An Empirical Study for the Multi-class Imbalance Problem with Neural Networks
The latest research in neural networks demonstrates that the class imbalance problem is a critical factor in the classifiers performance when working with multi-class datasets. This occurs when the number of samples of some classes is much smaller compared to other classes. In this work, four different options to reduce the influence of the class imbalance problem in the neural networks are studied. These options consist of introducing several cost functions in the learning algorithm in order to improve the generalization ability of the networks and speed up the convergence process.
KeywordsMulti-class imbalance backpropagation cost function
- 1.Kotsiantis, S., Pintelas, P.: Mixture of expert agents for handling imbalanced data sets. Annals of Mathematics and Computing & TeleInformatics 1(1), 46–55 (2003)Google Scholar
- 6.Kukar, M., Kononenko, I.: Cost-sensitive learning with neural networks. In: 13th European Conference on Artificial Intelligence, pp. 445–449 (1998)Google Scholar
- 7.Looney, C.: Pattern Recognition Using Neuronal Networks - theory and algorithms for engineers and scientists, 1st edn. Oxford University Press, New York (1997)Google Scholar
- 8.Ding, C., Xiang, S.Q.: From multilayer perceptrons to radial basis function networks: a comparative study. In: IEEE Conference on Cybernetics and Intelligent Systems, vol. 1, pp. 69–74 (2004)Google Scholar
- 10.Alejo, R., García, V., Sotoca, J.M., Mollineda, R.A., Sánchez, J.S.: Improving the performance of the rbf neural networks with imbalanced samples. In: Sandoval, F., Gonzalez Prieto, A., Cabestany, J., Graña, M. (eds.) IWANN 2007. LNCS, vol. 4507, pp. 162–169. Springer, Heidelberg (2007)CrossRefGoogle Scholar