Using Domain Knowledge to Improve Inductive Learning Algorithms for Diagnosis

Gerhard Friedrich, Wolfgang Nejdl

Abstract: Publisher Summary This chapter presents an important similarity based technique, ID3-like classification tree generation, which can be improved using additional domain theory axioms. The constructed decision trees use generalized expression labels based on the domain theory instead of constants yielding a more knowledge intensive and compact representation. The generalized attributes of the tree representation include all knowledge present in the domain theory and are, therefore, information theoretically more relevant than simple constant attributes. The generated classification tree is much more compact than the original one. Also, the system knowledge is not only used for generating diagnosis examples but also for transforming these examples into a form with information theoretically more relevant attributes. The algorithm can be applied to compile diagnosis knowledge into efficient decision tree format. This approach extends model-based diagnosis systems that also use an information theoretic function to select measurement points.