In the previous seminar, S. Matwin showed the need for dimensionality reduction in relational learning. He also presented a change of representation from relational to attribute-value (AV) learning which enables to perform feature selection in relational learning.

This new AV learning problem is in fact a set of so-called multi-instance (MI) problems with some particular properties. More specifically the conversion process introduces inherent attribute and class noise. Furthermore, within this framework, the feature selection task is equivalent to the learning task. In this presentation, we will outline several algorithms used to learn in this noisy MI context. Some experiments have been performed on both artificial and real datasets. They show that good results can be achieved with a simple greedy algorithm.