Decision-making on chatter-free machining parameters is essential for process planning since chatter significantly affects production quality and efficiency. Stability lobe diagram (SLD) is commonly used for selecting chatter-free machining parameters, but its analytical prediction often has poor accuracy and experiment-based prediction is time-consuming. This paper proposes a multi-fidelity (MF) surrogate model and transfer learning-based method to improve the milling stability analysis. Firstly, an analytical stability model is constructed to predict low-fidelity (LF) SLDs for key combinations of radial cutting width (a e) and feed rate per tooth (ft ). A few spindle speeds (n s) are selected from each key LF SLD to detect high-fidelity (HF) stability limits (ap lim) through milling experiments. Subsequently, sufficient LF and limited HF combinations of n s, a e, ft , and ap lim are taken to construct additive scaling function-based MF stability models. Predicted MF combinations of n s, a e, ft , and ap lim are combined with limited HF combinations to construct more accurate stability models through transfer learning. Then, a neural network is ultimately trained to predict ap lim values for arbitrary combinations of n s, a e, and ft . A detailed experimental validation indicates that the proposed method can provide more accurate lobe boundaries for machining parameters selection by introducing fewer experimental samples.