Deep Learning to the Test: an Application to Traffic Data Streams

Deep learning is a broad class of machine learning techniques based on learning data representation through multiple levels of abstraction. It has been successfully applied in several areas of research, but very few literature addressed the problem of traffic flow forecasting. Thus, driven by the belief that deep learning algorithms may capture the sharp traffic data non-linearities, we aimed to develop a deep architecture, namely a feed-forward neural network, and evaluate its performances in predicting short-term traffic streams. We illustrate our methodology, consisting in a predictors selection step and a subsequent training step, using traffic speed data of the Grande Raccordo Anulare road of Rome for the month of June 2016.