabilistic graphical models; I will present an overview of our
framework, the challenges in query evaluation over uncertain data, and the
algorithms we have developed for efficient query evaluation over large uncertain
databases.