Behavior Modeling Network with Inter-layer Attention for Human Mobility Prediction

With the increasing popularity of mobile smart devices and advancements in positioning technology, the prediction of human mobility has emerged as a popular research area. The objective of predicting human mobility is to forecast users' next place choices. This research has wide-ranging applications in various domains, including tourism recommendation, urban planning, and traffic prediction. In recent years, researchers have approached mobility prediction as a sequential prediction task, aiming to predict users' next place by modeling their movement patterns. Furthermore, some researchers have incorporated personalized information and employed attention layers to capture individual preferences. However, these methods often overlook the potential interactions among visit records and fail to fully leverage semantic information from historical trajectories, leading to incomplete modeling of user. To tackle these issues, we propose a Behavior Modeling Network with Inter-layer Attention (BMIA) for comprehensive modeling of user behavior. Specifically, we employ bi-directional GRU and self-attention layer to capture the potential interactions between visit records. We design a user-based fusion layer to leverage three types of semantic information effectively. We model user behaviors using two parallel modules. Additionally, we introduce a filtering mechanism to emphasize significant places, ultimately enabling accurate prediction of the user's next place. We evaluate our method using HR and MRR metrics, and conduct comparative experiments and ablation studies on two public real-world datasets collected from the Foursquare platform in New York and Tokyo. The experimental results demonstrate that our method outperforms four existing baseline methods.