Extracting the geometric backbone of bipartite networks

Abstract

Real bipartite networks combine degree-constrained random mixing with structured, locality-like rules. We introduce a statistical filter that benchmarks node-level bipartite clustering against degree-preserving randomizations to classify nodes as geometric (signal) or random-like (noise). In synthetic mixtures with known ground truth, the filter achieves high F-scores and sharpens inference of latent geometric parameters. Applied to four empirical systems -- metabolism, online group membership, plant-pollinator interactions, and languages -- it isolates recurrent neighborhoods while removing ubiquitous or weakly co-occurring entities. Filtering exposes a compact geometric backbone that disproportionately sustains connectivity under percolation and preserves downstream classifier accuracy in node-feature tasks, offering a simple, scalable way to disentangle structure from noise in bipartite networks.