Having accurate spatial prediction models of air pollutant concentrations can be very helpful to alleviate the shortage of monitoring stations, specially in low-to-middle income countries. However, given the large diversity of model types, both statistical, numerical and machine learning (ML) based, it is not clear which of them are most suitable for this task. In this paper we study the predictive capabilities of common machine learning methods for the spatial prediction of PM2.5 concentration level. Three relevant factors were scrutinized: the extent to which meteorological variables impact the prediction performance; the effect of variable normalization by inverse distance weighting (IDW); and the number of neighborhood stations needed to maximize predictive performance. Results in a dataset from Beijing monitoring network show that simple models like Linear Regresors trained on IDW normalized variables can cope with this task. Some knowledge have been derived to guide the construction of competent models for spatial prediction of PM2.5 concentrations with ML-based methods.