This paper proposes a novel method for early action prediction based on 3D skeleton data. Our method combines the advantages of graph convolutional networks (GCNs) and adversarial learning to avoid the problems of insufficient spatio-temporal feature extraction and difficulty in predicting actions in the early execution stage of actions. In our method, GCNs, which have outstanding performance in the field of action recognition, are used to extract the spatio-temporal features of the skeleton. The model learns how to optimize the feature distribution of partial videos from the features of full videos through adversarial learning. Experiments on two challenging action prediction datasets show that our method performs well on skeleton-based early action prediction. State-of-the-art performance is reported in some observation ratios.