An Improved Two-Step Supervised Learning Artificial Neural Network for Imbalanced Dataset Problems

Abstract

An improved two-step supervised learning algorithm of Artificial Neural Networks (ANN) for imbalanced dataset problems is proposed in this paper. Particle swarm optimization (PSO) is utilized as ANN learning mechanism for first step and second step. The fitness function for both steps is Geometric Mean (G-Mean). Firstly, the best weights on network are determined with a decision threshold is set to 0.5. After the first step learning is accomplished, the best weights will be used for second step learning. The best weights with the best value of decision threshold are obtained and can be used to predict an imbalanced dataset. Haberman's Survival datasets, which is available in UCI Machine Learning Repository, is chosen as a case study. G-Mean is chosen as the evaluation method to define the classifier's performance for a case study. Consequently, the proposed approach is able to overcome imbalanced dataset problems with better G-Mean value compared to the previously proposed ANN.