An adaptive Cost-sensitive Classifier
Xiaolin Chen, Enming Song, Guangzhi Ma · 2010
Balancing Recall and Precision of rare class in cost-sensitive classification is a general problem. In this paper, we propose a novel cost-sensitive learning algorithm, named Adaptive Cost Optimization (AdaCO), which uses the resampling and genetic algorithm to build convex combination composite classifiers. In every base classifier's building, we use G-mean over Recall and Precision of rare class as the fitness function to find the optimal balance point in a reasonable misclassification costs space. We empirically evaluate and compare AdaCO with Cost-sensitive SVM (C-SVM in short) and CostSensitiveClassifier (CSC in short) over 6 realistic imbalanced bi-class datasets from UCI. The experimental results show that AdaCO does not sacrifice one class for the sake of the other, but produces high predictions on both classes.