Understanding of mobile internet traffic patterns and capacity to estimate future traffic, particularly at high spatiotemporal granularity, is crucial for proactive decision making in emerging and future cognizant cellular networks enabled with self-organizing features. It becomes even more important in the world of 'Internet of Things' with machines communicating locally. In this paper, internet activity data from a mobile network operator Call Detail Records (CDRs) is analysed at high granularity to study the spatiotemporal variance and traffic patterns. To estimate future traffic at high granularity, a Support Vector Regression (SVR) based traffic model is trained and evaluated for the prediction of maximum, minimum and average internet traffic in the next hour based on the actual traffic in the last hour. Performance of the model is compared with that of the State-of-the-Art (SOTA) deep learning models recently proposed in the literature for the same data, same granularity, and same predicates. It is concluded that this SVR model outperforms the SOTA deep and non-deep learning methods used in the literature. © 2018 IEEE.