Exceptional Knowledge Discovery in Databases Based on an Information-Theoretic Approach

Abstract

This paper presents an algorithm for discovering exceptional knowledge from databases. Exceptional knowledge, which is defined as an exception to a general fact, exhibits unexpectedness and is sometimes extremely useful in spite of its obscurity. Previous discovery approaches for this type of knowledge employ either background knowledge or domain-specific criteria for evaluating the possible usefulness, i.e. the interestingness of the knowledge extracted from a database. It has been pointed out, however, that the use of background knowledge can cause overlooking of useful knowledge. Furthermore, it is difficult to find such criteria in some domains. In order to circumvent these difficulties, we propose an information-theoretic approach in which we obtain exceptional knowledge associated with general knowledge in the form of a rule pair using a depth-first search method. The product of the ACEs (Average Compressed Entropies) of the rule pair is introduced as the criterion for evaluating the interestingness of exceptional knowledge. The inefficiency of depth-first search is alleviated by a branch-and-bound method, which exploits the upper-bound for the product of the ACEs. MEPRO (database Miner based on the average compressed Entropy PROduct criterion), which is a knowledge discovery system based on our approach, has been validated using the benchmark databases in the machine learning community.