Our model is based on the graph convolutional network (GCN) with node feature (coordination and demand) and edge feature (the real distance between nodes) as input and embedded. Separate decoders are proposed to decode the representations of these two features. The output of one decoder is the supervision of the other decoder. We propose a strategy that combines the reinforcement learning manner with the supervised learning manner to train the model. Through comprehensive experiments on real-world data, we show that 1) the edge feature is important to be explicitly considered in the model; 2) the joint learning strategy can accelerate the convergence of the training and improve the solution quality; 3) our model significantly outperforms several well-known algorithms in the literature, especially when the problem size is large; 3) our method is generalized beyond the size of problem instances they were trained on.