Knowledge discovery using concept-class taxonomies

This paper describes the use of taxonomic hierarchies of concept-classes (dependent class values) for knowledge discovery. The approach allows evidence to accumulate for rules at different levels of generality and avoids the need for domain experts to predetermine which levels of concepts should be learned. In particular, higher-level rules can be learned automatically when the data doesn't support more specific learning, and higher level rules can be used to predict a particular case when the data is not detailed enough for a more specific rule. The process introduces difficulties concerning how to heuristically select rules during the learning process, since accuracy alone is not adequate. This paper explains the algorithm for using concept-class taxonomies, as well as techniques for incorporating granularity (together with accuracy) in the heuristic selection process. Empirical results on three data sets are summarized to highlight the tradeoff between predictive accuracy and predictive granularity.