A Spatiotemporal Multi-View-Based Learning Method for Short-Term Traffic Forecasting

Short-term traffic forecasting plays an important part in intelligent transportation systems. Spatiotemporal k-nearest neighbor models (ST-KNNs) have been widely adopted for short-term traffic forecasting in which spatiotemporal matrices are constructed to describe traffic conditions. The performance of the models is closely related to the spatial dependencies, the temporal dependencies, and the interaction of spatiotemporal dependencies. However, these models use distance functions and correlation coefficients to identify spatial neighbors and measure the temporal interaction by only considering the temporal closeness of traffic, which result in existing ST-KNNs that cannot fully reflect the essential features of road traffic. This study proposes an improved spatiotemporal k-nearest neighbor model for short-term traffic forecasting by utilizing a multi-view learning algorithm named MVL-STKNN that fully considers the spatiotemporal dependencies of traffic data. First, the spatial neighbors for each road segment are automatically determined using cross-correlation under different temporal dependencies. Three spatiotemporal views are built on the constructed spatiotemporal closeness, periodic, and trend matrices to represent spatially heterogeneous traffic states. Second, a spatiotemporal weighting matrix is introduced into the ST-KNN model to recognize similar traffic patterns in the three spatiotemporal views. Finally, the results of traffic pattern recognition under these three spatiotemporal views are aggregated by using a neural network algorithm to describe the interaction of spatiotemporal dependencies. Extensive experiments were conducted using real vehicular-speed datasets collected on city roads and expressways. In comparison with baseline methods, the results show that the MVL-STKNN model greatly improves short-term traffic forecasting by lowering the mean absolute percentage error between 28.24% and 46.86% for the city road dataset and, between 53.80% and 90.29%, for the expressway dataset. The results suggest that multi-view learning merits further attention for traffic-related data mining under such a dynamic and data-intensive environment, which owes to its comprehensive consideration of spatial correlation and heterogeneity as well as temporal fluctuation and regularity in road traffic.