Exploring of STGNN for Traffic Forecasting at Expanding Traffic Network

Traffic forecasting is a spatio-temporal forecasting task and plays an important role in traffic management and transportation resource allocation. In this field, many spatio-temporal graph neural network (STGNN) models have been proposed and have achieved high prediction accuracy. However, when the traffic network expands, these models need to be re-trained to optimize for the new traffic network, but the problem of insufficient data arises. We therefore address this problem by using transfer learning between different time periods within the same city. This solves the training data shortage by using knowledge derived from data before the expansion of the traffic network. Specifically, we utilize a simple transfer learning method known as fine-tuning, along with the STGNN model that can store information about a city and consider long-term time series. For evaluation, we perform experiments on traffic data with expanding the traffic network. Experimental results show that our model can effectively address the problem.