Handling Different Levels of Granularity within Naive Bayes Classifiers

Abstract

Data mining techniques usually require a flat data table as input. For categorical attributes, there is often no canonical flat data table, since they can often be considered in different levels of granularity (like continent, country or local region). The choice of the best level of granularity for a data mining task can be very tedious, especially when a larger number of attributes with different levels of granularities is involved. In this paper we propose two approaches to automatically select the granularity levels in the context of a naive Bayes classifier. The two approaches are based on the χ ² independence test including correction for multiple testing and the minimum description length principle.