In this study was shown that random forest (RF) can be used as a sensible new data mining tool for variable importance measurements (VIMs) through various coal properties for prediction of coke quality (Free Swelling Index (FSI)). The VIMs of RF within coal analyses (proximate, ultimate, and petrographic analyses) were applied for the selection of the best predictors of FSI over a wide range of Kentucky coal samples. VIMs assisted by Pearson correlation through proximate, ultimate, and petrographic analyses indicated that volatile matter, carbon, vitrinite, and Rmax (coal rank parameters) are the most effective variables for the prediction of FSI. These important predictors have been used as inputs of RF model for the FSI prediction. Outputs in the testing stage of the model indicated that RF can predict FSI quite satisfactorily; the R2 was 0.93 and mean square error from actual FSIs was 0.15 (had less than interval unit of FSI; 0.5). According to the result, by providing nonlinear inter-dependence approximation among parameters for variable selection and also non-parametric predictive model RF can potentially be further employed as a reliable and accurate technique for the determination of complex relationship through fuel and energy investigations.