HBay: Predicting Human Mobility via Hyperspherical Bayesian Learning

Abstract

Accurate human mobility prediction is an essential but critical task in location-based services. Although existing deep learning solutions such as deep recurrent neural networks have remarkable achievements for this task, The diversity of check-in preferences and the sparsity of trajectory representations still prevent us from effectively capturing the richness of human mobility intentions and patterns. To this end, this study introduces a novel Hyperspherical Bayesian learning approach for mobility prediction problem, i.e., HBay. As a generative model, HBay considers multiple contextual semantics underlying check-ins to maximize human diverse preferences and encodes human trajectories in a latent space to mimic complex mobility patterns. In contrast to traditional generative models, HBay operates the latent variables derived from human trajectories in the hyperspherical space to avoid the concern of posterior collapse. In addition, HBay couples with an attentive layer to capture human long-term check-in preferences. The experimental results conducted on four real-world datasets demonstrate our HBay significantly outperforms the state-of-the-art baselines.