Improving adaptation and interpretability of a short-term traffic forecasting system

Traffic management is being more important than ever, especially in overcrowded big cities with over-pollution problems and with new unprecedented mobility changes. In this scenario, road-traffic prediction plays a key role within Intelligent Transportation Systems, allowing traffic managers to be able to anticipate and take the proper decisions. This paper aims to analyse the situation in a commercial real-time prediction system with its current problems and limitations. The analysis unveils the trade-off between simple parsimonious models and more complex models. Finally, we propose an enriched machine learning framework, Adarules, for the traffic prediction in real-time facing the problem as continuously incoming data streams with all the commonly occurring problems in such volatile scenario, namely changes in the network infrastructure and demand, new detection stations or failure ones, among others. The framework is also able to infer automatically the most relevant features to our end-task, including the relationships within the road network. Although the intention with the proposed framework is to evolve and grow with new incoming big data, however there is no limitation in starting to use it without any prior knowledge as it can starts learning the structure and parameters automatically from data. We test this predictive system in different real-work scenarios, and evaluate its performance integrating a multi-task learning paradigm for the sake of the traffic prediction task.Peer Reviewe