To address the difficulty in predicting recovery factor of ultra-deep fractured edge-water gas reservoirs controlled by both fracture conductivity and edge-water invasion, a machine learning-based prediction method was developed. Taking the KS gas reservoir as a case study, a fractured edge-water gas reservoir model was established using numerical simulation based on geological understanding, production data, and PVT properties. Six key parameters, including water body multiple, matrix permeability, fracture permeability, fracture density, fracture lateral penetration ratio, and fracture angle, were selected to design 180 simulation cases and construct a recovery factor dataset.Three machine learning models, including random forest (RF), least squares boosting (LSBoost), and Gaussian process regression (GPR), were developed. An ensemble model combining LSBoost, RF, and GPR was further proposed. Results show that LSBoost achieved the best performance among single models, with an R² of 0.9126 and RMSE of 4.4099. The optimized LSBoost-RF-GPR model further improved prediction accuracy, achieving R², RMSE, and MAE values of 0.9440, 3.0298, and 2.4984, respectively.The proposed method effectively captures the nonlinear relationship between reservoir parameters and recovery factor, providing an efficient approach for rapid recovery factor evaluation of ultra-deep fractured edge-water gas reservoirs.