Imbalanced data classification is a common problem in applications related to the detection of anomalies, failures, and risks. Since previous problem-solving approaches were basically heuristic and task dependent, we propose a novel imbalanced data classifier with a theoretical problem-solving approach. Our proposed method fine-tunes the parameters of kernel logistic regression using the harmonic mean of such criteria as sensitivity and positive predictive value, which are derived based on a confusion matrix and are essential for multilateral evaluation. This paper presents the formulation of our proposed method and reports our empirical evaluation results.
展开▼
