EMP-GAN: Encoder-Decoder Generative Adversarial Network for Mobility Prediction

Ultra-dense cell deployments in Beyond 5G and 6G result in extensive overlapping between cells. This makes current reactive handover mechanism inadequate due to availability of multiple strong signals at a position. Moreover, recently proposed predictive mobility management schemes are also not suitable as they may lead to unnecessary handovers. A predictive path-based mobility management scheme can solve these issues, but forecasting User Equipment (UE) paths with high accuracy is a challenging task. This paper proposes Encoder-Decoder Generative Adversarial Network (EMP-GAN) for forecasting multi-step ahead UE path. EMP-GAN architecture consists of generator and discriminator neural networks, where the generator predicts mobility (next multi-step target sequence) and the discriminator classifies between the predicted target sequence and the ground truth in adversarial learning. Besides adversarial learning, feature matching and fact forcing training methods are employed for fast convergence of GAN and performance improvement. EMP-GAN is evaluated on mobility dataset collected from the wireless network of Pangyo ICT Research Center, Korea, and results show that it outperforms state-of-the-art prediction models. In particular, EMP-GAN achieves 95.55%, 94.70%, 93.50%, and 92.39% accuracies for 3, 5, 7, and 9-step predictions, respectively.