Redefining POI Popularity: Integrating User Preferences and Recency for Enhanced Recommendations
Alif Al Hasan, Md. Musfique Anwar, M. Arifur Rahman
TL;DR
The paper redefines POI popularity by jointly considering recent activity and the number of unique users, addressing biases from highly active individuals. It proposes a graph-based transformer model GETNext that uses a trajectory flow map, POI embeddings, and a transition attention map, augmented by contextual embeddings that fuse POI-user and time-category information. A multivariate popularity measure with parameters $\alpha$ and $\beta$ guides recommendations, and a loss with $L_{final} = L_{poi} + 10 \times L_{time} + L_{cat}$ balances objectives. Experiments on FourSquare-NYC show improved top-$k$ accuracy and MRR under various settings, demonstrating that recency-aware popularity yields more timely and diverse recommendations with practical impact for real-world POI systems.
Abstract
The task of point-of-interest (POI) recommendation is to predict users' immediate future movements based on their previous records and present circumstances. Popularity is considered as one of the primary deciding factors for selecting the next place to visit. Existing approaches mainly focused on the number of check-ins to model the popularity of a POI. However, not enough attention is paid to the temporal impact or number of people check-ins for a particular POI. Thus, to prioritize more on recent check-ins, we propose recency-oriented definition of POI's popularity by considering the temporal effect of the POIs, the number of check-ins, as well as the number of users who registered in those check-ins. Our experimental results on real dataset show the efficacy of the proposed approach.