RC-TL: Reinforcement Convolutional Transfer Learning for Large-scale Trajectory Prediction

Anticipating future locations of mobile users plays a pivotal role in intelligent services supporting mobile networks. Predicting user trajectories is a crucial task not only from the perspective of facilitating smart cities but also of significant importance in network management, such as handover optimization, service migration, and the caching of services in a mobile and edge-computing network. Convolutional Neural Networks (CNNs) have proven to be successful to tackle the forecasting of mobile users’ future locations. However, designing effective CNN architectures is challenging due to their large hyper-parameter space. Reinforcement Learning (RL)-based Neural Architecture Search (NAS) mechanisms have been proposed to optimize the neural network design process, but they are computationally expensive and they have not been used to predict user mobility. In large urban scenarios, the rate at which mobility information is generated makes it a challenge to optimize, train, and maintain prediction models for individual users. However, considering that user trajectories are not independent, a common trajectory-prediction model can be built and shared among a set of users characterized by similar mobility features. In the present work, we introduce Reinforcement Convolutional Transfer Learning (RC-TL), a CNN-based trajectory-prediction system that clusters users with similar trajectories, dedicates a single RL agent per cluster to optimize a CNN neural architecture, trains one model per cluster using the data of a small user subset, and transfers it to the other users in the cluster. Experimental results on a large-scale dataset show that our proposed RL-based CNN achieves up to 12% higher trajectory-prediction accuracy, with no training speed reduction, over other state-of-the-art approaches on a large-scale, real-world mobility dataset. Moreover, RC-TL’s clustering strategy saves up to 90% of the computational resources needed for training compared to single-user models, in exchange for a 3% accuracy reduction