Abstract
The Gromov-Wasserstein (GW) variant of optimal transport, designed to compareprobability densities defined over distinct metric spaces, has emerged as animportant tool for the analysis of data with complex structure, such asensembles of point clouds or networks. To overcome certain limitations, such asthe restriction to comparisons of measures of equal mass and sensitivity tooutliers, several unbalanced or partial transport relaxations of the GWdistance have been introduced in the recent literature. This paper is concernedwith the Conic Gromov-Wasserstein (CGW) distance introduced byS\'{e}journ\'{e}, Vialard, and Peyr\'{e}. We provide a novel formulation interms of semi-couplings, and extend the framework beyond the metric measurespace setting, to compare more general network and hypernetwork structures.With this new formulation, we establish several fundamental properties of theCGW metric, including its scaling behavior under dilation, variationalconvergence in the limit of volume growth constraints, and comparison boundswith established optimal transport metrics. We further derive quantitativebounds that characterize the robustness of the CGW metric to perturbations inthe underlying measures. The hypernetwork formulation of CGW admits a simpleand provably convergent block coordinate ascent algorithm for its estimation,and we demonstrate the computational tractability and scalability of ourapproach through experiments on synthetic and real-world high-dimensional andstructured datasets.