Research map / Label-efficient and robust learning / Recognition and long-tail
Cross-Domain Trajectory Association Based on Hierarchical Spatiotemporal Enhanced Attention Hypergraph
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes StarNet, a hierarchical spatiotemporal attention hypergraph network with KAN fusion for cross-domain trajectory-user identity linkage on LBSN datasets.
Problem
Linking the same users across different LBSN platforms from mobility trajectories, where RNN-based trajectory encoders miss high-order relations and linear fusion of local and global information is weak; data differ in check-in frequency, precision and distribution.
Approach
Combines a local GCN with dynamic composable multi-head attention over grid-cell trajectory graphs, a global hypergraph network with cross-attention and contrastive normalization linking trajectories and users, and a KAN layer that fuses local and global representations for user classification.
Key results
- On Foursquare-Twitter, StarNet reaches ACC@1 0.6187 versus 0.5485 for AttnTUL (12.80% improvement) and Macro-F1 0.4696 versus 0.4006 (17.22%).
- On Ins-Twitter, StarNet reaches ACC@1 0.7390 and Macro-F1 0.6154, improving ACC@1 by 7.16% over AttnTUL.
- Ablations show removing group, temporal, or semantic features, the attention modules, the contrastive normalization layer, or replacing KAN with MLP each reduces performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use StarNet instead of AttnTUL | Cross-domain trajectory association · Foursquare-Twitter | ACC@1: 12.80% | Helped |
| Use StarNet instead of AttnTUL | Cross-domain trajectory association · Foursquare-Twitter | Macro-F1: 17.22% | Helped |
| Use StarNet instead of AttnTUL | Cross-domain trajectory association · Ins-Twitter | ACC@1: 7.16% | Helped |
| Use StarNet instead of AttnTUL | Cross-domain trajectory association · Ins-Twitter | Macro-F1: 7.68% | Helped |
| Add temporal features to trajectory representations (ablation) | StarNet · Foursquare-Twitter and Ins-Twitter | trajectory association performance | Helped |
| Add group features to trajectory representations (ablation) | StarNet · Foursquare-Twitter and Ins-Twitter | trajectory association performance | Helped |
| Add semantic features to trajectory representations (ablation) | StarNet · Foursquare-Twitter and Ins-Twitter | trajectory association performance | Helped |
| Add dynamic composable multi-head attention to the local graph (ablation) | StarNet · Foursquare-Twitter and Ins-Twitter | trajectory association performance | Helped |
| Add cross-attention to the global graph (ablation) | StarNet · Foursquare-Twitter and Ins-Twitter | ACC@1 | Helped |
| Add the contrastive normalization layer (ablation) | StarNet · Foursquare-Twitter and Ins-Twitter | trajectory association performance | Helped |
| Use KAN instead of MLP to fuse local and global information (ablation) | StarNet · Foursquare-Twitter and Ins-Twitter | trajectory association performance | Helped |
Limitations
- Evaluated only on two cross-domain LBSN dataset pairs.
- Relies on many hand-tuned components (composable attention, contrastive normalization, KAN) with limited analysis of their individual contributions beyond ablations.
Experimental setup
Datasets: Foursquare-Twitter Instagram-Twitter
Benchmarks: Foursquare-Twitter Instagram-Twitter
Metrics: ACC@1 ACC@5 Macro-P Macro-R Macro-F1
Baselines: TULER-GRU TULER-LSTM TULER-BiLSTM T3S DPLink TULAM AttnTUL
Builds on: TULER DPLink AttnTUL Kolmogorov-Arnold Networks
Closest papers by meaning
- Hierarchical Attention Network with Correction for Cross-Domain User Association (AAAI 2026 · similarity 0.91)
- Correlation-Attention Masked Temporal Transformer for User Identity Linkage Using Heterogeneous Mobility Data (AAAI 2025 · similarity 0.85)
- TAU: Trajectory Data Augmentation with Uncertainty for Next POI Recommendation (AAAI 2024 · similarity 0.84)
- Scalable Trajectory-User Linking with Dual-Stream Representation Networks (AAAI 2025 · similarity 0.83)
- Higher-order Relational Reasoning for Pedestrian Trajectory Prediction (CVPR 2024 · similarity 0.83)
- THGNets: Constrained Temporal Hypergraphs and Graph Neural Networks in Hyperbolic Space for Information Diffusion Prediction (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.