Improving human mobility identification with trajectory augmentation

Many location-based social networks (LBSNs) applications such as customized Point-Of-Interest (POI) recommendation, preference-based trip planning, travel time estimation, etc., involve an important task of understanding human trajectory patterns. In particular, identifying and linking trajectories to users who generate them – a problem called Trajectory-User Linking (TUL) – has become a focus of many recent works. TUL is usually studied as a multi-class classification problem and has gained recent attention because: (1) the number of labels/classes (i.e., users) is way larger than the number of motion patterns among various trajectories; and (2) the location-based trajectory data, especially the check-ins – i.e., events of reporting a location at particular Point of Interest (POI) with known semantics – are often extremely sparse. Towards addressing these challenges, we introduce a Trajectory Generative Adversarial Network (TGAN) as an approach to enable learning users motion patterns and location distribution, and to eventually identify human mobility. TGAN consists of two jointly trained neural networks, playing a Minimax game to (iteratively) optimize both components. The first one is the generator, learning trajectory representation by a Recurrent Neural Network (RNN) based model, aiming at fitting the underlying trajectory distribution of a particular individual and generate synthetic trajectories with intrinsic invariance and global coherence. The second one is the discriminator – a Convolutional Neural Network (CNN) based model that discriminates the generated trajectory from the real ones and provides guidance to train the generator model. We demonstrate that the above two models can be well tuned together to improve the TUL performance, while achieving superior accuracy when compared to existing approaches.