The large volume of data flowing throughout location-based social networks (LBSNs) provides an opportunity for human mobility behavior understanding and prediction. However, data quality issues (e.g., historical check-in POI missing, data sparsity) limit the effectiveness of existing LBSN-oriented studies, e.g., Point-of-Interest (POI) recommendation or prediction. Contrary to previous efforts in next POI recommendation or prediction, we focus on identifying the missing POI which the user has visited at a past specific time and proposed a multi-network Embedding (MNE) method. Specifically, the model jointly captures temporal cyclic effect, user preference and sequence transition influence in a unified way by embedding five relational information graphs into a shared dimensional space from both POI- and category-instance levels. The proposed model also incorporates region-level spatial proximity to explore geographical influence, and derives the ranking score list of candidates for missing POI identification. We conduct extensive experiments to evaluate the performance of our model on two real large-scale datasets, and the experimental results show its superiority over other competitors. Significantly, it also proves that the proposed model can be naturally transferred to general next POI recommendation and prediction tasks with competitive performances.