Spatial-Temporal Interval Aware Individual Future Trajectory Prediction

Yiheng Jiang, Yongjian Yang, Yuanbo Xu, En Wang

Abstract: In this work, we present STiSAN$^+$ as an end-to-end mobility trajectory prediction framework. It is a Self-Attention Network (SAN)-based model augmented with two lightweight approaches: Rotary Time Aware Position Encoder (RoTAPE) and multi-head Interval Aware Attention Block (IAAB). These methods allow for the explicit, efficient and effective processing of spatial-temporal intervals among the user's historical trajectory, without the need for additional parameters or a substantial computational burden. On the one hand, RoTAPE simultaneously encodes day- and hour-level timestamps into the sequence representation via a sinusoidal encoding matrix. Notably, the multi-level temporal differences operate in a mutually independent manner to reflect the periodical pattern, and collectively measure the absolute time interval. On the other hand, IAAB, point-wise injecting the historical spatial-temporal intervals into the attention map, can promote SAN attaching importance to the spatial relations under the constraints of time conditions. Moreover, we equip STiSAN$^+$ with a novel MLP-based module, namely Spatial-Temporal Relation Memory (STR Memory). It endows the interactions inside historical intervals along different directions, and converts them into spatial-temporal relations in future trajectories for accurate predictions. The empirical study on six public LBSN shows that from Next Location Recommendation to Multi-location Future Trajectory Prediction, our STiSAN$^+$ gains average 15.05% and 18.35% improvements against several state-of-the-art sequential models, respectively. We demonstrate the effectiveness of proposed module with an ablation study, and validate the extensibility and interpretability of RoTAPE and IAAB through non-sampled metric evaluation and visualization.