Abstract
Community detection provides a principled lens on mesoscale organization in functional brain networks, yet many widely used methods presume assortative structure and depend on arbitrary thresholding, which complicates the selection of the community count K . We conducted a systematic benchmark of three assumption lean approaches that operate directly on weighted functional connectivity matrices: the Weighted Stochastic Block Model, Spectral Clustering, and K-means. Performance was assessed on synthetic networks with known ground truth and on three neuroimaging cohorts spanning development, namely the Human Connectome Project, Washington University 120, and the Baby Connectome Project. We compared strategies for choosing K , including post hoc indices such as silhouette, Calinski-Harabasz, C index, modularity, variation of information, Normalized Mutual Information, and zRand, together with a likelihood-based criterion for the Weighted Stochastic Block Model that uses bootstrap confidence intervals for differences in log likelihood between successive values of K . In simulations all methods recovered stable partitions, but the post hoc indices favored incorrect values of K under weak signal and nonassortative mixing. In adult datasets the indices do not yield a unique optimum, whereas the likelihood-based criterion selects a parsimonious range centered near K = 11 , which is consistent with established sensory and association systems. In infants and toddlers, the same procedure supports a larger K around 15 and reveals developmentally distinct mesoscale architecture, including anterior and posterior subdivisions within default mode and fronto parietal systems. A consensus relabeling scheme based on Hungarian matching with Hamming distance further stabilizes solutions across runs and across values of K . Overall, threshold free weighted methods mitigate assortative bias and the likelihood-based comparison provides a reproducible path to selecting K .