Learning with Non-uniform Class and Cost Distributions: Eeects and a Distributed Multi-classiier Approach

Many factors innuence a learning process and the performance of a learned classiier. In this paper we investigate the eeects of class distribution in the training set on performance. We also study diierent methods of measuring performance based on cost models and the performance eeects of training class distribution with respect to the diierent cost models. Observations from these eeects help us devise a distributed multi-classiier 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 our approach can signiicantly reduce loss due to illegitimate transactions.