Multi-Relational Graph Attention Network for Social Relationship Inference from Human Mobility Data

Guangming Qin, Jianpeng Qi, Bin Wang, Guiyuan Jiang, Yanwei Yu, Junyu DongIJCAI 2024

Introduction

Inferring social ties from location-based social networks (LBSNs) like Foursquare or Yelp is a fundamental task in spatio-temporal data mining. Traditional methods often rely on direct "co-location" events—instances where two individuals check in at the same place at the same time. However, such signals are frequently sparse and partially visible. The Multi-Relational Graph Attention Network (MRGAN) addresses these limitations by explicitly modeling both direct social interactions and indirect mobility preferences, capturing the subtle social signals hidden in sparse check-in data.

Problem Statement: Beyond Co-location

While physical proximity is a strong predictor of friendship, many real-world meetings go unrecorded. Conversely, many check-ins occur at different times, masking potential social connections. Existing models often fail to distinguish between different types of interactions or miss the "homophily" effect: the tendency for individuals with similar interests to visit similar types of venues, even without physical overlap. MRGAN bridges this gap by incorporating indirect relations, such as shared location preferences, into the inference process.

The MRGAN Framework

The MRGAN architecture operates through a hierarchical attention-based approach across four distinct stages:

The MRGAN architecture, illustrating the multi-relational graph construction, influence encoding, and cross-relation aggregation.
  1. Multi-Relational Graph Construction: The model builds a multiplex graph with four layers:
  2. Friendship: Known social ties.
  3. Meeting: Direct spatio-temporal overlaps (within a 2-hour window).
  4. Personalized Preference (PP): Links between users visiting the same Point of Interest (POI) at different times.
  5. Common Preference (CP): Links between users visiting different but semantically related POIs (e.g., a gym and a health food store).
  6. Relation-Specific Influence Encoder: A node-level attention mechanism identifies which neighbors within each layer exert the most social influence on a user's representation.
  7. Cross-Relation Embedding Aggregator: A second level of attention learns to weight the four relation layers, acknowledging that different behaviors (e.g., a single meeting vs. long-term shared interests) carry different weights for different users.
  8. Optimization: The framework is trained using binary cross-entropy loss with negative sampling to distinguish likely friends from random pairs.

Experimental Setup

MRGAN was evaluated on three real-world datasets: Gowalla (Austin), Brightkite (San Francisco/LA), and Foursquare (NYC). Crucially, the authors did not filter out "inactive" users with few check-ins, ensuring the model's performance is representative of the noisy, sparse data found in real-world applications.

Results and Discussion

MRGAN consistently outperformed 14 state-of-the-art baselines, including GCN, HAN, and SRINet. The model achieved an average improvement of 5.51% in ROCAUC and 4.28% in PRAUC across all datasets.

Model performance (PRAUC) across different training set sizes, demonstrating robustness even with limited labels.

Ablation studies confirmed that every component—the meeting layer, preference layers, and hierarchical attention—is vital for peak performance. Furthermore, as shown in the figure above, MRGAN reaches high accuracy with as little as 25% of known friendships in the training set, indicating high label efficiency and resilience to data sparsity.

Conclusion and Future Work

By leveraging indirect mobility relations and hierarchical attention, MRGAN provides a robust solution for social tie inference. Future research may focus on automating the tuning of the time threshold ($\tau$) for different POI types, optimizing the model for million-user scales, and integrating temporal graph learning to capture the evolution of social and mobility patterns over time.

Summary updated

BibTeX
@inproceedings{qin2024multi,
  author = {Guangming Qin and Jianpeng Qi and Bin Wang and Guiyuan Jiang and Yanwei Yu and Junyu Dong},
  title = {{Multi-Relational Graph Attention Network for Social Relationship Inference from Human Mobility Data}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure