Many binary prediction situations involve imbalanced datasets where the ratio of the minority class over the majority class is very low. This is especially true when dealing with problems looking to use machine learning to better detect fraud, errors or exceptions. In this paper, we address the problem of extreme imbalance, i.e. where the imbalance ratio of majority over minority instances exceeds 500. Given the scarcity of minority examples, oversampling is not sensible due to expensive computational cost. Hence, we explore and expand undersampling approaches. Specifically, we propose a modeling framework (i.e., sequence of modeling steps) that seeks to leverage as much training data as possible. Our results indicate the better trade-off between the false positives and false negatives, which makes it more suitable for real-life application.
展开▼
