© 2019, ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering. Network Embedding aims to learn latent representations and effectively preserves structure of network and information of vertices. Recently, networks with rich side information such as vertex’s label and links between vertices have attracted significant interest due to its wide applications such as node classification and link prediction. It’s well known that, in real world applications, network always contains mislabeled vertices and edges, which will cause the embedding preserves mistake information. However, current semi-supervised graph embedding algorithms assume the vertex label is ground-truth. Manually relabel all mislabeled vertices is always inapplicable, therefore, how to effective reduce noise so as to maximize the graph analysis task performance is extremely important. In this paper, we focus on reducing label noise ratio in dataset to obtain more reasonable embedding. We proposed two methods for any semi-supervised network embedding algorithm to tackle it: first approach uses a model to identify potential noise vertices and correct them, second approach uses two voting strategy to precisely relabel vertex. To the best of our knowledge, we are the first to tackle this issue in network embedding. Our experiments are conducted on three public data sets.