Cross-Domain Trajectory Association Based on Hierarchical Spatiotemporal Enhanced Attention Hypergraph

Chenlong Wu, Ze Wang, Keqing Cen, Yude Bai, Jin HaoAAAI 2025 · DOI

TL;DR

This paper introduces StarNet, a novel framework for linking user identities across different location-based social networks (LBSNs). By employing a hierarchical approach that combines local spatio-temporal graph networks with global hypergraph attention, StarNet effectively captures complex user mobility patterns and cross-domain relationships, leading to improved accuracy in user identity association.

What Problem Does This Paper Address?

Identifying and linking the same user across different online platforms, especially location-based social networks (LBSNs), is crucial for understanding user behavior and providing personalized services. However, this task is challenging due to significant differences in user check-in frequencies, data precision, and distribution across various platforms. Existing methods often struggle to capture the intricate, multi-dimensional relationships within trajectories and the higher-order associations between trajectories from different domains. This paper aims to bridge this gap by proposing a model that can dynamically learn rich trajectory representations and effectively associate them across diverse LBSN datasets.

Methods

StarNet is built upon three key components:

  1. Local Spatio-Temporal Enhanced Graph Neural Network Module: This module focuses on capturing local trajectory characteristics. It uses a spatio-temporal graph to represent grid cells within trajectories and applies graph convolutional networks (GCNs) to learn local embeddings. To enrich these embeddings, it integrates temporal, semantic, and group features using sparse linear embedding layers. A dynamic composable multi-head attention mechanism is then employed to capture long-term dependencies and adaptively fuse these multi-dimensional features.
  2. Global Spatio-Temporal Enhanced Hypergraph Network Module: This component addresses the challenge of modeling complex, higher-order relationships between trajectories and users across different domains. It constructs a global spatio-temporal hypergraph where trajectory nodes and user nodes are connected via hyperedges. A hypergraph neural network is used to process these complex relationships. Crucially, it incorporates a cross-attention mechanism to deeply learn cross-domain trajectory features and a contrastive normalization layer to mitigate information dilution and improve feature uniformity.
  3. Fusion Enhancement Correlation Module: This module integrates the local and global representations to achieve precise user identity linkage. It utilizes Kolmogorov-Arnold Networks (KANs), which are known for their ability to adaptively fuse local and global information, to combine the enhanced trajectory representations. This fusion process generates feature vectors that reflect user identity associations, which are then used for prediction.

Datasets

The paper evaluates StarNet on two real-world cross-domain LBSN datasets:

  • Foursquare-Twitter: A sparse dataset.
  • Instagram-Twitter: A dense dataset.

These datasets are standard benchmarks for cross-domain user identity linkage tasks.

Key Insights and Ideas

  • Hierarchical Representation Learning: The paper highlights the effectiveness of a hierarchical approach, first learning local trajectory features and then modeling global, cross-domain relationships.
  • Hypergraph for Higher-Order Relations: Using hypergraphs is shown to be effective in capturing complex, non-pairwise relationships between trajectories and users, which is crucial for cross-domain association.
  • Dynamic Attention for Feature Fusion: The dynamic composable multi-head attention mechanism allows the model to adaptively weigh different aspects of trajectory data (temporal, semantic, group), leading to richer representations.
  • KANs for Adaptive Fusion: Kolmogorov-Arnold Networks (KANs) prove to be a powerful tool for adaptively fusing local and global trajectory information, outperforming traditional MLP-based fusion.
  • Contrastive Normalization for Robustness: The contrastive normalization layer helps to improve feature uniformity and mitigate dimensional collapse, enhancing model generalization.

Key Results

StarNet significantly outperforms several state-of-the-art baselines on both sparse and dense LBSN cross-domain datasets.

  • On sparse datasets (Foursquare-Twitter), StarNet achieved improvements of 12.80% in ACC@1 and 17.22% in Macro-F1 compared to the baseline AttnTUL.
  • On dense datasets (Instagram-Twitter), StarNet showed an improvement of 7.16% in ACC@1 and 7.68% in Macro-F1 compared to AttnTUL. The ablation studies also confirmed the importance of each proposed component, with the attention mechanisms and KAN layer showing significant contributions to performance.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the model performs well, the inherent sparsity and noise in LBSN data can still pose challenges. The computational complexity of hypergraph networks, though addressed through specific mechanisms, might still be a consideration for extremely large-scale deployments.
  • Future Work: The authors suggest further exploration into more sophisticated attention mechanisms and hypergraph structures. Investigating methods to further enhance the model's robustness against data noise and sparsity, and exploring its applicability to other cross-domain tasks beyond user identity linkage, are potential future directions.

Summary updated

BibTeX
@inproceedings{wu2025cross,
  author = {Chenlong Wu and Ze Wang and Keqing Cen and Yude Bai and Jin Hao},
  title = {{Cross-Domain Trajectory Association Based on Hierarchical Spatiotemporal Enhanced Attention Hypergraph}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i12.33402},
}
Figure