Multi-relational classification in imbalanced domains

Abstract

This paper discusses the problem of multi-relational classification on imbalanced datasets. To solve the class imbalance problem, a new multi-relational Naive Bayesian classifier named R-NB is proposed, the attribute filter criterion based on mutual information is upgraded to deal with multi-relational data directly and the basic sampling methods include under-sampling and over-sampling are adopted to eliminate or minimize rarity by altering the distribution of relational examples. Experiments show, with the help of attribute filter method, R-NB can get better results than those without that. And, experiments show that multi-relational classifiers with under-sampling methods can provide more accurate results than that with over-sampling methods considering the ROC curve. © 2008 Springer Berlin Heidelberg.