This paper examines the idea of incorporating machine learning algorithms into a database system for monitoring its stream of incoming queries and generating hierarchies with the most important concepts expressed in those queries. The goal is for these hierarchies to provide valuable input to the database administrator for dynamically modifying the physical and external schemas of a database for improved system performance and user productivity. The criteria for choosing the appropriate learning algorithms are analyzed, and based on them, two such algorithms, UNIMEM and COBWEB, are selected as the most suitable ones for the task. Standard UNIMEM and COBWEB implementations have been modified to support queries as input. Based on the results of experiments with these modified implementations, the whole approach appears to be quite promising, expecially if the concept hierarchy from which the learning algorithms start their processing is initialized with some of the most obvious concepts captured in the database.