Abstract: In order to solve the problems of information mining of complex residential polygons, a multi-level graph partition clustering method is proposed to construct the graph model of residential polygons, and the clustering results are obtained by coarsen, matching and reconstruction, initialization and refinement of the graph model. Firstly, the Delaunay triangular network of residential buildings in the study area is constructed to generate the adjacent information graph including the research objects. Then, the similarity of the neighborhood graph is measured by five indexes of shape narrow length, size, convexity, distance and connectivity combined with the characteristics of spatial cognition and human cognition in this paper. Finally, the clustering results are obtained by using the multi-level graph partition method. In the experiment, the vector data of residential buildings in Shanghai are used for clustering analysis, and the silhouette coefficients and visual effects of improved k-Means algorithm (k-Means + +), density-based spatial clustering of applications with noise (DBSCAN) and minimum spanning tree (MST) clustering algorithms with noise robustness are compared. The experimental results show that the results of polygonal clustering analysis based on multi-level graph partition are more consistent with human cognition.