Graph Convolutional Neural Networks (Graph CNNs) are generalizations ofclassical CNNs to handle graph data such as molecular data, point could andsocial networks. Current filters in graph CNNs are built for fixed and sharedgraph structure. However, for most real data, the graph structures varies inboth size and connectivity. The paper proposes a generalized and flexible graphCNN taking data of arbitrary graph structure as input. In that way atask-driven adaptive graph is learned for each graph data while training. Toefficiently learn the graph, a distance metric learning is proposed. Extensiveexperiments on nine graph-structured datasets have demonstrated the superiorperformance improvement on both convergence speed and predictive accuracy.