Detecting Dynamic Critical Links within Large Scale Network for Traffic State Prediction

Can we expose the relationship between the physical dynamics of a network and its predictability? To contribute to this point, we propose a dimensionality reduction method for network states prediction based on spatiotemporal data. The method is intended to deal with large scale networks, where only a subset of critical links can be relevant for accurate multidimensional prediction (MIMO) performances. The algorithm is based on Latent Dirichlet Allocation (LDA) to highlight relevant topics in terms of networks dynamics. The feature selection trick relies on the assumption that the most representative links of the most dominant topics are critical links for short term prediction. The method is fully implemented to an original application field: short term road traffic prediction on large scale urban networks based on GPS data. Results highlight significant reductions in dimensionality and execution time, a global improvement of prediction performances as well as a better resilience to non recurrent traffic flow conditions.