Graph-enhanced clustering of late modern churches via multi-dimensional semantic feature integration

Typological classification of late-modern church buildings provides a basis for understanding the spatial evolution of religious architecture and heritage characteristics. In densely distributed regions like Shanxi, church forms, locations, and institutional backgrounds are highly heterogeneous, making single-attribute clustering inadequate. This study proposes a graph-enhanced clustering method integrating GIS-based spatial adjacency, architectural configuration, and institutional affiliation, with graph neural networks extracting node embeddings for single- and multi-relation clustering. Three representative types—Standardized Basilica Expansion, Localized Hybrid Adaptation, and Highland Western Composite—reflect Shanxi’s historical transition from isolation to openness and from traditional to hybrid local-imported forms, highlighting interactions among religious dissemination, topography, and construction practices. The findings demonstrate the value of combining spatial adjacency and semantic information in graph-based models, providing a transferable approach for typological classification and supporting pattern recognition in heritage studies.