Scholarlink

With the increasing complexity of transportation systems, improving traffic flow prediction accuracy is crucial. Traffic data’s spatiotemporal heterogeneity poses challenges to existing models. This paper proposes HeteroMoE, a novel approach combining a Transformer model to capture complex spatiotemporal features and a MoE model to handle data heterogeneity. By integrating the MoE framework with adaptive spatiotemporal feature matrices, HeteroMoE implicitly captures spatiotemporal heterogeneity. A gating mechanism, constructed using future data’s historical information, assigns experts based on data characteristics to enhance learning of spatiotemporal heterogeneity. Experiments on four real-world datasets show that HeteroMoE outperforms traditional methods and has great potential in real-time traffic forecasting.

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

HeteroMoE: Spatio-Temporal Heterogeneity Learning via Gated Mixture-of-Experts Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kaizhen Tang,
  • Zihao Zhong,
  • Hao Wu,
  • Haiyan Ding,
  • Gaosong Lv,
  • Zhili Zeng

摘要

With the increasing complexity of transportation systems, improving traffic flow prediction accuracy is crucial. Traffic data’s spatiotemporal heterogeneity poses challenges to existing models. This paper proposes HeteroMoE, a novel approach combining a Transformer model to capture complex spatiotemporal features and a MoE model to handle data heterogeneity. By integrating the MoE framework with adaptive spatiotemporal feature matrices, HeteroMoE implicitly captures spatiotemporal heterogeneity. A gating mechanism, constructed using future data’s historical information, assigns experts based on data characteristics to enhance learning of spatiotemporal heterogeneity. Experiments on four real-world datasets show that HeteroMoE outperforms traditional methods and has great potential in real-time traffic forecasting.