Scholarlink

<p>Traffic forecasting poses critical challenges in artificial intelligence when deployed across heterogeneous urban networks with limited data. We propose AST-GNN, an Adaptive Spatio-Temporal Graph Neural Network for transferable traffic forecasting that addresses three key limitations of existing methods: reliance on static graph structures, insufficient handling of cross-domain distribution shifts, and instability under low-data regimes. AST-GNN integrates three components: (1) an adaptive graph learning module that dynamically refines spatial dependencies without predefined topology, (2) a domain alignment mechanism using maximum mean discrepancy to reduce feature distribution shifts between source and target networks, and (3) a multi-stage transfer strategy that progressively fine-tunes parameters to prevent overfitting. Experiments on METR-LA and PeMSD7 benchmarks demonstrate that AST-GNN achieves superior accuracy and stable adaptation in cross-network scenarios, particularly with only 5–25% target data. Compared to baseline methods including Graph WaveNet and TL-GPSTGN, our approach consistently reduces prediction errors by 12–18% in transfer settings. This work advances spatio-temporal graph learning for data-scarce networked systems, offering insights into adaptive AI frameworks for urban computing and beyond.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive Spatio-Temporal Graph Neural Networks for Transferable Traffic Forecasting: Insights into AI-Driven Computing and Informatics

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Jiandong Liu,
  • Lina Qi

摘要

Traffic forecasting poses critical challenges in artificial intelligence when deployed across heterogeneous urban networks with limited data. We propose AST-GNN, an Adaptive Spatio-Temporal Graph Neural Network for transferable traffic forecasting that addresses three key limitations of existing methods: reliance on static graph structures, insufficient handling of cross-domain distribution shifts, and instability under low-data regimes. AST-GNN integrates three components: (1) an adaptive graph learning module that dynamically refines spatial dependencies without predefined topology, (2) a domain alignment mechanism using maximum mean discrepancy to reduce feature distribution shifts between source and target networks, and (3) a multi-stage transfer strategy that progressively fine-tunes parameters to prevent overfitting. Experiments on METR-LA and PeMSD7 benchmarks demonstrate that AST-GNN achieves superior accuracy and stable adaptation in cross-network scenarios, particularly with only 5–25% target data. Compared to baseline methods including Graph WaveNet and TL-GPSTGN, our approach consistently reduces prediction errors by 12–18% in transfer settings. This work advances spatio-temporal graph learning for data-scarce networked systems, offering insights into adaptive AI frameworks for urban computing and beyond.