Flexibly Exploiting Prior Knowledge in Empirical Learning

This paper presents a method to incorporate knowledge from possibly imperfect models and domain theories into inductive learning of decision trees for classification The approach assumes that a model or domain theory reflects useful prior knowledge of th< task Thus the default bias should accept the model s predictions as accurate even in the face of somewhat contradictory data which may be unrepresenlative or noisy However our approach allows the svslem to abandon the model or domain theorv, or portions thereof in the fact of sufficientlv contradictory data In particular we use C4 5 to induce decision trees from data that ha\t heen augmented b\ model or domaintheory-denvcd features' We weakly bias the svslem to select model-derived features dur ing decision tree induction but this preference is not dogmatically applied Our experiments vary imperfection in a model the representa tiveness of data and the veracitv with which modf l -demed feature are preferred 1 I n t r o d u c t i o n When human expertise is nonexistent or very weak relative to a particular domain/task and when data is plentiful machine induction from data mav be the only reasonable approach to task automation In contrast, when expertise is strong, then encoding the expert s model or domain theory via traditional knowledge acquisition strategies ma> be the best approach In fact, this human expertise may stem from induction over a much larger data sample than is available at the time task automation is undertaken In many cases, however, conditions are indeterminate as to whether sole reliance on machine induction or human expertise is most appropriate human expertise may not be 'perfect and/or data may not be as plentiful as desired in cases where some data is available and human expertise is less than perfect an advantageous strategy may be to exploit both in an appropriate way There is a growing body of work that combines modelbased or domain-theory knowledge with empirical learning from data Clark and Matwin [1993] assume that D o u g F i s h e r Computer Science Department Vanderbilt University Nashville, Tennessee 37235