Effective Rule Induction from Molecular Structures Represented by Labeled Graphs

Acyclic conjunctive queries form a polynomially evaluable fragment of definite nonrecursive first-order Horn clauses. Labeled graphs, a special class of relational structures, provide a natural way for representing chemical compounds. We propose an algorithm specific to learning acyclic conjunctive queries predicting certain properties of molecules represented by labeled graphs. To compensate for the reduced expressive power of the hypothesis language and thus the potential decrease in classification accuracy, we combine acyclic conjunctive queries with confidence-rated boosting. This approach leads to excellent prediction accuracy on the domain of mutagenicity.