|
Abstract : |
This paper describes a study of dierent adaptations of boosting algorithms for costsensitive classi cation. The purpose of the study is to improve our understanding of the behavior of various cost-sensitive boosting algorithms and how variations in the boosting procedure aect misclassication cost and high cost error. We nd that boosting can be simplied for cost-sensitive classication. A new variant, which excludes a factor used in ordinary boosting, performs best at minimizing high cost errors and it almost always performs better than AdaBoost. We also nd that cost-sensitive boosting seeks to minimize high cost errors rather than cost, and a minimum expected cost criterion, applied during classi cation, greatly enhances the performance of all cost-sensitive adaptations of boosting algorithms. We show a strong correlation between an algorithm that produces small model size and its success in reducing high cost errors. For a recently proposed method, AdaCost, we nd that a poor selection of cost adjustment factor induces poor performance, and suggest a modication that improves performance substantially. 1., |
