<p>The rapid growth of digital news platforms has brought about more challenges than ever before for the comprehension and anticipation of the spread of news content across geographic regions over time. The identification of several patterns of hotspots related to the diffusion of the news is of paramount importance to media organizations, policy-makers, and social scientists who aim to explain the dynamics of information flows. Conventional methods have been based on basic statistical modeling or simple machine learning methods that do not consider the complex spatio-temporal interactions built in the news diffusion process. The current approaches have weaknesses such as failure to deal with non-linear relationships, lack of generalization across geographic scales and lack of features representation to elicit temporal dynamics. The paper describes XGBoost-News, a new machine learning model that uses extreme gradient boosting to predict hotspots of news diffusion spatio-temporally. Our model presents three main novelties, namely, (1) an explicit spatio-temporal feature engineering pipeline that encodes the importance of geographic proximity, temporal decay, and factors of cross-regional influence; (2) a dynamic mechanism used to weight spatial and temporal features; and (3) a cascading prediction model that reflects the sequential nature of news diffusion between interconnected regions. Through wide-spread experiments on three large-scale real-world datasets, it is demonstrated that the performance of XGBoost-News is significantly better than prevailing state-of-the-art methods, with accuracy gains of up to 23.5%, F1 gains of up to 33.9% and AUC-ROC gains from 3.3 to 13.0%. Furthermore, the proposed framework is more efficient in terms of computation; we achieve a 73% training time saving with respect to the deep-learning alternatives, and it can be run on CPU hardware.</p>