Research map / Video / Video segmentation and tracking
Correlation-Attention Masked Temporal Transformer for User Identity Linkage Using Heterogeneous Mobility Data
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 MT-Link, a transformer framework for cross-platform user identity linkage that uses a correlation-attention block to find spatio-temporal co-occurrence points and attention-guided masking to filter noise.
Problem
Cross-platform user identity linkage from sparse, noisy, heterogeneous check-in data, where existing methods model each user's sequence independently and miss spatio-temporal co-occurrence between platforms.
Approach
A temporal transformer encodes each platform's check-in sequence, a bidirectional cross-attention block produces attention maps between platforms, and the lowest-attention tokens are masked (with a learnable mask embedding) so a masked transformer encoder keeps co-occurring points and drops noise before binary classification.
Key results
- On Twitter-Foursquare, MT-Link reaches 68.06% Macro-F1 and 82.43% AUC, versus 62.01% and 78.96% for the strongest baseline CPLink+.
- On Twitter-Facebook, MT-Link reaches 73.82% Macro-F1 and 87.86% AUC, versus 65.44% and 80.67% for CPLink+.
- On ISP-Weibo, MT-Link reaches 79.85% Macro-F1 and 86.89% AUC, versus 66.67% and 80.17% for CPLink.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MT-Link instead of CPLink+ | User identity linkage · Twitter-Foursquare | Macro-F1 (%): 62.01 -> 68.06 | Helped |
| Use MT-Link instead of CPLink+ | User identity linkage · Twitter-Facebook | AUC (%): 80.67 -> 87.86 | Helped |
| Use MT-Link instead of CPLink | User identity linkage · Foursquare-Facebook | Macro-F1 (%): 53.40 -> 62.87 | Helped |
| Use MT-Link instead of CPLink | User identity linkage · ISP-Weibo | AUC (%): 80.17 -> 86.89 | Helped |
| Add the masked transformer encoder (ablation) | MT-Link · Twitter-Foursquare | F1: 62.05 -> 68.06 | Helped |
| Add the masked transformer encoder (ablation) | MT-Link · Twitter-Foursquare | AUC: 64.33 -> 82.43 | Helped |
| Add the masked transformer encoder (ablation) | MT-Link · Twitter-Facebook | F1: 71.63 -> 73.82 | Helped |
| Add the masked transformer encoder (ablation) | MT-Link · Twitter-Facebook | AUC: 83.33 -> 87.86 | Helped |
| Add the correlation attention block instead of random masking (ablation) | MT-Link · Twitter-Foursquare | F1: 56.17 -> 68.06 | Helped |
| Add the correlation attention block instead of random masking (ablation) | MT-Link · Twitter-Facebook | AUC: 82.20 -> 87.86 | Helped |
| Use temporal transformer encoder instead of standard transformer encoder (ablation) | MT-Link · Twitter-Foursquare | AUC: 58.38 -> 82.43 | Helped |
| Use temporal transformer encoder instead of standard transformer encoder (ablation) | MT-Link · Twitter-Facebook | AUC: 84.98 -> 87.86 | Helped |
Surprising findings
- Removing the temporal transformer encoder's temporal encoding drops AUC by about 41% on Twitter-Foursquare, showing that temporal positional encoding matters most for long sequences.
- Performance on Twitter-Foursquare drops sharply as correlation attention layers increase, while Twitter-Facebook stays stable.
Limitations
- Runtime is similar to DPLink variants despite no pre-training, due to the added masking and correlation attention cost.
- Sensitivity results show performance on Twitter-Foursquare falls sharply with more correlation attention layers.
- Sequences longer than 200 or 400 and shorter than 3 were removed during preprocessing, which limits evaluation coverage.
Experimental setup
Datasets: XSiteTraj ISP-Weibo
Benchmarks: Twitter-Foursquare Twitter-Facebook Foursquare-Facebook ISP-Weibo
Metrics: Macro-Precision Macro-Recall Macro-F1 AUC
Baselines: DPLink DPLink-SM CPLink CPLink+
Builds on: transformer DPLink CPLink
Closest papers by meaning
- Cross-Domain Trajectory Association Based on Hierarchical Spatiotemporal Enhanced Attention Hypergraph (AAAI 2025 · similarity 0.85)
- TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond (NeurIPS 2026 · similarity 0.82)
- Scalable Trajectory-User Linking with Dual-Stream Representation Networks (AAAI 2025 · similarity 0.81)
- Hidden Follower Detection: How Is the Gaze-Spacing Pattern Embodied in Frequency Domain? (AAAI 2024 · similarity 0.81)
- Link Prediction in Multilayer Networks via Cross-Network Embedding (AAAI 2024 · similarity 0.81)
- Decoupling User Relationships Guides Information Diffusion Prediction (Student Abstract) (AAAI 2024 · similarity 0.80)
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.