Abstract: A data-driven approach was proposed to predict gas saturation based on random forests machine learning algorithm. This method was applied to predict and analyze gas saturation in a complex natural gas reservoir in western China from single well data and two-dimensional seismic data respectively. The results show that：（1）The method extracts the pre-stack seismic inversion results of three elastic parameters from the uphole trace（compressional wave velocity，shear wave velocity，and density）from well log data as basic feature variables. It employs the boundary synthetic minority oversampling technique to balance the basic feature variables and corresponding gas saturation samples，generates a series of extended variables by combining the extended elastic impedance with mathematical transformations，and then uses random forests to rank the importance of the feature variables for gas saturation prediction，finally selects feature variables with higher importance for gas saturation random forests training.（2）This method significantly reduces the manual workload for feature variables extraction andselection，effectively reduces information redundancy and training bias caused by imbalanced gas saturation samples，and effectively enhances the capability of the random forests algorithm in predicting gas saturation. （3）In practical applications，the predicted gas saturation using this method shows a high correlation coefficient of 0.985 5 with the gas saturation by log interpretation. In the case of two-dimensional data，it achieves higher accuracy in gas saturation prediction compared to using 11 conventional unbalanced elastic parameters as inputs for random forests.

Key words: gas saturation, random forests, compressional wave velocity, shear wave velocity, density, feature variable, imbalance data, machine learning, gas layer prediction, seismic prediction