Latent space models for grouped multiplex networks

A team of researchers from the University of Michigan and other institutions has published a new paper on arXiv introducing 'GroupMultiNeSS,' a latent space model designed for grouped multiplex networks. Complex multilayer network data is ubiquitous in fields like neuroscience, social sciences, and economics, but existing models often fail to capture patterns shared only within specific subsets of layers, such as treatment versus control groups in medical studies. This new model directly addresses that gap, enabling the simultaneous extraction of shared, group-specific, and individual latent structures from a sample of networks, which is crucial for uncovering systematic differences and improving downstream tasks like hypothesis testing and visualization.

The technical core of GroupMultiNeSS involves a fitting procedure using convex optimization combined with a nuclear norm penalty, with proven recovery guarantees for latent positions given sufficient separation between the different latent subspaces. In synthetic tests, it showed apparent improvements in modeling accuracy over its predecessor, MultiNeSS, and other models. Its real-world utility was demonstrated on a Parkinson's disease brain connectivity dataset, where it successfully provided superior node-level insights into biological differences between treatment and control patient groups. This advancement represents a significant step in network analysis, allowing researchers to move beyond modeling just universal or individual variations and finally quantify the intermediate, group-level signals that are critical in comparative scientific studies.