A Multi-Source Information Attention Network for Short Sequential Location Prediction

Sequential location prediction has been studied extensively for various applications, including mobility prediction, route planning, and location-based advertising. The existing models generally work on sequences containing hundreds or even hundreds of check-in records but do not perform well on short sequences with only a dozen check-in records. To solve this problem, we design a Multi-Source Information Attention Network (MSIAN). In this network, we consider information from various sources, including global spatio-temporal information and local personalized information, while using a time indicator with personal preference - stay-time. Specifically, we construct a global graph to use spatio-temporal information from the global. The location and visit time are then used in conjunction to extract the personalization information when modeling the check-in records. Finally, utilize stay-time with personalized preferences when performing the attention operation to improve the prediction effect. The experimental results prove that it makes sense to consider multi-source information and shows that our model significantly outperforms the baselines.