High crime rates have become a public health problem in many important cities, according to World Health Organization. Many researchers have been developing algorithms to predict crime occurrences to tackle this problem. The smart cities' environment can provide us enough ubiquitous data, e.g., traffic flow, human mobility, and Points of Interest (POI) information, to feed those predictive policing algorithms and reflect city dynamics. POIs data provide essential information such as geographical location, category, customer reviews, and busy hours. Recent studies have shown that POI geographical locations are useful for predictive policing. In this paper, we aim at predicting crimes in a delimited region around the POIs of a city with new environmental features. We investigate the relevance of POIs location and the semantic and the temporal features from POIs data in our problem. We also propose and analyze different machine learning approaches to train prediction functions based on these features and conduct experiments on real crime data over multiple years. The experiments demonstrate that the popular time feature is more relevant than the historical information about the number of crimes around a POI, but both information is much less critical than the spatio-temporal information. This work is the first that studies the popular time feature extracted from POIs data and historical criminal information for predictive policing from the authors' knowledge.