Time-aware Graph Flashback Network for Next Location Recommendation

Next Point-of-Interest (POI) recommendation predicts a user’s next likely destination based on historical check-ins, enhancing trip planning and location discovery. Current models, including sequence-based and graph-based approaches, often lack adaptability to temporal variations in relationships and treat graph construction as an isolated pre-training step, limiting their effectiveness. To address these challenges, we propose the Time-aware Graph Flashback Network (TGFN), introducing a Spatial-Temporal Knowledge Graph (STKG) that captures dynamic, time-evolving POI relationships. Our Time-TransH model learns both temporal edge variations and core feature representations, enabling real-time updates through weighted convolutions on location nodes and neighbors. By integrating relationship learning in an end-to-end framework, TGFN ensures accurate node representations. Experiments on real-world datasets show that TGFN significantly outperforms existing methods, achieving higher accuracy across multiple metrics.