Optimizing hyperparameters in the k-NN classifier to maximize F_β score

Classification methods typically aim at achieving desirable performance in multiple accuracy measures. When correct classification of certain class is more important or misclassifying a particular class is more costly, the precision or recall measure would be more relevant. The F_β score combines precision and recall into an aggregate accuracy measure by a weight (β) indicating their relative importance and is appropriate for imbalanced datasets. The classical k-NN algorithm requires hyperparameters including a set of variables (or features), neighbourhood size (k) and a distance metric as input. This study aims to maximize the F_β score by optimizing the choice of variables and k in the k-NN classifier for binary classification. The problem is formulated by a mixed integer linear fractional program (MILFP) and solved by parametric algorithm. Results are compared with the ensemble approach k-NN, a prominent k-NN variant. The Hassanat distance is adopted in both classifiers due to its better performance on disease prediction datasets than other k-NN variants in recent studies. Their performances of the F_β scores are tested on six publicly available datasets. © 2025 Author(s).