Machine scheduling serves as a vital function for industrial and service operations, and uncertainties always pose a significant challenge in real-world scheduling practices. In this paper, we propose to solve the stochastic machine scheduling problems with uncertain processing times by an improved prescriptive tree-based (IPTB) model. Our approach includes a novel way of combining historical processing time data with current scheduling constraints to strengthen the quality of historical decisions. We apply these improved historical decisions and incorporate an improved model for calculating the optimisation loss and accelerate the training of our IPTB model. Our trained model can directly prescribe downstream scheduling solutions with high robustness in the face of uncertainties. We evaluate the proposed IPTB method on a stochastic parallel machine scheduling problem originating from printed circuit board assembly lines. Through a series of comparative experiments, our findings demonstrate the IPTB method’s superior accuracy and robustness, highlighting its resilience in noisy data environments. Additionally, we interpret the model through feature importance analysis and examine the model’s behaviours under noisy conditions.