Abstract
M3 addresses the challenge of sustainable transportation and leisure planning by proposing a multi-stakeholder recommendation framework that integrates attention-based graph clustering with Q-Learning. The system models spatial heterogeneity at multiple scales — local, district, and city-level — to balance user preferences, operator objectives, and environmental impact when recommending green attractions.
Published at ACM SIGSPATIAL 2025, the work contributes a novel reinforcement learning formulation that treats spatial recommendation as a sequential decision process over heterogeneous graph structures.
Key Contributions
- Multi-scale spatial heterogeneity encoding using hierarchical graph attention networks (HAT-GNN).
- Q-Learning agent that optimizes multi-stakeholder reward (user satisfaction, operator throughput, eco-score).
- Cluster-based action space reduction for scalable real-world deployment.
- Evaluation on real geospatial datasets from Tainan and Taipei metropolitan areas.
Technologies
BibTeX
@inproceedings{lai2025m3,
title={M3: Recommendation via Attention-Graph Cluster Q-Learning with Multi-Scale Spatial Heterogeneity for Multi-Purpose, Multi-Stakeholder Green Attractions in Transportation},
author={Lai, Shih-Yu and Hsieh, Tzu-Hsin and Tsai, Pei-Chi and Kung, Chao-Chun and Ling, Sing-Kai and Hsieh, Hsun-Ping},
booktitle={Proceedings of the 33rd ACM International Conference on Advances in Geographic Information Systems},
pages={39--51},
year={2025}
}