This paper discusses deep graph clustering (DGC), which unsupervisedly classifies nodes in attribute graphs into multiple clusters. To address the challenges of real-world attribute graphs, which are often large and often lack attributes, we propose a novel DGC method, "Complementary Multi-View Neighbor Differentiation (CMV-ND)." CMV-ND preprocesses graph structural information into multiple perspectives in a complete and non-redundant manner. Specifically, it fully expands the local structure of the graph through recursive neighbor search and removes redundancy between neighbors with different hop distances through a neighbor differentiation strategy. Then, it constructs K+1 complementary perspectives from differential hop representations and target node features, and applies existing multi-view clustering or DGC methods. Experimental results on six widely used graph datasets demonstrate that CMV-ND significantly improves the performance of various methods.
