Skimmed Classifiers

Rogerio Salvini, Universidade Federal do Rio de Janeiro, Brazil
Eduardo Aguilar, Universidade Federal do Rio de Janeiro, Brazil
Inês Dutra, Universidade do Porto, Portugal

Most Inductive Logic Programming (ILP) systems use a greedy covering algorithm to find a set of clauses that best model positive examples. This set of clauses is called a theory and can be seen as an ensemble of clauses. It turns out that the search for a theory within the ILP system is very time consuming and often yields overly complex classifiers. One alternative approach to obtain a theory is to use the ILP system to non deterministically learn one clause at a time, several times, and to combine the obtained clauses using ensemble methods.