An improved dynamic sampling back-propagation algorithm based on mean square error to face the multi-class imbalance problem

作者：Alejo R^*; Monroy de Jesus J; Ambriz Polo J C; Pacheco Sanchez J H

来源：Neural Computing & Applications, 2017, 28(10): 2843-2857.

DOI：10.1007/s00521-017-2938-3

• MLP
• Back-propagation
• Multi-class imbalance problem
• Mean square error
• Re-sampling methods
• Dynamic sampling

In this paper, we present an improved dynamic sampling approach (I-SDSA) for facing the multi-class imbalance problem. I-SDSA is a modification of the back-propagation algorithm, which is focused to make a better use of the training samples for improving the classification performance of the multilayer perceptron (MLP). I-SDSA uses the mean square error and a Gaussian function to identify the best samples to train the neural network. Results shown in this article stand out that I-SDSA makes better exploitation of the training dataset and improves the MLP classification performance. In others words, I-SDSA is a successful technique for dealing with the multi-class imbalance problem. In addition, results presented in this work indicate that the proposed method is very competitive in terms of classification performance with respect to classical over-sampling methods (also, combined with well-known features selection methods) and other dynamic sampling approaches, even in training time and size it is better than the over-sampling methods .

• 出版日期2017-10