An intelligent distance-driven community detection framework for complex networks using the longest distance node head technique (LDNHT)

Community detection in complex networks needs methods that are not only structurally good, but also computationally efficient, stable, and interpretable. This study introduces a centroid-based community detection framework called Longest Distance Node Head Technique (LDNHT), where the first centroid is chosen randomly, and the others are determined by maximizing the total shortest-path distances from existing centroids. The method was tested on Facebook, Power-Grid, Wiki-Vote, and Gnutella networks in comparison with Random Node Head Technique (RNHT), Highest Degree Node Head Technique (HDNHT), classical max-min distance seeding, Louvain, and Fast-Greedy community detection. The stochastic methods were tested on more than 35 seeded runs, and the deterministic baselines were performed under the same experimental protocol. The modularity, External Connection Ratio (ECR), conductance, coverage, and runtime were used to assess partition quality. On Power-Grid and Wiki-Vote, its performance was significantly better than RNHT, HDNHT, and MAXMIN in terms of ECR and coverage, while Louvain and Fast-Greedy generally achieved higher modularity. Sensitivity analysis showed that the number of repeated initializations had a small impact, with the dependence on the first random centroid decreasing significantly with repeated initializations, and the benefit diminishing after about 20 internal iterations. The empirical scalability experiments on sparse networks showed approximately linear behavior, with a log-log slope of 0.911. Overall, LDNHT offers a comprehensible cumulative geodesic-dispersion strategy whose effectiveness is dependent on network structure and parameter configuration.