Graph Convolutional Networks (GCNs) have been widely used in recommendation systems due to the graph-structured nature of recommendation data. However, existing GCN-based recommendation methods face two key challenges: 1) data sparsity problem, where sparse datasets make it hard for algorithms to extract useful information; 2) over-smoothing problem, where stacking convolutional layers makes embedding representations too similar, causing dissimilar users to have similar interests and weakening recommendation performance. To address these problems, we propose a novel dual-channel social graph convolutional networks, dubbed DCS-GCN, for recommendation. Building on traditional GCN-based recommendation frameworks, DCS-GCN enriches user information through social graph-based message passing to alleviate data sparsity problem. It also diversifies the aggregation mechanism by mining and passing social relationship information in the user-item interaction graph, to mitigate over-smoothing problem. Finally, the model fuses these features from two social channels at each graph convolutional layer to generate final user and item representations for preference prediction. Experiments on two public datasets show that DCS-GCN outperforms several state-of-the-art recommendation methods (Our code is available at https://github.com/anonymity73/DCS-GCN ).