BioMetaEvo-GNN: Bayesian Fourier Message Passing with Immune Memory for Robust Community Detection
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Community detection in real-world graphs is challenged by noisy edges, ambiguous node features, perturbation-induced assignment drift, and unstable community boundaries. Existing GNN-based methods typically optimize clustering accuracy or local robustness, but rarely unify uncertainty modeling, memory preservation, and adaptive defense. We propose BioMetaEvo-GNN, which integrates Bayesian Fourier learning, immune-memory defense, and UCB-based meta-governance. Bayesian Fourier learning identifies spectral instability and drift-prone boundaries, immune memory preserves reliable structural evidence and suppresses unstable transitions, and meta-governance adaptively selects strategies using uncertainty, stability, and historical rewards. Experiments on real-world and synthetic graphs show consistent gains in clustering accuracy, perturbation robustness, uncertainty calibration, assignment stability, and cross-dataset generalization over representative GNN and robust graph learning baselines. These results establish community detection as a joint problem of representation learning, uncertainty estimation, and adaptive robustness.
- Primary area:
- learning on graphs and other geometries & topologies