If data mining is to be performed on multiple database tables and using foreign keys, then relational approaches, rather than attribute-value (AV) methods, are necessary. Relational approaches, however, suffer from the curse of dimensionality. Therefore, there is a strong interest in being able to reduce the dimensionality of relational tasks. By analogy with AV learning, selecting only the features likely to be relevant in the learning task seems an attractive direction. But it is far from clear how to perform this task in relational learning, as even the notion of attribute is not well defined when examples are expressed in first-order notation. In this presentation, the first in a series of two, we outline a change of representation from relational to AV learning. Applying this change allows us to use the successful AV feature selection algorithms. We show some preliminary results. Continuation of this work will be presented by N. Stroppa in the next seminar.