Traditional milling stability analysis has relatively low prediction accuracy under real working conditions for using the static tool tip frequency response functions (FRFs) and average cutting force coefficients. Therefore, a milling stability prediction method based on a small number of experimental samples is proposed by introducing transfer learning. First,the tool tip FRFs at idle state and the average cutting force coefficients are measured to generate multiple series of random values within the spindle speed range. An optimal series is determined by comparing the limited experimental stability limits and their related predicted values, and it is used to further construct sufficient source stability data close to the real data. On the basis, a multi-layer perceptron model for predicting the stability limits is formulated by the source data, and it is globally fine-tuned by the limited target experimental samples for adapting to the real machining scene. Forty groups of chatter experimental samples are used to develop a validation case study. The prediction accuracy of the proposed method is 32% higher than that of the model constructed only using the 40 samples. In addition, accuracies of different types of prediction models trained by different target data sizes are also compared to evaluate the effectiveness of the proposed method.