Classification of High Dimensional Imbalanced Dataset via Game Theory-based Generative Adversarial Networks

Game theory uses mathematical models to analyze the methods of cooperation or competition of intelligent beings. Game theory attempts to model the mathematical behavior of strategic interaction among rational decision-makers. The ultimate goal of this knowledge is to find the optimal strategy for the players. One of the newest ideas in the application of Game theory in the field of artificial intelligence and machine learning is Generative Adversarial Networks. GANs consist of two parts, use Game theory and compete with each other, making it possible for unsupervised or semi-supervised learning. In addition to generating data, these networks are also used to identify malicious software and software security, machine translation, and natural language processing, and to build a three-dimensional model of an image. However, GANs have a very long training time due to the high number of epochs and input parameters. In this paper, in order to solve the problem of long training time of these networks in the classification of imbalanced high-dimensional datasets, a solution is presented that first, GAN-based oversampling on minority classes. Then in order to improve the efficiency of the designed GAN, the mentioned network is parallelized and ensemble classification is done. The different scenarios performed on the classification of diabetic retinopathy dataset by the proposed method.The results showed the classification accuracy of 87%, the training time is reduced by 74%, which shows higher accuracy than the latest scientific advances.