A graph neural network incorporating spatio-temporal information for location recommendation

Location recommendation is at the core of location-based service, while recommendation based on graph neural networks (GNNs) has recently flourished, and for location recommendation tasks, GNN-based approaches are equally applicable. To provide fair location recommendation services for multi-users,...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:World wide web (Bussum) 2023-09, Vol.26 (5), p.3633-3654
Hauptverfasser: Chen, Yunliang, Huang, Guoquan, Wang, Yuewei, Huang, Xiaohui, Min, Geyong
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Location recommendation is at the core of location-based service, while recommendation based on graph neural networks (GNNs) has recently flourished, and for location recommendation tasks, GNN-based approaches are equally applicable. To provide fair location recommendation services for multi-users, correlation information between non-adjacent locations and non-consecutive visits is essential information in understanding user behavior. The key to GNN-based location recommendation is how to use GNNs to learn embedding representations for users and locations according to their neighbors. Existing approaches usually focus on how to aggregate information from the perspective of spatial structural information, but temporal information about neighboring nodes in the graph has not been fully exploited. In this paper, a GNN location recommendation model, STAGNN, is proposed to incorporate spatio-temporal information to support fairness-driven location-based service. STAGNN facilitates the progression from spatial to spatio-temporal by generating spatio-temporal embeddings from the perspective of spatial structural information and temporal information. STAGNN also explicitly uses spatio-temporal information of all check-ins through an extended attention layer, an improvement that incorporates non-adjacent locations and non-consecutive visits between point-to-point interactions into the learning of user/location embedding representations with significant spatio-temporal effects. STAGNN also employs a multi-head attention mechanism. Experimental results demonstrate that STAGNN brings a good improvement in GNN-based location recommendation, outperforming the optimal baseline by 6%-11% on the three datasets under the HR@20 evaluation metric.
ISSN:1386-145X
1573-1413
DOI:10.1007/s11280-023-01193-9