The use of big data in transportation research is increasing and this leads to new approaches in modeling the traffic flow, especially for what concerns metropolitan areas. An open and interesting research issue is city-wide traffic representation, correlating both spatial and time patterns and using them to predict the traffic flow through the whole urban network. In this paper we present a machine learning based methodology to model traffic flow in metropolitan areas with the final aim to address short-term traffic forecasting at various time horizons. Specifically, we introduce an ad-hoc neural network model (GBNN, Graph Based Neural Network) that mirrors the topology of the urban graph: neurons corresponds to intersections, connections to roads, and signals to traffic flow. Furthermore, we enrich each neuron with a memory buffer and a recurrent self loop, to model congestion and allow each neuron to base its prediction on previous local data. We created a GBNN model for a major Italian city and fed it one year worth of fine-grained real data. Experimental results demonstrate the effectiveness of the proposed methodology in performing accurate lookahead predictions, obtaining 3% and 16% MAAPE error for 5 min and 1 h forecasting respectively.