AdaCC: cumulative cost-sensitive boosting for imbalanced classification

Abstract Class imbalance poses a major challenge for machine learning as most supervised models might exhibit bias towards the majority class and under-perform in minority class. Cost-sensitive tackles this problem by treating classes differently, formulated typically via user-defined fixed misclassification cost matrix provided input to learner. Such parameter tuning is challenging task that requires domain knowledge moreover, wrong adjustments lead overall predictive performance deterioration. In work, we propose novel cost-sensitive boosting approach imbalanced data dynamically adjusts costs over rounds response model’s instead of using matrix. Our method, called AdaCC, parameter-free it relies on cumulative behavior model order adjust next round comes with theoretical guarantees regarding training error. Experiments 27 real-world datasets from different domains high demonstrate superiority our method 12 state-of-the-art approaches exhibiting consistent improvements measures, instance, range [0.3–28.56%] AUC, [3.4–21.4%] balanced accuracy, [4.8–45%] gmean [7.4–85.5%] recall.