Abstract: The existing self-training few-shot node classification methods have two shortcomings: (1) pseudo labels learned from small support set are prone to errors or overconfidence issues; (2) fixed one-way teacher model is adopted to supervise student model, which cannot dynamically adjust knowledge transfer mode according to student learning progress, all of which weaken self-training effectiveness. Therefore, a self-training few-shot node classification method based on mutual learning knowledge distillation(CMLCKD-FSNC) is proposed. This method constructed dual teacher model: pre-training a graph convolutional network as main teacher on basic dataset, pre-training a graph convolutional network as auxiliary teacher on high-order topological graph of basic dataset. A student model consistent with main teacher architecture was initialized. By using representation distillation technology, all graph convolution layer information of main teacher was transferred to student model. Mutual learning pseudo label distillation technology was adopted to supervise student model pseudo label soft prediction with dual teacher prediction enhanced soft labels, improving pseudo label quality, and teacher model can dynamically adjust knowledge distribution according to student model to enhance student model acceptance ability. The superiority of this method was verified through experiments on six graph datasets.