Short-Term Traffic Prediction with Vicinity Gaussian Process in the Presence of Missing Data

This paper considers the problem of short-term traffic flow prediction in the context of missing data and other measurement errors. These can be caused by many factors due to the complexity of the large scale city road network, such as sensors not being operational and communication failures. The proposed method called vicinity Gaussian Processes provides a flexible framework for dealing with missing data and prediction in vehicular traffic network. First, a weighted directed graph of the network is built up. Next, a dissimilarity matrix is derived that accounts for the selection of training subsets. A suitable cost function to find the best subsets is also defined. Experimental results show that with appropriately selected subsets, the prediction root mean square error of the traffic flow obtained by the vicinity Gaussian Processes method reaches 18.9% average improvement with lower costs, which is with comparison to inappropriately chosen training subsets.