This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).
Traditional database systems are designed to provide the user with a clear picture of past states of the world as is represented in the database. Recently, stream processing systems are introduced to produce near real-time answers for those applications applications that require up-to-date information. Thus, we see a trend toward shrinking the ?reality gap? to zero. But for some applications, even real-time is not good enough. There is often a desire to get out in front of the present by delivering predictions of future events to take advantage of opportunities or to avert calamity. Security applications are a good example since they are typically interested in preventing a breach rather than simply reporting that one has happened. There is currently no database system that can effectively serve as a generic platform to support such predictive applications.
This project aims to fill this gap by designing and building a prototype database system called "Longview" to enable data-centric predictive analytics. Longview facilitates the use of statistical models to analyze historical and current data and make predictions about future data values and events. Users can plug new predictive models into the system along with a modest amount of meta-data, and the system uses those models to efficiently evaluate predictive queries.
Longview treats predictive models as first-class citizens by intelligently managing them in the process of data management and query optimization. This involves automatically building models and determining when and which model(s) to apply to answer predictive queries. This also involves creating and using the proper physical data structures to facilitate efficient model building, selection, and execution. Longview handles both streaming and historical queries. In fact, many streaming queries need efficient access to an archive of past values, making it necessary to seamlessly combine both stream and historical processing. Finally, Longview investigates "white-box" model support, in which the database leverages the operational semantics and representation of models to improve performance.
Longview's goal is to make it much easier to build predictive analytics applications in data intensive situations. Seamlessly combining data and model management is key to make the process of computing with predictions far easier to express and more efficient than the current ad-hoc application-level approaches. The resulting technology also allows for a better understanding and support for user-defined functions in database systems.
Longview is initially used for a real-world sensor-based tracking application and a predictive web portal for easy experimentation with different models and data sets. Further information on the project can be found on the project web page: http://database.cs.brown.edu/projects/longview/
