Aggregation versus selection bias, and relational neural networks
Hendrik Blockeel, Maurice Bruynooghe · Lirias · 2003
Current relational learners handle sets either by aggregating over them or by selecting specific elements, but do not combine both. This imposes a significant, possibly undesirable bias on these learners. We discuss this bias, as well as some ideas on how to lift it. In the process, we introduce the notion of relational neural networks. 1 Biases of Relational Learners Among the many approaches to relational model learning that currently exist, a distinction can be made with respect to how they handle one-to-many and many-to-many relations, or, equivalently, how they handle sets of objects. To illustrate this, consider a database with just a single relation “Person ” with attributes Mother, Father, and Sex.