SFB-GNN: Spectral Fourier–Bayesian Framework forDynamic Community Detection in Graph Neural Networks

Daozheng Qu, Yanfei Ma, Yibo WangOpenReview 2026 · OpenReview

Abstract

Community discovery in attributed graphs is acutely responsive to slight structural and feature alterations, frequently resultingin sudden assignment modifications despite strong clustering accuracy in pristine conditions. This deceptive instabilitycompromises the dependability of graph learning models in practical situations, when noise, temporal changes, and structuralambiguity are inevitable. This study introduces SFB-GNN, a Spectral Fourier–Bayesian Graph Neural Network that amalgamatesspectral stability regulation, Bayesian uncertainty modeling, and community-aware message propagation within a cohesiveframework. The model restricts the spectrum response of graph propagation operators using polynomial Fourier filtering, whilevariational Bayesian inference is utilized to measure epistemic uncertainty. Comprehensive experiments on various real-worldgraphs, ranging from weak to strong homophily, sparse to dense connectivity, and small to large scales, reveal that SFB-GNNconsistently mitigates assignment drift amid both structural and feature perturbations while maintaining competitive clusteringaccuracy. The sensitivity analysis conducted under regulated spectral gap conditions confirms that the proposed frameworkattains significant stability improvements without causing representation collapse or excessive smoothing. These findingsconfirm stability-aware community detection as a systematic extension of graph representation learning, integrating spectralgraph theory, probabilistic inference, and resilient graph modeling.
BibTeX
@misc{qu2026sfb,
  author = {Daozheng Qu and Yanfei Ma and Yibo Wang},
  title = {{SFB-GNN: Spectral Fourier–Bayesian Framework forDynamic Community Detection in Graph Neural Networks}},
  year = {2026},
  url = {https://openreview.net/forum?id=Ao9nKWpiIO},
  howpublished = {OpenReview},
}

Loading related content...

Figure