Handling Class Imbalance in k-Nearest Neighbor Classification by Balancing Prior Probabilities

It is well known that recall rather than precision is the performance measure to optimize in imbalanced classification problems, yet most existing methods that adjust for class imbalance do not particularly address the optimization of recall. Here we propose an elegant and straightforward variation of the k-nearest neighbor classifier to balance imbalanced classification problems internally in a probabilistic interpretation and show how this relates to the optimization of the recall. We evaluate this novel method against popular k-nearest neighbor-based class imbalance handling algorithms and compare them to general oversampling and undersampling techniques. We demonstrate that the performance of the proposed method is on par with SMOTE yet our method is much simpler and outperforms several competitors over a large selection of real-world and synthetic datasets and parameter choices while having the same complexity as the regular k-nearest neighbor classifier.