Authors:
Érika G. Assis
;
Mark A. Song
;
Luis E. Zárate
and
Cristiane N. Nobre
Affiliation:
Department of Computing, Pontifical Catholic University of Minas Gerais University, Brazil
Keyword(s):
Congenital Syndrome, Zika, Generative Adversarial Networks, GAN, DCGAN.
Abstract:
Class imbalance is a common health care problem and often affects the performance of machine learning algorithms. Unfortunately, the minority class, generally the one with the most significant interest, has their learning affected to the detriment of the majority class. This article proposes using Deep Convolutional Generative Adversarial Networks (DCGAN) for minority class oversampling, generating synthetic instances. For this, the ’RESP-Microcephaly’ database was used, which records suspected cases of congenital alteration due to Zika virus (ZIKV) infection. The database presents unbalanced data with 2904 and 7606 instances with and without congenital alteration, respectively. To evaluate the performance of DCGAN, we compared this method with an undersampling and an oversampling approach, using SMOTE with three classification algorithms. The use of DCGAN for balancing demonstrates a significant improvement in classification indices, especially about the minority class.