An improved deep sequential model for context-aware POI recommendation
Issued Date
2023-01-01
Resource Type
ISSN
13807501
eISSN
15737721
Scopus ID
2-s2.0-85158926402
Journal Title
Multimedia Tools and Applications
Rights Holder(s)
SCOPUS
Bibliographic Citation
Multimedia Tools and Applications (2023)
Suggested Citation
Thaipisutikul T., Chen Y.N. An improved deep sequential model for context-aware POI recommendation. Multimedia Tools and Applications (2023). doi:10.1007/s11042-023-15540-5 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/82760
Title
An improved deep sequential model for context-aware POI recommendation
Author(s)
Author's Affiliation
Other Contributor(s)
Abstract
Nowadays, Location-based social networks (LSBNs) have been pervasive in our everyday lives since mobile users can share their experiences on point-of-interest (POIs) with friends anytime at hand. Currently, the recurrent neural network methods (RNN) have shown promising results for the next POI prediction task. However, they fail to incorporate important spatial-temporal contextual factors into the model learning process to distill the relevant long-term user preferences to the current user preference. Besides, traditional RNN-based models could only capture the consecutive check-ins relationship, which can obscure the complex motivations behind users' decisions in real-world scenarios. Therefore, in this paper, we propose An Improved Deep Sequential Model for Context-Aware POI Recommendation (SCR) for the next POI recommendation. Specifically, we model the user's short-term preference by a self multi-head attentive aggregation layer to capture the relationship of non-consecutive POIs under complex situations and aggregate all POI representations into a user's fine-grained short-term preference. Also, we model the users' long-term preferences by newly proposed three context-aware layers to discover the relevant past sessions on the current sessions based on temporal and spatial contexts. Finally, both users' short-and-long-term preferences are collaboratively fused for the final next POI prediction in a unified framework. The experimental results on two widely public POI datasets reveal the substantial enhancement of our proposed method over several state-of-the-art baselines from all evaluation metrics. Also, through the case study, we demonstrate the model's capability in delivering the interpretable recommendation results to the LSBN users.