| Concomitant with the rapid evolution of information technology and social networks,Location-based Social Networks(LBSNs)have progressively become an integral component of modern social life.LBSNs enable users to check in at specific geographic locations,share information,and interact,thereby fostering social interaction and information exchange in the physical world.In this context,Points of Interest(POI)recommendation systems have emerged.These systems endeavor to provide personalized POI recommendations for users based on their geographic locations,preferences,and social network relationships,thereby fulfilling their demands for high-quality,intriguing,and practical information.Existing POI recommendation systems primarily rely on user check-in data,location data,and social relationships among users.Although this information helps capture users’ interests and preferences to some extent,it lacks comprehensive consideration of social contextual information.Another challenge lies in users’ interests and needs changing over time and with environmental fluctuations,resulting in recommendation systems’ inadequacies in capturing users’ dynamic interests.This study investigates POI recommendation methods integrating contextual information and deep learning approaches to address these limitations.(1)To address the insufficient consideration of contextual information,specifically social relationships in POI recommendations,we propose a POI recommendation algorithm based on Gated Recurrent Unit(GRU)that integrates social relationship strength and geographic location information—SG-GRU.First,we construct a weighted heterogeneous graph between users and POIs and devise a new edge weight calculation method based on user social relationship strength,check-in sequences,and geographic location information to optimize the heterogeneous graph.Through graph embedding,we obtain formal representations of users and POIs,subsequently utilizing gated recurrent neural networks to establish user sequence interest preference models.Finally,we recommend the top-K POI list to users based on each user’s final POI score.Experimental validation on two public datasets,Gowalla and Brightkite,demonstrates that our proposed SG-GRU method outperforms the five given POI benchmark recommendation models.Compared to the best-performing benchmark model GEAPR,the F1-score@10 on both datasets is increased by 5.2% and 7.2%,respectively,validating the effectiveness of the recommendation model.(2)Addressing the issue of insufficient consideration of the temporal interval between user check-in records and spatial distance between POIs,resulting in lower recommendation accuracy in existing recommendation algorithms based on unidirectional neural network structures,we propose a POI recommendation model integrating user spatiotemporal preferences based on bidirectional self-attention mechanism(Point-of-Interest Recommendation based on Bidirectional Self-Attention Mechanism with Spatio-Temporal Preference,BSA-ST-Rec).First,we extract features such as check-in sequences,temporal intervals,and spatial intervals based on the chronological order of user check-ins.Next,we embed POI feature information to represent sequence embedding and spatiotemporal fusion embedding.Subsequently,we integrate these feature information with bidirectional self-attention mechanism as improved information sources,enabling better understanding of users’ dynamic interest preferences.Ultimately,we predict users’ next POI through this approach,thereby enhancing POI recommendation performance.Corresponding experiments were conducted on public datasets Foursquare and Gowalla,and benchmark methods were compared.Experimental results reveal that our proposed method significantly improves POI recommendation accuracy,further substantiating its effectiveness. |
