Research map / Video / Video segmentation and tracking
Scalable Trajectory-User Linking with Dual-Stream Representation Networks
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 ScaleTUL, a dual-stream (BiLSTM short-term + structured state space long-term) trajectory encoder trained with spatio-temporal augmentation and supervised contrastive learning, then a cosine-similarity TUL layer, for large-scale trajectory-user linking.
Problem
Trajectory-user linking (matching anonymous trajectories to users) is typically formulated as multi-class classification that does not scale to thousands of users; existing deep models are complex and memory-intensive.
Approach
Combine temporal (multi-day) and random location-masking augmentations with supervised contrastive learning using user labels, a dual-stream encoder (BiLSTM for short trajectories, structured state space model for long ones), and a cosine-similarity linking layer against precomputed user representations instead of a fixed classifier head.
Key results
- On TKY with 400 users, ScaleTUL reaches 65.13% Acc@1 vs. 63.53% for the best baseline S2TUL-HRSTS.
- On the US dataset with 5000 users, ScaleTUL achieves 54.94% Acc@1 vs. 51.40% for MainTUL; S2TUL variants run out of memory at this scale.
- Paper reports about 96% speed improvement over S2TUL-HRSTS on TKY and 10M fewer trainable parameters than MainTUL at 10k users.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ScaleTUL instead of S²TUL-HRSTS | TUL on TKY (400 users) · TKY (400 users) | Acc@1: 63.53% -> 65.13% | Helped |
| Use ScaleTUL instead of MainTUL | TUL on TKY (800 users) · TKY (800 users) | Acc@1: 57.02% -> 60.63% | Helped |
| Use ScaleTUL instead of S²TUL-R | TUL on NYC (400 users) · NYC (400 users) | Acc@1: 57.00% -> 59.43% | Helped |
| Use ScaleTUL instead of MainTUL | TUL on NYC (800 users) · NYC (800 users) | Acc@1: 52.43% -> 56.11% | Helped |
| Use ScaleTUL instead of MainTUL | TUL on Santiago (400 users) · Santiago (400 users) | Acc@1: 55.29% -> 58.44% | Helped |
| Use ScaleTUL instead of MainTUL | TUL on Santiago (800 users) · Santiago (800 users) | Acc@1: 51.82% -> 53.55% | Helped |
| Use ScaleTUL instead of MainTUL | TUL on US (5000 users) · US (5000 users) | Acc@1: 51.40% -> 54.94% | Helped |
| Use ScaleTUL instead of MainTUL | TUL on US (10000 users) · US (10000 users) | Acc@1: 43.85% -> 44.90% | Helped |
| Add temporal augmentation (ablation) | ScaleTUL · NYC and TKY (400 users) | model performance | Helped |
| Add random spatial masking augmentation (ablation) | ScaleTUL · NYC and TKY (400 users) | model performance | Helped |
| Add the full trajectory augmentation strategy (ablation) | ScaleTUL · NYC and TKY (400 users) | model performance | Helped |
| Add the long-term encoder (ablation) | ScaleTUL · NYC and TKY (400 users) | model performance | Helped |
Surprising findings
- Baseline S2TUL variants run out of memory on the 800-user TKY/Santiago and all US settings, illustrating scalability limits of graph-based methods.
Limitations
- Evaluated only on Foursquare check-in data, which is sparse and venue-level rather than GPS trajectories.
- Ablation and parameter claims are shown mainly on NYC and TKY with 400 users, limiting evidence at large scale.
- Stated comparisons rely on baselines tuned by the authors; statistical significance details are limited.
Experimental setup
Datasets: Foursquare check-in TKY Foursquare check-in NYC Foursquare check-in Santiago Foursquare check-in US
Benchmarks: Foursquare TKY Foursquare NYC Foursquare Santiago Foursquare US
Metrics: Acc@1 Acc@5 Macro-F1 Micro-F1 running time
Baselines: KNNTUL MainTUL S2TUL-R S2TUL-HRSTS
Builds on: supervised contrastive learning InfoNCE structured state space models Bi-LSTM
Closest papers by meaning
- Holistic Semantic Representation for Navigational Trajectory Generation (AAAI 2025 · similarity 0.85)
- UniTR: A Unified Framework for Joint Representation Learning of Trajectories and Road Networks (AAAI 2025 · similarity 0.84)
- GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.84)
- Generative Human Trajectory Recovery via Embedding-Space Conditional Diffusion (ICML 2025 · similarity 0.83)
- Cross-Domain Trajectory Association Based on Hierarchical Spatiotemporal Enhanced Attention Hypergraph (AAAI 2025 · similarity 0.83)
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 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.