An improvement to gravitational fixed radius nearest neighbor for imbalanced problem

Abstract

Mining of imbalanced data is one of the basic challenges in the field of machine learning and data mining. In the recent years, a lot of approaches have been proposed to handle imbalanced learning problem. A group of these methods are algorithmic level methods, which are adapted to the nature of imbalanced datasets. Gravitational fixed radius nearest neighbor algorithm (GFRNN) is an algorithmic level method, proposed in order to improve k nearest neighbor classifier when dealing with imbalanced datasets. This algorithm finds the fixed radius nearest neighbors as candidate set. Then, by computing the sum of gravitational forces on a query instance from candidate set, predict its label. Simplicity and no need for manually parameter setting during the run of algorithm are the main advantages of this method. In this paper, gravitational search algorithm (GSA) is used with the aim of finding the mass of training instances to improve the performance of GFRNN. Also we utilize the all training instances to make a decision about query instance. Experimental result on fifteen datasets show the superiority of it compared with four other algorithms.