DIFFUSION-INDUCED SPATIAL ATTENTION OVERLAPPING COMMUNITY DETECTION

Community detection is a fundamental task in network science, yet traditional graph neural networks (GNNs) often struggle with overlapping communities due to the limitations of local message passing. Standard GCNs and GATs rely on immediate neighbors, which can lead to oversmoothing and an inability to capture long-range structural dependencies. The authors introduce DISCO, a deep-learning framework designed to overcome these limitations by integrating a structural prior derived from influence-spreading dynamics into a sparse multi-head attention mechanism.

DISCO uses an Influence Spreading Model (ISM) to compute pairwise spreading probabilities, which serve as a structural prior. This prior biases the attention mechanism, enabling nodes to attend to structurally significant nonlocal nodes. The framework employs a Bernoulli-Poisson edge-reconstruction objective, allowing it to infer overlapping community memberships from node attributes, structural profiles, or a combination of both.

Benchmark experiments on Facebook ego networks demonstrate that DISCO performs competitively against established GCN and GAT approaches. The model consistently outperforms these baselines in both topology-driven and attribute-driven configurations. The ablation study reveals that the diffusion-induced prior is particularly beneficial when node attributes are used without explicit structural input, as it effectively injects topological information into the attention process.

Beyond static benchmarks, the authors present a proof-of-concept cybersecurity use case in an operational technology (OT) network. By monitoring temporal changes in inferred community assignments, DISCO provides an interpretable anomaly signal. The model successfully detects various simulated cyberattacks and malfunctions, with node-level contributions helping to localize the devices associated with the structural deviations.

DISCO provides a flexible, principled method for overlapping community detection that bridges the gap between local message passing and computationally expensive global attention. By using diffusion-derived priors, it offers a scalable way to incorporate higher-order graph structure into deep learning models. Furthermore, the application to dynamic OT networks highlights the potential of community-based analysis for real-time network monitoring and anomaly detection, offering a more interpretable alternative to black-box intrusion detection systems.