Using Neural Network to Weight the Partial Rules: Application to Classification of Dopamine Antagonist Molecules

In this paper, we propose an approach which can help Inductive Logic Programming (ILP) in multiclass domains and also its application to a real world domain, Classification of Dopamine Antagonist Molecules. When we classify an example by using the unordered rules constructed by standard ILP systems in multiclass domains, an example may match with the rules from different classes or may match with no rule in the rule set. Thus, using the rules alone is insufficient. We present the approach which utilises some matches in the rule to classify such examples. First, we extract the pieces of knowledge from the original rules, called partial rule. Then, we apply Neural Network to assign the importance to each partial rule. Finally, we use the weighted partial rules to classify unseen examples. Furthermore, the weights from Neural Network also show the importance of the piece of knowledge which is different from the knowledge originally represented in the form of First