Finding the Proverbial Needle: Improving Minority Class Identification Under Extreme Class Imbalance

Imbalanced learning problems typically consist of data with skewed class distributions, coupled with large misclassification costs for the rare events. For binary classification, logistic regression is a common supervised learning technique chosen to perform this task. Unfortunately, the model performs poorly on classification tasks when class distributions are highly imbalanced. To improve this generalization, we implement a novel instance-level weighting methodology for the minority class in the loss function. We build our method from a recently published, locally weighted log-likelihood objective function, where each of the minority class weights are learned from the data. We improve upon this previous approach by creating a convex and hyperparameter-free loss function that improves generalization performance for datasets exhibiting extreme class imbalance.