Block-pair edge community detection in bipartite networks via an adaptive spectral-modularity framework

Queueing Systems · Springer Nature · 2026-09-22 · DOI: 10.1007/s10588-026-09448-2

Аннотация

Bipartite networks represent interactions between two different types of vertices, such as users and items, authors and papers, documents and words, buyers and suppliers, or senders and receivers. Most bipartite community detection methods cluster the two vertex sides, whereas many edge-clustering methods treat edges as independent objects and may lose the reusable structure of endpoint roles. This paper studies block-pair edge community detection, where an edge community is interpreted as an interaction pattern between a row-side role and a column-side role. This formulation naturally represents diagonal, shifted, many-to-many, nested, off-diagonal, and star-shaped structures, without forcing a one-to-one matching between the two sides. We propose VR-CNBM-EdgePair, an adaptive spectral-modularity framework for detecting block-pair edge communities in bipartite networks. Here, CNBM stands for centered normalized bipartite modularity, and VR refers to validation-regularized selection of the centering and degree-regularization parameters. The method uses a single two-sided operator family that connects normalized bipartite spectral co-clustering with fully centered bipartite modularity, while allowing degree-regularized versions of both. A centering parameter controls how much of the bipartite configuration-model expectation is removed, while a degree-regularization parameter stabilizes sparse or heterogeneous degrees. Both parameters are selected internally by a held-out block likelihood criterion, so the method remains unsupervised and avoids graph-specific switching rules. The selected operator is decomposed by a truncated two-sided singular value decomposition. Its left and right singular vectors define spectral coordinate systems for the row side and the column side, respectively. Weighted spherical K-means is then applied separately to these two coordinate systems, and each observed edge receives the block-pair label induced by its two endpoint roles. Thus, the method performs constrained edge clustering through reusable endpoint roles, rather than direct clustering of independent edges. Synthetic experiments on matching, multi-pair, nested, and hub-like bipartite block-pair benchmarks show strong aggregate recovery, with particularly clear gains in several degree-corrected and hub-like settings. A repeated-split sensitivity analysis further shows high partition agreement in seven of eight representative configurations, while identifying weak degree-corrected nested structure as a difficult case. Experiments on real bipartite networks further support the usefulness of adapting between normalized biadjacency structure and centered bipartite modularity.

Читать у издателя →

Похожие статьи