Comprehensive Analysis of Different Techniques for Data Augmentation and Proposal of New Variants of BOSME and GAN

In many environments in which detection of minority class instances is critical, the available data intended for training Machine Learning models is poorly distributed. The data imbalance usually produces deterioration of the trained model by generalisation of instances belonging to minority class predicting as majority class instances. To avoid these, different techniques have been adopted in the literature and expand the original database such as Generative Adversarial Networks (GANs) or Bayesian network-based over-sampling method (BOSME). Starting from these two methods, in this work we propose three new variants of data augmentation to address data imbalance issue. We use traffic event data from three different areas of California divided in two subgroups attending their severity. Experiments show that top performance cases where reached after using our variants. The importance of data augmentation techniques as preprocessing tool has been proved as well, as a consequence of performance drop of systems in which original databases with imbalanced data where used.