The performances of GNNs for representation learning on the graph-structured data are generally limited to the issue that existing GNNs rely on one assumption, i.e., the original graph structure is reliable. However, since real-world graphs is inevitably noisy or incomplete, this assumption is often unrealistic. In this paper, we propose a structure learning graph convolutional networks (SLGCNs) to alleviate the issue from two aspects, and the proposed approach is applied to node classification. Specifically, the first is node features, we design a efficient-spectral-clustering with anchors (ESC-ANCH) approach to efficiently aggregate feature representationsfrom all similar nodes, no matter how far away they are. The second is edges, our approach generates a re-connected adjacency matrix according to the similarities between nodes and optimized for the downstream prediction task so as to make up for the shortcomings of original adjacency matrix, considering that the original adjacency matrix usually provides misleading information for aggregation step of GCN in the graphs with low level of homophily. Both the re-connected adjacency matrix and original adjacency matrix are applied to SLGCNs to aggregate feature representations from nearby nodes. Thus, SLGCNs can be applied to graphs with various levels of homophily. Experimental results on a wide range of benchmark datasets illustrate that the proposed SLGCNs outperform the stat-of-the-art GNN counterparts.