Correlation-Based Refinement of Rules with Numerical Attributes
André Melo, Martin Theobald, Johanna Völker · The Florida AI Research Society · 2014
Learning rules is a common way of extracting useful information from knowledge or data bases. Many of such data sets contain numerical attributes. However, approaches like ILP or association rule mining are optimized for data with categorical values, and considering numerical attributes is expensive. In this paper, we present an extension to top-down ILP algorithms such as FOIL, which enables an efficient discovery of rules from data with both numerical and categorical attributes. Our approach comprises a preprocessing phase for computing the correlations between numerical and categorical attributes, as well as an extension to the ILP refinement step, which enables us to detect interesting candidate rules and to suggest refinements with relevant attribute combinations. We report on experiments with U.S. Census data, Freebase and DBpedia, and show that our approach helps to efficiently discover rules with numerical intervals.