Deep Over-sampling Framework for Classifying Imbalanced Data

摘要

Class imbalance is a challenging issue in practical classification problemsfor deep learning models as well as traditional models. Traditionallysuccessful countermeasures such as synthetic over-sampling have had limitedsuccess with complex, structured data handled by deep learning models. In thispaper, we propose Deep Over-sampling (DOS), a framework for extending thesynthetic over-sampling method to exploit the deep feature space acquired by aconvolutional neural network (CNN). Its key feature is an explicit, supervisedrepresentation learning, for which the training data presents each raw inputsample with a synthetic embedding target in the deep feature space, which issampled from the linear subspace of in-class neighbors. We implement aniterative process of training the CNN and updating the targets, which inducessmaller in-class variance among the embeddings, to increase the discriminativepower of the deep representation. We present an empirical study using publicbenchmarks, which shows that the DOS framework not only counteracts classimbalance better than the existing method, but also improves the performance ofthe CNN in the standard, balanced settings.