Abstract

Many factors influence the performance of a learned classifier. In this paper we study different methods of measuring performance based on a unified set of cost models and the effects of training class distribution with respect to these models. Observations from these effects help us devise a distributed multi-classifier meta-learning approach to learn in domains with skewed class distributions, non-uniform cost per error, and large amounts of data. One such domain is credit card fraud detection and our empirical results indicate that, up to a certain degree of skewed distribution, our approach can significantly reduce loss due to illegitimate transactions. Introduction Inductive learning research has been focusing on devising algorithms that generate highly accurate classifiers. Many factors contribute to the quality of the learned classifier. One factor is the class distribution in the training set. Using the same algorithm, different training class distributions can generate classi...

Citations