DropEdge is a novel technique designed to address over-fitting and over-smoothing in deep Graph Convolutional Networks (GCNs) for node classification. By randomly removing edges from the input graph during training, DropEdge acts as a data augmenter and message passing reducer, enhancing the generalization ability of GCNs. Theoretical analysis shows that DropEdge either slows the convergence of over-smoothing or reduces information loss caused by it. DropEdge is a general technique that can be integrated with various GCN backbones, such as GCN, ResGCN, GraphSAGE, and JKNet, to improve performance. Extensive experiments on benchmark datasets demonstrate that DropEdge consistently improves the performance of both shallow and deep GCNs. The effectiveness of DropEdge in preventing over-smoothing is empirically validated. The method is implemented in Python and is available on GitHub. DropEdge is shown to be effective in preventing over-smoothing by reducing the convergence speed of over-smoothing or mitigating information loss. It is also compatible with Dropout and can be used in conjunction with other techniques like layer-wise DropEdge. The results show that DropEdge significantly improves the performance of GCNs on various tasks, including node classification on social networks and citation datasets. The method is expected to contribute to the development of deeper and more expressive GCNs for broader applications.DropEdge is a novel technique designed to address over-fitting and over-smoothing in deep Graph Convolutional Networks (GCNs) for node classification. By randomly removing edges from the input graph during training, DropEdge acts as a data augmenter and message passing reducer, enhancing the generalization ability of GCNs. Theoretical analysis shows that DropEdge either slows the convergence of over-smoothing or reduces information loss caused by it. DropEdge is a general technique that can be integrated with various GCN backbones, such as GCN, ResGCN, GraphSAGE, and JKNet, to improve performance. Extensive experiments on benchmark datasets demonstrate that DropEdge consistently improves the performance of both shallow and deep GCNs. The effectiveness of DropEdge in preventing over-smoothing is empirically validated. The method is implemented in Python and is available on GitHub. DropEdge is shown to be effective in preventing over-smoothing by reducing the convergence speed of over-smoothing or mitigating information loss. It is also compatible with Dropout and can be used in conjunction with other techniques like layer-wise DropEdge. The results show that DropEdge significantly improves the performance of GCNs on various tasks, including node classification on social networks and citation datasets. The method is expected to contribute to the development of deeper and more expressive GCNs for broader applications.