A Clustering-based Multi-Task Learning Method using Graph Attention Network for Short-term Traffic Forecasting
HongYuan Tan, Pan He, Xiaoyong Sun, Yongting Zhao · 2024
In modern intelligent transportation systems, accurate short-term traffic forecasting is pivotal for managers and travelers. While recent spatio-temporal traffic forecasting models excel in short-term predictions, data from different traffic node inherently possesses diverse characteristics, yet existing models tend to neglect the disparities between these nodes and treat neighboring nodes with different data in a uniform manner, overlooking their unique attributes, which can result in imprecise prediction results. To address these concerns, this paper introduces, the Clustering-based Multi-Task Learning Method using Graph Attention Network (CMTGAT). This model adopts a multi-task learning framework, enhanced with time-series clustering for nuanced pattern recognition and improved prediction accuracy in diverse traffic scenarios. Experimental results demonstrate that the CMTGAT model outperforms existing methods, almost doubling in forecasting precision. The integration of the time series clustering and the graph attention mechanism helps to reduce the MAPE approximately by 6%.