In the last years, municipal governments have started to share open data about citizens’ activities in smart city environments. In particular, the Open311 project has fostered the release of non-emergency call reports thus enabling the use of Machine Learning techniques to discover actionable knowledge. To delimit the intervention area and the period of analysis in response to specific non-emergency warning categories, police and municipality officers would need precise spatial and temporal indications. To this end, we present a new approach based on the discovery of multiple-level flipping correlations between non-emergency warning categories and contextual features. We define the new concept of SpatioTemporal Warning Context to characterize the spatial and temporal conditions in which a particular warning category or subcategory is likely (or unlikely) to occur. We carry out a preliminary validation of the proposed approach on real open data acquired from the smart cities of New York City (U.S.) and Turin (Italy).