Missing POI Check-in Identification Using Generative Adversarial Networks

The missing point-of-interest (POI) check-ins in real-life mobility data prevent advanced analysis of users’ preferences and mobile patterns. Existing approaches for missing POI check-in identification mainly focus on modelling spatio-temporal dependencies and memorising transition patterns through users’ check-in sequences. However, these methods cannot ensure that the generated missing records obey the same distribution as the observed check-ins. To this end, we propose a novel Bi-G $$ ^2 $$ AN model, which fuses the merits of generative adversarial network (GAN) and bi-directional gated recurrent unit (GRU), to identify the missing POI check-ins. Specifically, we develop a GAN-based method to mimic the overall distribution of a given check-in dataset, and it is further utilized to generate more reasonable missing POI check-ins. In order to capture bi-directional dependencies and historical impact, a modified bi-directional GRU is utilized in GAN. Moreover, both spatio-temporal influence and local motion information are employed to learn users’ dynamic preferences. Finally, experiments conducted on three real datasets demonstrate the competitiveness of the Bi-G $$ ^2 $$ AN model, outperforming state-of-the-art approaches.