Toward scalable learning with non-uniform class and cost distributions: a case study in credit card fraud detection

Abstract

Very large databases with skewed class distributions and non-uniform cost per error are not uncommon in real-world data mining tasks. One such task is credit card fraud detection: the number of fraudulent transactions is small compared to legitimate ones, the amount of financial loss for each fraudulent transaction depends on the transaction amount and other factors, and millions of transactions occur each day. We devised a multi-classifier meta-learning approach to address these three issues. Our empirical results indicate that the approach can significantly reduce loss due to illegitimate transactions. Keywords: skewed class distributions, non-uniform error cost, large amounts of data, credit card fraud detection. This work was partially funded by grants from DARPA (F30602-96-1-0311), NSF (IRI-96-32225 & CDA96 -25374), NYSSTF (423115-445). 1 Introduction Very large databases with skewed class distributions and non-uniform cost per error are not uncommon in real-world data mining...