Abstract
Transforming user check-in data into graph structure data is a popular and powerful way to analyze users' behaviors in the field of recommendation. Graph-based deep learning methods such as graph embeddings and graph neural networks have shown promising performance on the task of point-of-interest (POI) recommendation in recent years. Despite effectiveness, existing methods fail to capture deep graph structural information, leading to suboptimal representations. In addition, they lack the ability to learn the influences of both global preference and user preference on check-in behavior. To address these issues, we propose a general framework based on preference-aware graph diffusion, named PGD. We first construct two types of graphs to represent global and user preferences. Then, we apply a graph diffusion process to capture the structural information of the generated graphs, resulting in weighted adjacency matrices. Finally, graph neural network-based backbones learn the representations of users and POIs. A learnable aggregation module is developed to adaptively combine global and user preferences. Extensive experiments on four real-world datasets demonstrate the superiority of PGD on POI recommendation compared with mainstream graph-based deep learning methods.
Introduction
Recommending the next place a user might want to visit—a Point-of-Interest (POI)—is a cornerstone of Location-Based Social Networks (LBSNs) like Yelp or Foursquare. While Graph Neural Networks (GNNs) are adept at modeling these relationships, they face two major hurdles:
- The "Shallow" Problem: Standard GNNs typically only aggregate information from immediate neighbors. Stacking many layers to capture long-range dependencies often leads to "over-smoothing," where node representations become indistinguishable and lose their utility.
- The Preference Gap: Most models struggle to find the optimal balance between global popularity (where the crowd goes) and personalized history (where a specific individual goes).
The PGD Framework
The proposed Preference-aware Graph Diffusion (PGD) framework addresses these challenges through a three-stage process:
- Dual-Graph Construction: The model builds two distinct graphs: a Global Activity Graph, capturing collective movement patterns, and a Personalized Activity Graph, capturing unique user habits. The edges in these graphs are defined by temporal proximity in check-ins.
- Graph Diffusion Operation: To capture deep structural information without over-smoothing, PGD applies a diffusion process (specifically, a version of Personalized PageRank). This transforms sparse adjacency matrices into weighted matrices that reflect long-range reachability and structural similarity between POIs.
- Adaptive Aggregation: Weighted matrices are fed into GNN backbones (such as GCN or GAT). A learnable aggregation module then uses soft-max-normalized scalars to adaptively weight global versus personalized representations for each individual user.
Experimental Evaluation
The framework was evaluated on widely-used real-world LBSN datasets, including Yelp, Foursquare, and Gowalla. Data was pre-processed to remove users and POIs with fewer than 20 interactions to ensure a sufficient signal-to-noise ratio. The model was trained using Bayesian Personalized Ranking (BPR) loss.
Sensitivity Analysis
A key hyperparameter in PGD is the time threshold used during graph construction. As shown in the figures below, the model's performance (Precision and Recall) varies with the time window, peaking at specific intervals that reflect the underlying temporal dynamics of each dataset.
Key Results and Insights
PGD consistently outperformed state-of-the-art baselines like GeoMF, POI2Vec, and STA across all datasets.
- Performance Gains: On Yelp, PGD-GAT achieved a Precision@10 of 0.0509, significantly higher than STA's 0.0439. On Foursquare, it reached a Recall@10 of 0.0798, compared to STA’s 0.0732.
- Diffusion as a Shortcut: Ablation studies confirmed that removing the diffusion process (the "PGD-GCN-RW" variant) led to a noticeable drop in performance, proving that capturing high-order relations is vital for accurate POI prediction.
- Backbone Agnostic: The framework proved effective regardless of whether GCN or GAT was used as the underlying engine, demonstrating its versatility.
Conclusion and Future Work
PGD provides a robust solution for POI recommendation by bypassing the limitations of deep GNNs and adaptively balancing preference types. While powerful, the model's reliance on rich check-in data makes "cold-start" scenarios a remaining challenge. Future research will focus on incorporating semantic relevance (e.g., knowing that two POIs are both "Italian Restaurants") and optimizing the computational efficiency of the diffusion process for massive scales.
Summary updated