Scholarlink

Accurate short-term traffic forecasting is very important for urban traffic management and planning. In this paper, we propose a new method to apply SSL into STGNN for improving traffic prediction accuracy. We use the GCN-LSTM model to learn complex spatiotemporal dependence in traffic flow data. We further improve the model representation learning through a self-supervised learning framework called SimCLR, hence enabling it to learn more invariant and generalized features in the traffic pattern without huge amounts of labeled data. Therefore, our GCN-LSTM model outperforms the traditional ones concerning the precision of the forecast reflected by both the RMSE and MAE scores. It also indicates that this model does more robustly to variation in traffic and captures the intricate traffic dynamics better. Specially, our model achieved a test loss of 135.01, test RMSE of 11.61, test MAE of 4.73. Extensive experiments on real traffic datasets prove the effectiveness of our approach and also provide an effective scalable solution for real modern traffic management.

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

Self-Supervised Learning for Short-Term Traffic Prediction Using Spatiotemporal Graph Neural Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Idriss Moumen,
  • Jaafar Abouchabaka,
  • Najat Rafalia

摘要

Accurate short-term traffic forecasting is very important for urban traffic management and planning. In this paper, we propose a new method to apply SSL into STGNN for improving traffic prediction accuracy. We use the GCN-LSTM model to learn complex spatiotemporal dependence in traffic flow data. We further improve the model representation learning through a self-supervised learning framework called SimCLR, hence enabling it to learn more invariant and generalized features in the traffic pattern without huge amounts of labeled data. Therefore, our GCN-LSTM model outperforms the traditional ones concerning the precision of the forecast reflected by both the RMSE and MAE scores. It also indicates that this model does more robustly to variation in traffic and captures the intricate traffic dynamics better. Specially, our model achieved a test loss of 135.01, test RMSE of 11.61, test MAE of 4.73. Extensive experiments on real traffic datasets prove the effectiveness of our approach and also provide an effective scalable solution for real modern traffic management.