Research map / Graph learning / Knowledge and temporal graphs
TAU: Trajectory Data Augmentation with Uncertainty for Next POI Recommendation
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 TAU, a graph-based trajectory augmentation method that inserts likely missing intermediate POIs (found by BFS on a global transition graph) and calibrates their uncertainty with confidence-aware evidential deep learning for next POI recommendation.
Problem
Next POI recommendation suffers from incomplete and sparse check-in trajectories, since users do not share every visit, which limits sequence models' use of historical context.
Approach
Use BFS over a global trajectory flow graph to find intermediate POIs between consecutive check-ins, sample them weighted by edge transition probabilities to complete sequences, then calibrate the uncertainty of these pseudo-labeled check-ins with confidence-aware evidential deep learning loss.
Key results
- On Gowalla, TAU reaches Acc@1 0.1604, Acc@10 0.4568 and MRR 0.2571, above Graph-Flashback (0.1495 / 0.4242 / 0.2401).
- On Foursquare, TAU reaches Acc@1 0.2931 and MRR 0.4410, versus 0.2786 and 0.4109 for Graph-Flashback.
- Plugging the augmentation into Flashback raises Gowalla Acc@1 from 0.1158 to 0.1415 and into SASRec from 0.0787 to 0.1041.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TAU instead of Graph-Flashback | Next POI recommendation · Gowalla | Acc@1: 0.1495 -> 0.1604 | Helped |
| use TAU instead of Graph-Flashback | Next POI recommendation · Gowalla | Acc@5: 0.3399 -> 0.3672 | Helped |
| use TAU instead of Graph-Flashback | Next POI recommendation · Gowalla | Acc@10: 0.4242 -> 0.4568 | Helped |
| use TAU instead of Graph-Flashback | Next POI recommendation · Gowalla | MRR: 0.2401 -> 0.2571 | Helped |
| use TAU instead of Graph-Flashback | Next POI recommendation · Foursquare | Acc@1: 0.2786 -> 0.2931 | Helped |
| use TAU instead of Graph-Flashback | Next POI recommendation · Foursquare | Acc@5: 0.5733 -> 0.6174 | Helped |
| use TAU instead of Graph-Flashback | Next POI recommendation · Foursquare | Acc@10: 0.6501 -> 0.6992 | Helped |
| use TAU instead of Graph-Flashback | Next POI recommendation · Foursquare | MRR: 0.4109 -> 0.4410 | Helped |
| add Uncertainty Calibration Module (ablation) | TAU · Gowalla | Acc@1: 0.1487 -> 0.1604 | Helped |
| add confidence thresholds to uncertainty calibration (ablation) | TAU · Gowalla | Acc@1: 0.1576 -> 0.1604 | Helped |
| add Trajectory Completion Module (ablation) | TAU · Gowalla | Acc@1: 0.1536 -> 0.1604 | Helped |
| use spatial-aware self-attention instead of Graph-Flashback spatial-temporal LSTM (ablation) | TAU · Gowalla | Acc@1: 0.1541 -> 0.1604 | Helped |
Surprising findings
- Random weighted sampling of completed check-ins still improves performance despite noise in the augmented trajectories, per the authors' ablation.
Limitations
- Missing-check-in candidates are sampled from a transition graph built only from the training data, so augmentation relies on past transition statistics.
- Hyperparameters (BFS depth m, threshold gamma) are tuned on Gowalla and sensitivity is only reported there.
- Gains rely on a fixed BFS depth and random sampling, and the paper notes sampling noise remains.
Experimental setup
Datasets: Gowalla Foursquare
Benchmarks: Gowalla Foursquare
Metrics: Acc@1 Acc@5 Acc@10 MRR
Baselines: PRME STRNN DeepMove STGN SASRec LSTPM Flashback STAN GETNext Graph-Flashback
Builds on: Graph-Flashback Flashback evidential deep learning Transformer (self-attention)
Closest papers by meaning
- GeoMamba: Towards Multi-granular POI Recommendation with Geographical State Space Model (AAAI 2025 · similarity 0.87)
- Integrating Personalized Spatio-Temporal Clustering for Next POI Recommendation (AAAI 2025 · similarity 0.86)
- Learning Time Slot Preferences via Mobility Tree for Next POI Recommendation (AAAI 2024 · similarity 0.86)
- Taming the Long Tail in Human Mobility Prediction (NeurIPS 2024 · similarity 0.85)
- Cross-Domain Trajectory Association Based on Hierarchical Spatiotemporal Enhanced Attention Hypergraph (AAAI 2025 · similarity 0.84)
- Enhancing Long-and Short-Term Representations for Next POI Recommendations via Frequency and Hierarchical Contrastive Learning (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.