Abstract: The problem of data with noise and sparsity of heterogeneous information networks can not be solved by the traditional feature extraction methods efficiently due to their semantics and complicated structure. Stacked denoising auto encoder was introduced to learn the features of sample. The relax strategy was employed to construct class hierarchy with high-quality, and then the nodes of the heterogeneous information network were classified and ranked. Experimental results on the dataset of DBLP (digital bibliography & library project) show that the method is effective, and the precision of classification is 86.3%.