A New Traffic-Mining Approach for Unveiling Typical Global Evolutions of Large-Scale Road Networks

In this paper, we present a new traffic-mining approach for automatic unveiling of typical global evolution of large-scale road networks. Our method uses as input a history of continuous traffic states (typically measured by travel times) of *all* links of the road graph. This historical data concatenated in a link/time matrix is then approximated with a locality-preserving Non-negative Matrix Factorization (NMF) method. The network-level traffic state similarity takes into account the graph topology by systematically combining link-wise comparisons with same measure on adjacent links. Based on the obtained matrix factorization, we project original high-dimensional network-level traffic information into a feature space (that of NMF components) of much lower dimensionality than original data. Importantly, because we use a modified NMF ensuring locality-preserving property (LP-NMF), the proximity of data-points in low-dim projected space correspond to proximity also in original high-dim space. We can therefore apply standard clustering methods easily in low-dim space, and directly deduce from its output pertinent categorization of global network traffic states and dynamics. Experimentations on simulated data with a large realistic network of more than 13000 links have been done, and show that our method allows to easily obtain meaningful partition of the attained global traffic states, and to deduce a categorization of the global daily evolution.