DynaPosGNN: Dynamic-Positional GNN for Next POI Recommendation
Junbeom Kim, Sihyun Jeong, Goeon Park, Kihoon Cha, Ilhyun Suh, Byungkook Oh · 2021 International Conference on Data Mining Workshops (ICDMW) · 2021
Location-based offline user activity data plays a key role to understand user interest and context in various user-friendly services. Especially, the next Point-Of-Interest (POI) recommendation task is focusing on the user’s next location based on previous user trajectory. Previous studies for the next POI recommendations generally utilized the order of each visit as temporal characteristic and positions as spatial characteristic based on the Recurrent Neural Networks (RNN) mechanism. However, they are just interested in the next location right after a user’s current trajectory not considering specific arrival time or the prediction time. In this paper, we propose Dynamic-Positional Graph Neural Network (DynaPosGNN), a novel next POI recommendation model considering specific arrival time in offline user activities. DynaPosGNN can predict users’ next location by analyzing the correlation between arrival time and two spatial dynamic graphs called ’User-POI graph’ and ’POI-POI graph’. Our experiments conducted on two real-world datasets show that DynaPosGNN outperforms existing next POI recommendation models.