Aiming at the problem that the uniform sampling function is widely used in aggregate graph embedding methods to construct neighborhoods for nodes in a graph, i.e. neighbor nodes are sampled randomly, and the differences of their properties are neglected, a non-uniform neighbor nodes sampling method was proposed. The neighbor nodes of the target node with larger degrees are sampled preferentially. Some neighbor nodes with lower degrees are hidden so that they do not appear during the sampling process. The remaining nodes in the neighbor node set are randomly sampled to preserve sampling randomness, then these randomly sampled nodes and the preferentially sampled nodes form the neighborhood of the target node. The proposed non-uniform neighbor nodes sampling method was applied to the graph embedding process. Experimental results showed that the proposed method can be used to improve the classification F1 score of graph embedding to 91.7% on the Reddit dataset, and the results was superior than that of several known graph embedding methods. Experiments on the overlapping community dataset PPI also confirmed that the proposed method can be used to generate embedding with higher quality for graph data.