Transfer Learning in Traffic Prediction with Graph Neural Networks

Statistics on urban traffic speed flows are essential for thoughtful city planning. Recently, data-driven traffic prediction methods have become the state-of-the-art for a wide range of traffic forecasting tasks. However, many small cities have a limited amount of traffic data available for building data-driven models due to lack of data collection methods. With the acceleration of urbanization, the need for traffic construction of small and medium-sized cities is imminent. To tackle the above problems, we propose a TransfEr lEarning approach with graPh nEural nEtworks (TEEPEE) for traffic prediction that can forecast the traffic speed in data-scarce areas with massive value data from developed cities. In particular, TEEPEE uses graph clustering to divide the traffic network map into multiple sub-graphs. Graph clustering captures more spatial information in the transfer process. To evaluate the effectiveness of TEEPEE, we conduct experiments on two realworld datasets and compare them with other baseline models. The results demonstrate that TEEPEE is among the best efforts of baseline models. We provide a comprehensive analysis of the experimental results in this work.