Improving the fuel efficiency of vehicles is imperative to reduce costs and protect the environment. While the efficient engine and vehicle designs, as well as intelligent route planning, are well-known solutions to enhance the fuel efficiency, research has also demonstrated that the adoption of fuel-efficient driving behaviors could lead to further savings. In this work, we propose a novel framework to promote fuel-efficient driving behaviors through real-time automatic monitoring and driver feedback. In this framework, a random-forest based classification model developed using historical data to identifies fuel-inefficient driving behaviors. The classifier considers driver-dependent parameters such as speed and acceleration/deceleration pattern, as well as environmental parameters such as traffic, road topography, and weather to evaluate the fuel efficiency of one-minute driving events. When an inefficient driving action is detected, a fuzzy logic inference system is used to determine what the driver should do to maintain fuel-efficient driving behavior. The decided action is then conveyed to the driver via a smartphone in a non-intrusive manner. Using a dataset from a long-distance bus, we demonstrate that the proposed classification model yields an accuracy of 85.2% while increasing the fuel efficiency up to 16.4%.