Scholarlink

To accelerate the urbanization, accurate traffic prediction is crucial for reducing traffic congestion and improving road usage efficiency. In reality, newly developed or rapidly expanded urban areas face data scarcity due to their limited sensors. Existing methods face challenges in selecting source nodes and handling data shifts between source and target nodes. To address this issue, we introduce a novel framework called NL-ATD. This framework comprises two key components: a denoising module and an attention transfer module. Selecting source nodes highly similar to the target nodes, the denoising module enhances prediction accuracy by calculating spatio-temporal similarities among those nodes. Meanwhile, through dynamic weight allocation, the Attention Transfer (AT) module optimizes knowledge transfer between source and target nodes and strengthens the model’s generalization capabilities. We evaluated our NL-ATD framework using real-world datasets. The results demonstrate that our method outperforms baseline approaches, showing a 7% improvement in accuracy.

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

NL-ATD: Spatio-Temporal Few-Shot Learning via Attention Transfer and Denoising Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuwei Gu,
  • Weilong Ding,
  • Maoxiang Sun,
  • Qi Yu

摘要

To accelerate the urbanization, accurate traffic prediction is crucial for reducing traffic congestion and improving road usage efficiency. In reality, newly developed or rapidly expanded urban areas face data scarcity due to their limited sensors. Existing methods face challenges in selecting source nodes and handling data shifts between source and target nodes. To address this issue, we introduce a novel framework called NL-ATD. This framework comprises two key components: a denoising module and an attention transfer module. Selecting source nodes highly similar to the target nodes, the denoising module enhances prediction accuracy by calculating spatio-temporal similarities among those nodes. Meanwhile, through dynamic weight allocation, the Attention Transfer (AT) module optimizes knowledge transfer between source and target nodes and strengthens the model’s generalization capabilities. We evaluated our NL-ATD framework using real-world datasets. The results demonstrate that our method outperforms baseline approaches, showing a 7% improvement in accuracy.