[Paper Review] RELINE: Point-of-Interest Recommendations using Multiple Network Embeddings

Giannis Christoforidis, Pavlos Kefalas+2 more|arXiv (Cornell University)|Feb 2, 2019

Recommender Systems and Techniques4 citations

TL;DR

RELINE proposes a unified model for Point-of-Interest (POI) recommendations by jointly learning user and POI dynamics through multiple network embeddings. It integrates social, geographical, temporal, and preference dynamics by embedding eight relational graphs into a shared latent space, achieving significant accuracy gains over state-of-the-art methods, especially in cold-start scenarios.

ABSTRACT

The rapid growth of users' involvement in Location-Based Social Networks (LBSNs) has led to the expeditious growth of the data on a global scale. The need of accessing and retrieving relevant information close to users' preferences is an open problem which continuously raises new challenges for recommendation systems. The exploitation of Points-of-Interest (POIs) recommendation by existing models is inadequate due to the sparsity and the cold start problems. To overcome these problems many models were proposed in the literature, but most of them ignore important factors such as: geographical proximity, social influence, or temporal and preference dynamics, which tackle their accuracy while personalize their recommendations. In this work, we investigate these problems and present a unified model that jointly learns users and POI dynamics. Our proposal is termed RELINE (REcommendations with muLtIple Network Embeddings). More specifically, RELINE captures: i) the social, ii) the geographical, iii) the temporal influence, and iv) the users' preference dynamics, by embedding eight relational graphs into one shared latent space. We have evaluated our approach against state-of-the-art methods with three large real-world datasets in terms of accuracy. Additionally, we have examined the effectiveness of our approach against the cold-start problem. Performance evaluation results demonstrate that significant performance improvement is achieved in comparison to existing state-of-the-art methods.

Motivation & Objective

• Address the limitations of existing POI recommendation models that fail to capture sparsity, cold-start issues, and multidimensional user behavior.
• Integrate social influence, geographical proximity, temporal patterns, and evolving user preferences into a single unified recommendation framework.
• Overcome sparsity and cold-start problems by modeling dynamic user and POI behavior through relational graph embeddings.
• Develop a probabilistic, graph-based approach that jointly learns embeddings from multiple heterogeneous networks in a shared latent space.

Proposed method

• RELINE constructs eight relational graphs—user-social, user-geographical, user-temporal, POI-temporal, user-route, POI-route, user-preference, and POI-preference—to model diverse user-POI interactions.
• It applies a probabilistic weighting strategy to balance contributions from each graph, with learnable parameters α, β, γ, and δ for social, geographical, temporal, and preference dynamics, respectively.
• A graph-based learning framework jointly optimizes user and POI embeddings in a shared latent space of dimension d, using stochastic sampling to improve training efficiency.
• The model uses time intervals ΔT to define temporal segments, enabling the construction of time-aware graphs such as User-Time period and POI-Time period.
• Hyperparameters such as dimensionality (d) and number of samples (S) are tuned empirically to maximize accuracy, with d=100–140 and S=100–300 across datasets.
• The model is trained end-to-end using negative sampling and stochastic optimization to minimize prediction loss on user-POI interaction sequences.

Experimental results

Research questions

• RQ1 How can multiple network embeddings be effectively combined to improve POI recommendation accuracy?
• RQ2 To what extent do social, geographical, temporal, and preference dynamics collectively reduce sparsity and cold-start issues in POI recommendation?
• RQ3 What is the optimal balance between different influence factors (social, geographic, temporal, preference) in a unified recommendation model?
• RQ4 How does the model perform on cold-start users and locations compared to state-of-the-art baselines?

Key findings

• RELINE achieves significant performance improvements over state-of-the-art models in terms of accuracy across all three real-world datasets: Foursquare, Weeplaces, and Gowalla.
• The model’s accuracy peaks at dimensionality d=100 for Foursquare and Weeplaces, and d=140 for Gowalla, with no further gains beyond these values.
• Optimal time interval ΔT is 20 for Foursquare, 40 for Weeplaces, and 15 for Gowalla, as larger or smaller intervals reduce accuracy due to data sparsity or overfitting.
• Parameter tuning reveals that social influence dominates in Foursquare, geographical influence in Weeplaces, and preference dynamics in Gowalla, reflecting dataset-specific behavioral patterns.
• The model demonstrates strong robustness to cold-start users and locations, significantly outperforming baselines in these challenging scenarios.
• The integration of all four dynamics (social, geographical, temporal, preference) leads to higher accuracy than models relying on only one or two factors.