MLGAN: Addressing Imbalance in Multilabel Learning Using Generative Adversarial Networks
Aatif Nisar Dar, Reshma Rastogi · 2023
A common problem while training supervised deep learning models is the lack of labeled training data. Often real-life datasets such as multilabel datasets suffer from class Imbalance problem, which is inescapable. The limited minority data may not be sufficient for efficient learning and often can cause the networks to overfit. This paper considers the potential appli-cation of Generative Adversarial Networks to restore balance in imbalanced multilabel datasets. We will generate new data for the minority labels and use multilabel learning algorithms to handle multilabel data. We compare our model with MLSMOTE with local label imbalance to validate the effectiveness of our model. Experiments over six real datasets using five different multilabel learning algorithms and five evaluation measures show that our strategy of resampling the multilabel data constantly outperforms MLSMOTE with local label imbalance. Results indicate that imbalance in multilabel datasets is reduced in a classifier-independent way; that is, the classifier should have a deplorable influence on the effectiveness of the resampling strategy. While generating new samples from GAN architecture, we use representative class samples to represent each class distribution which further reduces the training time.