Point-of-interest recommendation based on non-adjacent trajectory interaction model

Abstract

Next Point-of-Interest(POl)recommendation aims to predict users'future behaviors based on their historical trajectories, providing significant value toboth users and service providers. Most models fail to capture users'non-adjacent trajectory features, leading to insufficient modeling of users'long-term preferences. Therefore, this paper proposes a Non-Adjacent Trajectory Interaction(NATI) model. The NATI model first uses a multi-dimensional embedding layer to represent user trajectories, then employs multi-head self-attention to capture non-adjacent spatio-temporal features across different subspaces, updating users'long-term preferences.Finally, matching attention is used to match potential locations and predict users'possible POls. Validation on two public datasets demonstrates that the proposed model outperforms baseline models by 8%-16%.