Spatio-Temporal Variable Selection Based Support Vector Regression for Urban Traffic Flow Prediction

Short-term urban traffic flow prediction remains a difficult yet important problem in the intelligent transportation systems (ITS). Most previous spatio-temporal based urban traffic flow prediction techniques just pay attention to building the relationship between the adjacent upstream and downstream road segments using various models. While in this paper, we take advantage of the spatial and temporal information from all available road segments in the road network to predict the short-term traffic volume accurately. However, the available traffic states can be high-dimensional for high-density or large scale road networks. Therefore, we present a spatio-temporal variable selection based support vector regression (VS-SVR) model fed with the high-dimensional traffic data collected from all available road segments. Our prediction framework can be presented as a two-stage model. In the first stage, we employ the multivariate adaptive regression splines (MARS) model to select a set of predictors most related to the target one from the high-dimensional spatio-temporal variables, and reasonable weights are assigned to the selected predictors. In the second stage, the kernel learning method, support vector regression (SVR), is trained on the weighted variables in the second stage for prediction. In the experiments, we employ the actual traffic volume collected from a subarea of Shanghai, China, every 10 minutes. The experimental results indicate that the proposed spatio-temporal variable selection based support vector regression model can generate preferable results in contrast with the time series based autoregression (AR) method, the separate MARS model, and the SVR model.