Detect Rare Events via MICE Algorithm with Optimal Threshold
Sung-Chiang Lin, Charlotte I. Wang, Zhenyu Wu, Yu‐Fang Chung · 2013
Class imbalanced classifications are important issues in machine learning since class imbalanced problems usually happen in real applications such as intrusion detection, medical diagnostic/monitoring, oil-spill detection, and credit card fraud detection. It is hard to identify rare events correctly if a learning algorithm is just established based on optimal accuracy, as all samples will be classified into the major group. Many algorithms were proposed to deal with class imbalance problems. In this paper, we focus on MICE algorithm proposed by [15] and improve the algorithm by choosing the optimal threshold based on the posterior probabilities. In addition, we illustrate the reason why the logistic transformation works in MICE. The empirical results show that choosing the optimal threshold vis posterior probabilities can improve the performance of the MICE algorithm.