Next Point-of-Interest Recommendation for Cold-start Users with Spatial-temporal Meta-Learning
Shaoqing Zhang, Jianzhong Guo, Chun Liu, Zheng Li, Ruijing Li · 2023
Next point-of-interest (POI) recommendation is to recommend the next POI that the users may want to visit according to their check-in trajectories. When there are few check-ins in the trajectories, it will cause the cold-start problem that traditional next POI recommendation methods can't capture users' preferences accurately and no longer has the ability to forecast the next POI. Based on the few-shot meta-learning method of MAML, this paper proposes a new method named STMeta to recommend the next POI for cold-start users. Following few-shot learning, the main idea of the proposed method is to train a model based on the long trajectories of other users with the aim to learn some transferrable and generalized knowledge which can be reused under the cold-start environment. For this purpose, it constructs a lot of short trajectories with only two POI checkins next to each other from the long trajectories to simulate the cold-start environment, and uses them to train a MLP with the ability of measuring the spatial-temporal transition information between two POI check-ins. This information indicates the probability that the users visit the next POI from current POI. And by computing such transition information between users' current POI and all candidate POIs, the most nearest POI can be recommended to users. Based on the public datasets of Gowalla and Brightkite, our experiments have shown that the proposed method achieves better performance in recommending next POIs for cold-start users when compared with related methods.