|
Abstract:
|
Boosting is a very powerful method in classification. Among these methods, a logit boost is very useful because we can derive a discriminant score as a probability. However, because minimizing logistic loss is equivalent to maximizing accuracy, using logistic loss is inadequate when considering a certain evaluation metric. For example, using accuracy is inadequate when applying imbalanced data in which majority classes dominate over minority classes. This study developed an adaptive cost-sensitive logit boost, which is the extension method of cost-sensitive logit boost and misclassification costs can be adjusted in each step. We then proposed the misclassification cost to optimize the metrics. To define the misclassification costs for optimizing each performance metric, we estimated the classification model along with motivated performance metrics. Experimental results using real-world data reveal that the proposed method is efficient for optimizing each performance metric.
|
