Different Sampling Techniques in Neural Networks
Open AccessThis study compared several resampling methods for both the minority and majority classes to address the problematic issue of a highly skewed unbalanced data. In the initial experiments, deep learning models outperformed other machine learning approaches, these final experiments all used the same deep learning architecture of 9 layers (1 input, 7 hidden, 1 output). While Experiment 1 used the adam optimizer, 100 epochs, batch size 1000, 75%/25% train-test split, Experiment 2 used the nadam optimizer, 700 epochs, batch size 1000, 64%/16%/20% train-validation-test split. In addition, Experiment 2 found results on the original data produced by the techniques unlike Experiment 1. The purpose of Experiment 1 was to build basic models to initially test all the sampling methods. Experiment 2, which included a validation set, was structured to be a more rigorous examination.These experiments demonstrate that the simple resampling methods, minority resampling and the minority and majority bootstrap, outperform the pre-built methods in python. These results suggest that majority of the prebuilt methods, as well as doing nothing (‘No balancing’) over-train on one of the two classes.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.