Entropy and gravitation based dynamic radius nearest neighbor classification for imbalanced problem

Abstract In imbalanced problems, the asymmetric number of samples in different classes brings great challenges to traditional classifiers, especially to the Nearest Neighbors (NN) classifiers. When NN-based classifier deals with imbalanced problems, the criterion of itself makes the classification result data-dependent, thus biasing towards the majority class. To overcome the drawback in NN-based classifiers, a meta heuristic NN-based algorithm named Gravitational Fixed Radius Nearest Neighbor classifier (GFRNN) is proposed to solve imbalanced problems by drawing on Newton’s law of universal gravitation. However, GFRNN still has three major problems including negligence of the distribution of samples, unreasonable calculation of data mass and improper distance metric. To this end, this paper proposes an Entropy and Gravitation based Dynamic Radius Nearest Neighbor algorithm (EGDRNN). Different from GFRNN, EGDRNN determines the radius in a dynamic and rapid way. EGDRNN uses entropy information to make samples at different locations have different importance. Finally, by utilizing a general L p -norm to calculate the distance between two samples, the classification performance is greatly improved. The experimental result validates that the proposed EGDRNN not only achieves the highest classification accuracy but also takes the lowest time consuming among all comparison algorithms.