Critical node identification is an important topic in the study of complex networks. Traditional methods have problems such as insufficient node differentiation and single evaluation dimension, while machine learning-based methods face the challenges of long running time and high training cost. In order to solve these problems, a three-stage progressive learning model (GAMAL) with “pre-training-adaptive transfer-prediction” was proposed to realize the identification of key nodes. Firstly, based on the improved graph attention network (GATv2), a multi-scale encoder was constructed to aggregate local neighborhood and global topological features, and the synthetic network pre-training model was used to establish the initial feature mapping relationship. Then, through the active learning strategy of AP(affinity propagation) clustering and maximum uncertainty sampling, representative nodes in the real network were selected to realize the adaptive migration of the model. Finally, the influence score of the nodes in the real network was predicted according to the model, and the key nodes were identified. Experimental results on 10 real networks show that the proposed method performs well in terms of discrimination, accuracy and running time, and can accurately and efficiently identify the key nodes of complex networks.