Urban Traffic Prediction using Congestion Diffusion Model

Abstract

Traffic prediction is an essential task in reducing traffic congestions and improving transportation. However, this task is challenging due to the complex spatio-temporal dynamics of urban traffic networks which are difficult to model. Previous approaches principally concentrate on modeling the Euclidean correlations among spatially adjacent sensors in a road network. In this paper, we propose a new weight modeling technique for the adjacency matrix using the path distance metric for the graph signals to provide accurate spatial properties according to the connection information of the urban road network. We exploit a diffusion-based traffic prediction method for modeling spatial dependency and capturing the temporal dynamics. The experimental result shows that the recent deep learning techniques with the proposed spatial model are promising solutions to the traffic prediction.