The fifth generation of cellular networks (5G) will rely on edge cloud deployments to satisfy the ultra-low latency demand of future applications. In this paper, we argue that an edge-based deployment can also be used as an enabler of advanced Machine Learning (ML) applications in cellular networks, thanks to the balance it strikes between a completely distributed and a centralized approach. First, we will present an edge-controller-based architecture for cellular networks. Second, by using real data from hundreds of base stations of a major U.S. national operator, we will provide insights on how to dynamically cluster the base stations under the domain of each controller. Third, we will describe how these controllers can be used to run ML algorithms to predict the number of users, and a use case in which these predictions are used by a higher-layer application to route vehicular traffic according to network Key Performance Indicators (KPIs). We show that prediction accuracy improves when based on machine learning algorithms that exploit the controllers' view with respect to when it is based only on the local data of each single base station.