Network alignment is a key way to integrate data from different platforms. Obtaining node representations by using network representation learning and establishing node matching strategies is one of the current mainstream technologies for alignment of heterogeneous networks. In this kind of research， network representation model and computational complexity are two key problems. This paper proposes an unsupervised network alignment method based on multi-scale feature modeling and improved sampling strategy. Firstly， a node feature representation with different scales is proposed to extract node features. Then， a network embedding model is used to obtain the initial representation of the network. On this basis， a sampling strategy based on node importance is designed to select landmark nodes and improve the random sampling strategy. The similarity matrix of network nodes based on landmark nodes is established， and the low rank matrix approximation is introduced. Finally， the two networks are aligned according to the similarity of node representation. Experimental results on three data sets show that the proposed model is better than other baselines.