@Article{cmc.2026.087484,
AUTHOR = {Xiangmin Liu, Weizhi Xiong, Wei Zhang, Taojian Luo, Xiu Yao, Jian Hu, Peng Peng},
TITLE = {Relation-Aware Adaptive Anchor Spectral Clustering for Large-Scale Attributed and Heterogeneous Graphs},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/28272},
ISSN = {1546-2226},
ABSTRACT = {Spectral clustering (SC) effectively captures non-convex and graph-structured cluster patterns, but its application to large-scale attributed and heterogeneous graphs is hindered by the high costs of similarity graph construction and Laplacian eigendecomposition, as well as by insufficient use of relation semantics. To address these challenges, we propose Relation-aware Adaptive Anchor Spectral Clustering (RAASC). RAASC first identifies relation-aware local components in relation-specific subgraphs and allocates anchor budgets according to component difficulty. It then constructs an adaptive, load-aware sample-anchor graph by jointly considering node uncertainty, relation compatibility, and anchor-load balance. Using the resulting sparse sample-anchor matrices, RAASC builds a compact relation-aware reduced anchor graph and performs spectral learning, anchor-level discretization, and label recovery in the anchor space. This design avoids full similarity graph construction and reduces the scale of eigendecomposition while preserving heterogeneous structural information. Experiments on DBLP, IMDB, and OGBN-MAG show that RAASC achieves the highest mean Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI) among the representative clustering and spectral clustering baselines. Efficiency and scalability results further demonstrate that RAASC maintains practical runtime and memory usage on large-scale heterogeneous graphs and achieves a 3.72-fold speedup using 16 workers.},
DOI = {10.32604/cmc.2026.087484}
}