BiBoost for Asymmetric Learning

Although boosting methods have become an extremely important classification method, there has been little attention paid to boosting with asymmetric losses. In this paper we take a gradient descent view of boosting in order to motivate a new boosting variant called BiBoost which treats the two classes differently. This variant is likely to perform well when there is a different cost for false positive and false negative predictions. The variant is also appropriate when the data comes from multiple sources with different reliabilities or noise levels. Experiments show that BiBoost effectively reduces the number of false positive mistakes, and a more general algorithm is discussed.