In this paper, we concentrate on graph clustering by using graph spectral features. The leading eigenvectors or the spectrum of graphs and derived feature inter-mode adjacency matrix are used. The embedding methods are the Locality Preserving Projection(LPP) and the mixtures of LPP. The experiment results show that although both of the conventional LPP and the LPP mixtures can separate the different graphs into outstanding clusters, the conventional LPP outperforms the LPP mixtures in the sense of compactness for graph clustering.