On the LearnAbility of Abstraction Theories from Observations for Relational Learning

The most common methodology in symbolic learning consists in inducing, given a set of observations, a general concept definition. It is widely known that the choice of the right description language for a learning problem can affect the efficacy and effectiveness of the learning task. Furthermore, most of the real-world domain are contaminated by various kinds of imperfections in data such as inappropriateness of the description language which does not contain/facilitate an exact representation of the target concept. To deal with such kind of situations, Machine Learning approaches have moving from a framework exploiting a single inference mechanism, such as induction, towards one integrating multiple inference strategies such as abstraction. The literature so far assumed that the information needed to the learning systems to apply additional inference strategies is provided by the an expert domain. The objective in this work is the automatic inference of such information. The efficacy of the proposed method in generating effective theories to perform abstraction was tested by providing the generated abstraction theories to the learning system INTHELEX allowing it to exploit its multistrategy capabilities, in particular the abstraction one. Various experiments were carried out on a real-world application domain of scientific paper documents showing the validity of the proposed method.