In addition to being a fundamental infrastructure for communication, cellular networks are increasingly employed for outdoor positioning through signal fingerprinting. In this respect, the choice of the specific strategy used to obtain a position estimation from fingerprints plays a major role in determining the overall accuracy. In this paper, we propose a novel fingerprint comparison method, to be used in dynamic and large-scale contexts, such as the outdoor one, based on a machine learning approach. We explore two possible machine learning solutions, that make use of decision tree ensembles and support vector machines, respectively, and carefully contrast and evaluate them against a set of well-known, state-of-the-art fingerprint comparison functions from the literature. Tests are carried out with different tracking devices and environmental settings. It turns out that the machine learning approach, especially when implemented using decision tree ensembles, provides consistently better estimations than all the other considered strategies.
