Traffic Data Prediction with Geometric Algebra Convolutional Neural Network

Traffic congestion is a major problem faced by many cities at present. As one of the important indicators to measure traffic conditions, vehicle speed prediction is an important task of Intelligent Transportation System (ITS). Using traffic speed prediction can assist traffic management decision-making. In recent years, Convolutional Neural Network (CNN) has played a great role in the field of traffic prediction. It can well extract the characteristics of traffic data to predict the road conditions. But neurons in ordinary CNN are scalars, which can only represent one-dimensional information and it is difficult to learn the correlation between multidimensional inputs effectively. Geometric algebra is an extension of linear algebra, which can effectively represent the relationship between different dimensional data. By combining CNN and geometric algebra, we propose Geometric Algebraic Convolutional Neural Networks (GACNN). This network can make good use of the high-dimensional information contained in the data to achieve better prediction results. We used three days of time-dependent history as input to GACNN to make predictions for the next day. The data from an elevated highway in Shanghai were then used for training and testing, and the experiment showed that our proposed method had better performance than related methods.