Make It Flat: Multidimensional Scaling of Citywide Traffic Data

Citywide urban traffic forecasting is widely acknowledged as beneficial yet challenging approach. One of the main obstacles for discovering and utilising relationships of traffic flows at a city level is an extreme complexity and high-dimensionality of the resulting data structure. In this paper we propose multidimensional scaling of actual spatiotemporal traffic data into regular image-like (two-dimensional) and video-like (three-dimensional) structures. Further we adopted existing approaches to image and video processing for making conclusions on the predictability of scaled traffic data. Spatial correlation and filtering were used for analysis of image-like traffic representation and an artificial neural network of a specific architecture – for prediction of video-like traffic representation. The proposed approach was empirically tested on a large real-world urban traffic data set and demonstrated its practical utility for traffic forecasting. In addition, we analysed the effects of different distance definitions (geographical, travel time-based, cross correlation-based, and dynamic time wrapping distance) and concluded the preference of travel time-based and cross correlation-based distances for discovering the spatiotemporal structure of traffic flows.