ADAPT-Net: an adaptive dynamic attention and persistence-aware transformer for overlapping community detection in complex social networks

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Overlapping community detection in dynamic complex networks is challenging because memberships evolve over time, interactions arrive irregularly, and genuine bridge roles must be distinguished from noisy multi-community assignments. The aim of this study is to develop and evaluate adaptive dynamic attention and persistence-aware transformer (ADAPT-Net), an adaptive dynamic attention and persistence-aware transformer for overlapping community detection in complex social networks. The proposed framework integrates adaptive event-driven graph partitioning, multiscale graph attention encoding, learnable node-centric scale fusion, cross-snapshot link-aware temporal attention, and density-preserving overlap regularization. The primary controlled evaluation uses one processed benchmark instance derived from the Real-world Dynamic Networks (DynaMo) collection. Under a unified five-fold held-out protocol, ADAPT-Net records an F1-score of 96.70 ± 0.65, precision of 95.75 ± 1.50, recall of 97.69 ± 0.49, Jaccard similarity of 93.62 ± 1.22, Omega Index of 88.91 ± 3.01, normalized mutual information (NMI) of 84.43 ± 2.59, area under the receiver operating characteristic curve (AUC–ROC) of 99.77 ± 0.11, and area under the precision–recall curve (AUPRC) of 99.59 ± 0.17. Additional label-independent structural validation is reported for Reddit, dblp computer science bibliography (DBLP), and Enron using modularity, coverage, conductance, internal density, overlap rate, mean memberships per node, and temporal switch rate. The controlled and external analyses address different questions: the processed DynaMo-derived benchmark instance measures agreement with the reference memberships contained in the evaluation folds, whereas the external networks assess structural cohesion and temporal stability without equivalent reference overlaps. The current implementation is trained offline on snapshot sequences and supports snapshot-level inference rather than fully online parameter updates.