Scholarlink

<p>The precondition for a deep prediction model to achieve high prediction accuracy is sufficient data, which is not always available in practice. Thus, urban transfer learning methods (i.e., fine-tuning-based methods) are successively proposed to mitigate this issue. However, these existing approaches do not estimate source knowledge transferability and therefore easily lead to negative transfer. Spatial homogeneity (i.e., F1 score gained by link prediction) can provide fine-grained topological structure indication for source knowledge transfer. To this end, we propose a spatial homogeneity-aware transfer learning framework named SHTL for urban flow prediction. In particular, SHTL consists of a link prediction model and an urban flow prediction model. Firstly, the link prediction model is used to capture regional road network topology and regional spatial homogeneity is obtained by evaluating model predictability in each region. Secondly, the urban flow prediction model is optimized by selective source training and target fine-tuning based on spatial homogeneity. We evaluate SHTL on real-world taxi and bike datasets and the result shows that SHTL outperforms the state-of-the-art baselines.</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

Spatial homogeneity-aware transfer learning for urban flow prediction

  • 期刊论文

  • 发表日期：

    2025年2月24日

查看全文
  • Yinghui Liu,
  • Guojiang Shen,
  • Yanjie Fu,
  • Zehui Feng,
  • Zhenzhen Zhao,
  • Xiangjie Kong

摘要

The precondition for a deep prediction model to achieve high prediction accuracy is sufficient data, which is not always available in practice. Thus, urban transfer learning methods (i.e., fine-tuning-based methods) are successively proposed to mitigate this issue. However, these existing approaches do not estimate source knowledge transferability and therefore easily lead to negative transfer. Spatial homogeneity (i.e., F1 score gained by link prediction) can provide fine-grained topological structure indication for source knowledge transfer. To this end, we propose a spatial homogeneity-aware transfer learning framework named SHTL for urban flow prediction. In particular, SHTL consists of a link prediction model and an urban flow prediction model. Firstly, the link prediction model is used to capture regional road network topology and regional spatial homogeneity is obtained by evaluating model predictability in each region. Secondly, the urban flow prediction model is optimized by selective source training and target fine-tuning based on spatial homogeneity. We evaluate SHTL on real-world taxi and bike datasets and the result shows that SHTL outperforms the state-of-the-art baselines.