Abstract
With the wide adoption of mobile devices and web applications, location-based social networks (LBSNs) offer large-scale individual-level location-related activities and experiences. Next point-of-interest (POI) recommendation is one of the most important tasks in LBSNs, aiming to make personalized recommendations of next suitable locations to users by discovering preferences from users' historical activities. Noticeably, LBSNs have offered unparalleled access to abundant heterogeneous relational information about users and POIs (including user-user social relations, such as families or colleagues; and user-POI visiting relations). Such relational information holds great potential to facilitate the next POI recommendation. However, most existing methods either focus on merely the user-POI visits, or handle different relations based on over-simplified assumptions while neglecting relational heterogeneities. To fill these critical voids, we propose a novel framework, MEMO, which effectively utilizes the heterogeneous relations with a multi-network representation learning module, and explicitly incorporates the inter-temporal user-POI mutual influence with the coupled recurrent neural networks. Extensive experiments on real-world LBSN data validate the superiority of our framework over the state-of-the-art next POI recommendation methods.
Introduction
Predicting a user's next destination—whether it is a local gym, a shopping mall, or a specific restaurant—has become a cornerstone of Location-Based Social Networks (LBSNs). While early recommendation systems relied heavily on simple historical trajectories, modern approaches attempt to leverage the rich social and spatial context provided by mobile data. However, two significant challenges remain largely unaddressed by current state-of-the-art models.
First, there is the issue of relational heterogeneity. Human movement is influenced by diverse social circles; a user might visit a toy store based on a family member's recommendation but choose a professional training center based on a colleague's activity. Most existing models treat all social links as a uniform "friendship" graph, failing to distinguish between these distinct types of influence. Second, traditional models often ignore inter-temporal mutual influence. The relationship between a user and a POI is dynamic: a user’s visit can influence a POI’s reputation or popularity, which in turn affects future visitors. Conversely, a POI's evolving status changes the likelihood of a user returning.
To address these gaps, we introduce MEMO (Multi-Relational Modeling), a framework designed to capture the complex interplay of social context and spatio-temporal dynamics through multi-network representation and coupled recurrent neural networks.
The MEMO Framework
The MEMO framework is built upon two primary modules designed to handle the complexities of LBSN data.
Multi-Network Relation Modeling
To account for the diverse nature of social ties, MEMO utilizes multiple Graph Convolutional Networks (GCNs). Each GCN is dedicated to a specific relation type, such as family relations, colleague relations, and user-POI visiting histories. Rather than forcing these disparate relations into a single latent space immediately—which can lead to information loss—MEMO learns relation-specific representations.
These individual "views" of a user are then aggregated using a self-attention mechanism. This allows the model to dynamically weigh the importance of different social circles depending on the context of the recommendation, acknowledging that a colleague's influence might outweigh a family member's in certain professional or urban settings.
User-POI Mutual Influence Modeling
To capture the evolving feedback loop between users and locations, MEMO employs coupled Recurrent Neural Networks (RNNs). In this architecture, one RNN tracks the state transitions of the user, while a parallel RNN tracks the state transitions of the POI. These networks are mathematically linked: the user’s next state is updated based on the POI they just visited, and the POI’s state is updated based on the specific users who visited it. This ensures that the POI is treated as a dynamic entity with a changing "latent status" rather than a static coordinate.
Spatio-Temporal Integration
Beyond social and mutual influence, MEMO incorporates the fundamental physical constraints of movement. The model explicitly integrates Haversine distances and time intervals between consecutive visits. By using learnable vectors to represent short-term versus long-term transitions, the framework effectively balances immediate sequential patterns with broader historical preferences.
Experimental Evaluation
Datasets and Setup
We evaluated MEMO using large-scale, real-world location data covering approximately one-quarter of the U.S. population in 2019. We focused on three specific geographic subsets: Baltimore (July), DC (July), and DC (August).
Because explicit social media friend lists are often unavailable or noisy, social relations (family and colleagues) were inferred using Infostop, an algorithm that detects home and work locations based on stay-point frequency. This allowed us to construct a multi-relational ground truth for our experiments.
Performance Results
MEMO was compared against several competitive baselines, including Markov chains, LightGCN, and self-attentive models like STAN. MEMO consistently achieved the highest scores across all metrics and datasets:
- Accuracy: On the Baltimore July dataset, MEMO reached a Recall@10 of 0.891, outperforming the strongest baseline, STAN (0.868).
- Ranking Quality: The Mean Reciprocal Rank (MRR@10) also saw consistent gains, reaching 0.446 for MEMO compared to 0.426 for STAN in the Baltimore subset.
- Ablation Studies: To understand the contribution of each component, we tested a version of the model without the relation modeling module (MEMO-NG). This led to the most significant drop in performance, confirming that the multi-relational approach is the primary driver of the framework's success.
Discussion and Future Work
The success of MEMO highlights two key insights for the future of POI recommendation. First, treating social relations as heterogeneous significantly boosts predictive accuracy by providing a "collaborative filtering" effect that helps predict visits to new locations where a user has no prior history. Second, acknowledging the reciprocity between users and locations allows the model to account for the dynamic nature of urban environments.
While the results are robust, there are opportunities for further refinement. The current reliance on inferred social ties (home/work locations) is a powerful proxy but may not capture all influential interactions. Future work could extend the framework to include explicit metadata, such as POI categories (e.g., "Nightlife" vs. "Education") or user demographics, to further enrich the multi-relational embeddings. Additionally, optimizing the computational complexity of the coupled RNNs and multiple GCNs remains a priority for scaling the system to even larger datasets.
Summary updated