Rule Discovery: Error Measures and Conditional Rule Probabilities

摘要

Many promising rule discovery algorithms use their proprietary methods to measure the goodness (or error) of rules. This paper first compares various goodness (or error) measures, including average error, mean square error, probability difference and prediction factor, to evaluate their appropriateness for rule discovery. Secondly, we study a method of estimating conditional probabilities for a single rule and as a novelty, for rule sets. Results regarding how conditional probabilities affect the goodness of the discovered knowledge are presented.