A hybrid graph-sequential framework for fusing long-term and short-term preferences in location-based POI recommendation

One-to-one recommendation systems have played a crucial role in enhancing user experience and facilitating informed decisions in the tourism industry. With the growth of the digital sphere in the travel process, an individualized approach can help achieve higher customer satisfaction, increase interactions, and provide a more sustainable approach to tourism. The current-day recommender systems, however, are constrained to some degree in their ability to model how the long-term preferences of the user relate to the short-term and trip-based intents. The present study will capitalize on this gap and propose a hybrid model, based on Graph Neural Networks (GNNs) and Transformer-based sequential models, that learns and combines adaptively long-term collaborative preferences and short-term sequential behavior. The suggested methodology involves data pre-processing to incorporate both time-based and geographical features, feature selection to identify key user attributes and POI traits, and model training within the framework of a hybrid approach. The empirical evaluation of the proposed framework is conducted using the Gowalla location-based social network (LBSN) dataset, which serves as a proxy benchmark for tourism-related mobility behavior, together with standard recommendation evaluation metrics including NDCG and HR@K. The outcomes indicate that the proposed framework achieves a mean NDCG@5 of 0.7776, substantially outperforming the baseline models LightGCN and SASRec, and delivers the highest Precision@5 and robust cold-start resilience. The hybrid frame also offers more flexibility in dynamic user behavior, becomes more interpretable, and is more scalable than existing practices. The study contributes to location-aware personalized recommendation research by proposing a unified framework capable of modeling both long-term and short-term mobility preferences with potential applicability to tourism recommendation scenarios.