Improved GAN with fact forcing for mobility prediction

Mobility prediction based on users movement history is a prerequisite for proactive mobility and several other management operations in wireless networks. Recent Deep Learning (DL) studies for mobility prediction have shown improved accuracies using variety of different models that have high complexity and massive training data requirements. These limitations restrict the viability of mobility prediction driven proactive management of wireless networks. This paper overcomes the mentioned limitations by proposing Improved Generative Adversarial Network with Fact Forcing (iGAN-FF) that enhances the existing improved GAN (iGAN). iGAN-FF architecture consists of generator and discriminator neural networks, where generator predicts mobility (next Point of Attachment (PoA)) and discriminator classifies between the predicted PoA and the ground truth in adversarial learning. The adversarial learning reduces the needed amount of training data and enables use of simplistic models, and novel compounded fact forcing and feature matching method achieves higher accuracy. Moreover, the effects of data representation on performance are evaluated by transforming the Campus Mobility Dataset (CMD) and Operator Mobility Dataset (OMD) to One-hot, Binary, and Embedded vectors representations. The results confirm the merits of iGAN-FF as it outperforms iGAN, GAN-FF, GAN, LSTM based Next PoA (LNP), and Transformer based Next PoA (TNP). In particular, the highest accuracies of 99.57% and 83.3% are achieved with CMD One-hot and OMD Embedded representations, respectively. Notably, iGAN-FF shows robustness under data scarcity and secures & SIM;98% accuracy with only 20% of CMD data.

Automated summary

This paper proposes an Improved Generative Adversarial Network with Fact Forcing (iGAN-FF) for mobility prediction, which reduces the need for training data and achieves higher accuracy through adversarial learning. Experimental results confirm the superiority of the iGAN-FF method.