We will present a general framework based on concepts of relational algebra for distance based learning over relational schemata. The advantage of the proposed framework is that it requires no transformation of the representation of data that come in the form of relational databases. It is directly applicable to any relational database without the need of type and mode definitions and conversions to logic programming as it is the case with most relational learning systems based on Inductive Logic Programming.

Our framework builds on the notions of tuples of relations and sets of tuples. We show how exploiting these elementary building blocks our learning examples are represented via tree like structures. In order to define distances between relational examples we will explore two avenues. Both of them are based on the definition of simple operators on tuples and sets of tuples which are subsequently combined in order to provide a global operator on the full relational structure. The first approach is based on the use of classical distances over tuples and sets of tuples and the second one on the definition of kernels.

The user of the system has at his disposal a number of possible distance operators from which he can choose, or alternatively, to what amounts to something like model selection, can let the system perform the selection automatically. Some results on well known relational datasets will be presented.