B0668
Title: Node clustering in large-scale graphical models
Authors: Andreas Alfons - Erasmus University Rotterdam (Netherlands) [presenting]
Daniel Touw - Erasmus University Rotterdam (Netherlands)
Ines Wilms - Maastricht University (Netherlands)
Abstract: Graphical models can represent conditional dependency structures among a large number of variables in a compact manner, especially when relying on edge-sparsity as a simplifying structure. Since data are nowadays measured and analyzed at ever-higher resolutions or more disaggregate levels, similarity in the conditional dependency structure among different nodes becomes a new, interesting guiding principle for dimensionality reduction. We develop an unsupervised node-clustering method for large-scale graphical models. Our aim is to find clusters of nodes in the network that share their conditional dependency structure. To this end, we leverage recent advances from convex clustering within a regularization framework, and we show how node-clustering can be efficiently combined with other popular penalization structures for graphical models such as edge-sparsity. We compare the methods estimation and clustering performance to several other state-of-the-art benchmark methods.
