A Bi-directional Category-Aware Multi-task Learning Framework for Missing Check-in POI Identification.

The prevalence of Location-based Social Networks (LBSNs) services makes next personalized Point-of-Interest (POI) prediction become a research topic. However, due to device failure or intention camouflage, geolocation information missing prevents existing POI-oriented studies for advanced user preference analysis. Herein, we proposed a Bi-directional category-aware multi-task learning (Bi-CatMTL) framework, which fuses bi-direction spatiotemporal transition patterns and personalized dynamic preference to identify where the user has been at a past specific time, namely missing POI identification. Specifically, Bi-CatMTL introduces: (1) a two-channel encoder, i.e., spatial-aware POI encoder and temporal-aware category encoder, to capture user bidirectional dual-grained mobility transition patterns; (2) a task-oriented decoder to fuse learned transition patterns and personalized preference for multi-task prediction; (3) a POI2Cat matrix to make full use of both types of sequential dependencies. Extensive experiments demonstrate the superiority of our model, and it can also be adaptively extended to next POI prediction task with the convincing performance.