Seismic events in Italy and worldwide have highlighted the high vulnerability of unreinforced masonry (URM) structures in small historical centres. A key feature of these settlements is to be mostly composed of buildings in aggregate, i.e., interconnected by a more or less structurally effective connection. The seismic assessment of such buildings is quite debated in the literature and no shared tools procedures are currently available. The difficulty of standardization derives from the fact that structural units can be characterized by multiple features and configurations that determine a vast number of vulnerability factors, whose interdependency is not straightforward to be identified. The paper addresses this issue by combining evidence-based damage data with the potential offered by Machine Learning (ML) technique. Real data are used in combination with state-of-the-art ML algorithms carefully tuned via an advanced statistical procedure for two main purposes. The first one will be able to predict possible URM damages based on the vulnerability factor in both interpolation and extrapolation scenarios. The second purpose of the ML-based techniques will be to predict the most important vulnerability factors in making these predictions, namely to make the ML-based model explainable and informative about the underlying phenomena and not just predictive. The small historic centre of Casentino, hit by the 2009 L'Aquila earthquake, is adopted in the paper as the first test case study. A large amount of data was collected after the earthquake through in-situ surveys made by the Universities of Genova, Catania and Rome. Data include both geometric and structural factors, i.e., the input data supplied to the ML algorithm, as well as the actual seismic damage mechanisms, i.e., the output data expected to be predicted by the ML algorithm. As first application, ML techniques are applied only to data acquired on out-of-plane mechanisms.