This paper introduces a novel Time-Location Aware Recommender System (TLARS) to effectively capture the intricate interplay between temporal, spatial, and user preference factors. TLARS leverages a heterogeneous graph neural network (HGNN) to model complex relationships between users, items, and locations. A dynamic multi-head attention mechanism is incorporated to focus on relevant information dynamically, enabling the model to adapt to evolving user preferences and contextual factors. Extensive experimental evaluations on the Foursquare-TKY dataset demonstrate TLARS’s superior performance over state-of-the-art baselines, achieving 15.2% improvement in accuracy over LBSN2Vec and significant gains in NDCG@10 (0.318) and MRR (0.438). Experiments were conducted using the Foursquare-TKY dataset (∼445,734 check-ins), implemented in PyTorch 2.4.0 and DGL 1.0.0, and trained on an NVIDIA GeForce GPU (24 GB VRAM). The model shows particular strength in handling cold-start scenarios and temporal variations, with notably strong performance during weekend periods (0.762 accuracy). Ablation studies validate the critical role of temporal encoding and multi-head attention mechanisms, with temporal information removal leading to a 10.9 percentage point decrease in accuracy. Future work will concentrate on developing real-time adaptation mechanisms and exploring generalization capabilities across diverse geographical datasets.
Nasralla et al. (Thu,) studied this question.