Imbalanced evolving self-organizing learning

In this paper, a hybrid learning model of imbalanced evolving self-organizing maps (IESOMs) is proposed to address the imbalanced learning problems. In our approach, we propose to modify the classic SOM learning rule to search the winner neuron based on energy function by minimally reducing local error in the competitive learning phase. The advantage of IESOM is that it can improve the classification performance through obtaining useful knowledge from the limited and underrepresented minority class data. The positive and negative SOMs are employed to train the minority and majority class, respectively. Based on the original minority class, the positive SOM evolves into a new stage that might discover novel knowledge. The purpose of convergent evolution is to recurrently search the fitness value via minimal mean quantization error in the feature space, which can motivate the offspring individuals to move toward the center of positive SOM so as to form more explicit boundary. The iterative learning procedure is used to adaptively update the incremental feature maps and create more minority instances to facilitate learning from imbalanced data. The effectiveness of the proposed algorithm is compared with several existing methods under various assessment metrics.