A multi-channel geometric algebra residual network for traffic data prediction

Abstract

Traffic data prediction offers a significant way to evaluate the future traffic congestion status; many deep learning based approaches have been widely applied in this field. Most current methods only consider short-term traffic data forecasting; however, long-term prediction, which supports the optimized distribution of traffic resources, is not well studied. Besides, multiple traffic parameters enable stronger constraints for the data estimation, but the correlation between them in both spatial and temporal domains has not been efficiently learned. Geometric algebra, as a generalization of linear algebra, provides a framework to encode multidimensional data and analyze the correlation. By combining the advantages of the deep neural network and geometric algebra, a multi-channel geometric algebra residual network (MGAResNet) is proposed to address the problem of long-term traffic data prediction. Traffic data obtained from two urban expressways are employed and experimental results demonstrate that the approach outperforms the state-of-the-art work.