A user trajectory simulation framework for next POI recommendation with uncertain check-ins.

Abstract

Next point-of-interest (POI) recommendation faces challenges from uncertain check-ins, especially within collective POIs (CPOIs)-venues like shopping malls containing multiple individual POIs (IPOIs). Due to unobserved user behaviors within CPOIs, existing methods treat CPOIs as proxies for their contained IPOIs, which limits next POI recommendation performance. To address this, we propose TraSim, a user Trajectory Simulation framework to simulate user behaviors within CPOIs based on Monte Carlo Tree Search, to approximate the potential trajectory within CPOIs, thus improving the recommendation accuracy under uncertain check-in scenarios. Specifically, to guide the simulation process, TraSim is equipped with a heuristic reward module that integrates (1) IPOI-level signals-temporal traits, transition patterns, and user personalized preferences across various activities, and (2) CPOI-level semantic constraints that narrow the simulation space based on the diversity of activities within each CPOI. Moreover, we construct a multi-candidate trajectory pool that retains diverse high-reward trajectories, enhancing simulation robustness. Furthermore, TraSim is model-agnostic and hyperparameter-free, allowing it to be a lightweight, plug-and-play module compatible with various baselines for next POI recommendation, thus providing a versatile and scalable solution. Extensive experiments on three real-world datasets (Calgary, Charlotte, and Phoenix) and nine competitive baselines demonstrate TraSim's effectiveness and efficiency, with average improvements of 49.1% in HR and 44.0% in MRR.