Efficient integration of data mining techniques in DBMSs

We propose in this paper a new approach for applying data mining algorithms, and more particularly supervised machine learning algorithms, to large databases, in acceptable response times. This goal is achieved by integrating these algorithms within a Database Management System. We are thus only limited by disk capacity, and not by available main memory. However, the disk accesses that are necessary to scan the database induce long response times. Hence, we propose in this paper an original method to reduce the size of the learning set by building its contingency table. The machine learning algorithms are then adapted to operate on this contingency table. In order to validate our approach, we implemented the ID3 decision tree construction method and showed that using the contingency table helped us obtaining response times equivalent to those of classical, in-memory software.