Spatio-temporal Disentangled representation learning for mobility prediction

Spatio-temporal (ST) prediction task like mobility forecasting is of great significance to traffic management and public safety. There is an increasing number of works proposed for mobility forecasting problems recently, and they typically focus on better extraction of the features from the spatial and temporal domains. Although prior works show promising results on more accurate predictions, they still suffer in characterising and separating the dynamic and static components, making it difficult to make further improvements. Disentangled representation learning separates the learnt latent representation into independent variables associated with semantic factors. It offers a better separation of the spatial and temporal features, which could improve the performance of mobility forecasting models. In this work, we propose a VAE-based architecture for learning the disentangled representation from real spatio-temporal data for mobility forecasting. Our deep generative model learns a latent representation that (i) separates the temporal dynamics of the data from the spatially varying component and generates effective reconstructions; (ii) is able to achieve state-of-the-art performance across multiple spatio-temporal datasets. Moreover, we investigate the effectiveness of our method by eliminating the non-informative features from the learnt representations, and the results show that models can benefit from this operation.