BBW: a batch balance wrapper for training deep neural networks on extremely imbalanced datasets with few minority samples

Abstract In recent years, Deep Neural Networks (DNNs) have achieved excellent performance on many tasks, but it is very difficult to train good models from imbalanced datasets. Creating balanced batches either by majority data down-sampling or minority up-sampling can solve the problem in certain cases. However, may lead learning process instability and overfitting. this paper, we propose Batch Balance Wrapper (BBW), a novel framework which adapt general DNN be well trained extremely datasets with few samples. BBW, two extra network layers are added start of DNN. The prevent overfitting samples improve expressiveness sample distribution Furthermore, (BB), class-based sampling algorithm, proposed make sure each batch always during process. We test BBW three well-known maximum imbalance ratio reaches 1167:1 only 16 positive Compared existing approaches, achieves better classification performance. addition, BBW-wrapped DNNs 16.39 times faster, relative unwrapped DNNs. Moreover, does not require preprocessing additional hyper-parameter tuning, operations that processing time. experiments prove applied common applications samples, such as EEG signals, medical images so on.