Research map / Video / Video segmentation and tracking
Spatial-Temporal Interplay in Human Mobility: A Hierarchical Reinforcement Learning Approach with Hypergraph Representation
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 STI-HRL, a two-level hierarchical RL framework with a Mobility Hypergraph state encoder that disentangles spatial and temporal agents for next-POI prediction, outperforming eight baselines on NYC and Tokyo check-in data.
Problem
Next-visit location prediction in human mobility, where the trade-off between spatial and temporal constraints varies across individuals and existing methods emphasize only one dimension.
Approach
A two-tiered hierarchical RL setup: low-level spatial and temporal agents learn separate preferences, and a high-level agent adaptively mixes their policies via a learned weight beta; states come from a multi-channel Mobility Hypergraph embedded with attention-based hypergraph convolution.
Key results
- On NYC, STI-HRL reaches Recall@5 0.4268 and Recall@10 0.5227, improving over the second-best baseline by 5.67% and 3.22% in Recall.
- On TKY, STI-HRL reports F1@10 of 0.1005 with improvements of 13.3% over the second-best baseline, and Recall@5 0.4425 (+7.30%).
- Ablations show removing either the spatial or temporal agent, or the high-level integration, degrades performance, and the hypergraph state outperforms a raw sequence input.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Full STI-HRL (spatial+temporal hierarchical agents with hypergraph state) versus strongest baselines | STI-HRL · New York and Tokyo next-location prediction | Recall, F1, MRR, NDCG: NYC: +5.67% Recall@5 over second-best; Tokyo: +13.10% F1@5 over second-best | Helped |
| Removing the hypergraph and feeding raw user sequences directly (STI-HRL') (ablation) | STI-HRL · New York and Tokyo | unspecified (shown in figure) | Helped |
| Using only POI, time, or zone hyperedges as state instead of all hyperedge types (ablation) | STI-HRL · New York and Tokyo | unspecified (shown in figure) | Helped |
| Removing spatial, temporal, or high-level integration agent (two-layer decoupling-integration) (ablation) | STI-HRL · New York and Tokyo | unspecified (shown in figure) | Helped |
Surprising findings
- On the TKY dataset, the STI-HRL F1@10 reported is 0.1005 while DeepMove gets 0.0887; however, in the NYC table the F1@5 for KNN Bandit (0.1414) is close to STI-HRL, suggesting metric variation across baselines.
Limitations
- Evaluated only on two check-in datasets (NYC and TKY) from a single time period.
- High-level reward weights are learned automatically and cannot be manually adjusted, reducing interpretability control.
- Some reported improvements are close to baselines on certain metrics (e.g., F1 and MRR at larger cutoffs).
Experimental setup
Datasets: NYC check-in dataset TKY check-in dataset
Benchmarks: NYC check-in TKY check-in
Metrics: Recall@5 Recall@10 Recall@20 F1@5 F1@10 F1@20 MRR@5 MRR@10 MRR@20 NDCG@5
Baselines: ST-RNN DeepMove LSTPM HST-LSTM STAN LBSN2vec FPMC KNN Bandit
Builds on: DeepMove policy gradient graph attention
Closest papers by meaning
- RHYTHM: Reasoning with Hierarchical Temporal Tokenization for Human Mobility (NeurIPS 2025 · similarity 0.85)
- Structural Information-based Hierarchical Diffusion for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
- Taming the Long Tail in Human Mobility Prediction (NeurIPS 2024 · similarity 0.84)
- Reconciling Spatial and Temporal Abstractions for Goal Representation (ICLR 2024 · similarity 0.84)
- Homogeneous Dynamics Space for Heterogeneous Humans (CVPR 2025 · similarity 0.84)
- STAIR: Addressing Stage Misalignment through Temporal-Aligned Preference Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
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.