Distributed predictive and descriptive data mining
Sabine McConnell · 2006
Over the past decade, data mining has gained a strong foothold in a variety of application areas, including including science, engineering, and commerce. Centralized techniques, assuming all data to be contained at a single site, have been successfully applied to a large variety of application domains. However, we currently observe a trend toward a distributed extraction of knowledge from datasets. The motivation for this trend includes the distributed nature of the data itself, the distributed nature of computational resources, privacy concerns, and pragmatic issues arising from the increasing amount of available data. The development of distributed data-mining techniques is therefore necessary to extend the success of centralized data-mining techniques to the distributed domain. We present our research in the area of distributed data-mining. Specifically, we introduce a light-weight technique, in which predictive and descriptive models are built locally from subsets of attributes, and combined to produce global descriptive or predictive results. We modify our technique to facilitate data-mining in sensor networks for static and dynamic data. Our approach enables us to use distributed predictive and descriptive data-mining techniques to investigate datasets that are too large to be examined using centralized methods, are sensitive to privacy issues, and are under the constraint of real-time deadlines.