Research map / Label-efficient and robust learning / Recognition and long-tail
Taming the Long Tail in Human Mobility Prediction
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 LoTNext, a next-POI prediction framework that handles the long-tail POI distribution via a denoising graph adjustment, a logit and sample-weight long-tail loss, and an auxiliary time-prediction task, outperforming 10 baselines on Gowalla and Foursquare.
Problem
Next POI prediction models are biased toward head (frequently visited) POIs, so they struggle to predict less-visited long-tail POIs in user check-in data.
Approach
Combines attention-based edge pruning on the user-POI graph with a global transition graph, a logit adjustment by POI frequency, and a cosine-similarity-based sample re-weighting loss (LTA) that generalizes focal-style reweighting, plus an auxiliary time-slot regression loss.
Key results
- On Gowalla, LoTNext reaches Acc@1 0.1668 and MRR 0.2591, versus 0.1593 and 0.2505 for the strongest baseline SNPM.
- On Foursquare, LoTNext reaches Acc@1 0.3155 and MRR 0.4469, versus 0.2899 and 0.4278 for SNPM.
- Ablation: removing the long-tailed graph adjustment drops Gowalla Acc@1 from 0.1668 to 0.1617; removing the auxiliary time loss drops Gowalla MRR from 0.2591 to 0.2523.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LoTNext instead of SNPM | Next POI prediction · Gowalla | Acc@1: 0.1593 -> 0.1668 | Helped |
| Use LoTNext instead of SNPM | Next POI prediction · Foursquare | Acc@1: 0.2899 -> 0.3155 | Helped |
| Add Long-Tailed Graph Adjustment (ablation) | LoTNext · Gowalla | Acc@1: 0.1617 -> 0.1668 | Helped |
| Add Long-Tailed Graph Adjustment (ablation) | LoTNext · Foursquare | Acc@1: 0.3020 -> 0.3155 | Helped |
| Add Long-Tailed Loss Adjustment instead of cross-entropy only (ablation) | LoTNext · Gowalla | Acc@1: 0.1544 -> 0.1668 | Helped |
| Add Long-Tailed Loss Adjustment instead of cross-entropy only (ablation) | LoTNext · Foursquare | Acc@1: 0.3014 -> 0.3155 | Helped |
| Add original cross-entropy loss alongside LTA loss (ablation) | LoTNext · Gowalla | Acc@1: 0.1550 -> 0.1668 | Helped |
| Add original cross-entropy loss alongside LTA loss (ablation) | LoTNext · Foursquare | Acc@1: 0.3029 -> 0.3155 | Helped |
| Add auxiliary time prediction task (ablation) | LoTNext · Gowalla | MRR: 0.2523 -> 0.2591 | Helped |
| Add auxiliary time prediction task (ablation) | LoTNext · Foursquare | MRR: 0.4370 -> 0.4469 | Helped |
| Use LoTNext instead of Graph-Flashback | Next POI prediction · Gowalla | inference time (10^-3 seconds): 0.0918 -> 0.257 | Hurt |
Surprising findings
- Using only the long-tail adjusted loss (without cross-entropy) gave lower performance than the full model, the authors attribute this to over-focus on tail data hurting head-POI accuracy.
- LightGCN scores much lower than other graph baselines on Foursquare (Acc@1 0.0540).
Limitations
- Does not model emergent events such as vehicle accidents or extreme weather that cause long-tail mobility.
- Reliance on extensive user trajectory data poses privacy risks.
- No error bars or statistical significance tests reported; results from fixed random seeds.
Experimental setup
Datasets: Gowalla Foursquare
Benchmarks: Gowalla Foursquare
Metrics: Acc@1 Acc@5 Acc@10 MRR
Baselines: ST-RNN DeepMove LBSN2Vec LightGCN LSTPM Flashback STAN GETNext Graph-Flashback SNPM
Builds on: Flashback GETNext Graph-Flashback Focal Loss logit adjustment GCN
Closest papers by meaning
- TAU: Trajectory Data Augmentation with Uncertainty for Next POI Recommendation (AAAI 2024 · similarity 0.85)
- Spatial-Temporal Interplay in Human Mobility: A Hierarchical Reinforcement Learning Approach with Hypergraph Representation (AAAI 2024 · similarity 0.84)
- Learning Time Slot Preferences via Mobility Tree for Next POI Recommendation (AAAI 2024 · similarity 0.84)
- TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond (NeurIPS 2026 · similarity 0.84)
- SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark (NeurIPS 2026 · similarity 0.84)
- Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking (NeurIPS 2024 · 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.