A1712
Title: Robustness in weighted networks
Authors: Dario Righelli - University of Naples Federico II (Italy)
Luisa Cutillo - University of Leeds (United Kingdom) [presenting]
Valeria Policastro - University of Naples Federico II (Italy)
Annamaria Carissimo - IAC-CNR (Italy)
Abstract: In network analysis, numerous community detection algorithms have been developed, yet their statistical validation remains underexplored. The R package ROBIN introduced a method for testing the robustness of unweighted networks using a configuration model as a null hypothesis to determine if detected communities result from random edge placement. We extend this approach to weighted networks and propose a new machine learning-based method for robustness analysis. The goal is to validate the robustness of community detection algorithms and compare their performance under various perturbation strategies. For weighted graphs, we employ the weighted configuration model (WCM), which preserves node strength sequences while randomizing edge positions and weights. This preserves the heterogeneity of node strengths, allowing for a statistical test of community structures. Our perturbation strategy rewires a percentage of the network's edges while maintaining the degree and weight distributions. This controlled perturbation ensures that key network properties are preserved. We tested this method using the LFR benchmark model, which simulates networks with power-law distributions for degree and community sizes. Various configurations were explored to evaluate algorithm robustness. Results demonstrate significant differences in algorithm stability, providing insights into which methods are more reliable for detecting communities in weighted networks under perturbation.
