Improving kNN classification under Unbalanced Data. A New Geometric Oversampling Approach
A. M. Carvalho, Ronaldo Cristiano Prati · 2018
Training classifiers with unbalanced data is one of the main challenges in the field of Machine Learning. Some techniques that try to get around this problem have been proposed, where one of most important is SMOTE, which artificially generates new instances by interpolating pairs of original instances. This paper proposes a new approach to the balancing of classifier training data. It is a geometric and spatial approach, that uses tetrahedralization by Delaunay tessellation to generate new artificial instances. This allow us to go beyond single pair interpolation. Applying this new method, we can notice an improvement in the classification quality (in terms of AUC) of the kNN classification algorithm when compared with SMOTE.