TNeT: Tensor-Based Neighborhood Discovery in Traffic Networks

Traffic networks comprise of sensors monitoring large numbers of roadways and highways with multiple lanes. Data from such sensors can be used for various monitoring tasks such as identifying high usage roads, traffic congestion and HOV lane demarcations. In this paper we propose a method to identify spatial and temporal neighborhoods in such traffic sensor networks. This approach can be used to demarcate HOV lane restrictions at certain time periods and at certain key locations on heavy usage highways. In many cases HOV lane restrictions are dynamic and our approach can provide automatic input to which time periods and locations should be designated as HOV lanes. We propose a spatio-temporal representation model for traffic networks, which models the spatio-temporal data using high-order tensor instead of the traditional vector model. We use tensor operations and tools, such as the High-order Singular Value Decomposition(HOSVD) for dimension reduction. Subsequently, a traditional clustering algorithm such as k-means is applied in the tensor subspace. For temporal neighborhood discovery we apply K-means to the subspace of time. Similarly, for spatial neighborhoods we apply K-means to the subspace of space. In real world traffic data we found that tensor based representations produce much more accurate results than traditional models. In this paper our focus is on traffic datasets which are typically spatio-temporal in nature as they measure a phenomenon at a particular location over a period of time, however this approach is generalizable to other spatiotemporal datasets as well.