Fast and Robust Comparison of Probability Measures in Heterogeneous Spaces

Ryoma Sato, Marco Cuturi, Makoto Yamada, Hisashi Kashima

Published 2020-02-05Version 1

The problem of comparing distributions endowed with their own geometry appears in various settings, e.g. when comparing graphs, high-dimensional point clouds, shapes, and generative models. Although the Gromov Wasserstein (GW) distance is usually presented as the natural geometry to handle such comparisons, computing it involves solving a NP-hard problem. In this paper, we propose the Anchor Energy (AE) and Anchor Wasserstein (AW) distances, simpler alternatives to GW that build upon the representation of each point in each distribution as the 1D distribution of its distances to all other points. We propose a sweep line algorithm to compute AE \emph{exactly} in $O(n^2 \log n)$, where $n$ is the size of each measure, compared to a naive implementation of AE requires $O(n^3)$ efforts. This is quasi-linear w.r.t. the description of the problem itself. AW can be pending a single $n^3$ effort, in addition to the $O(n^2)$ cost of running the Sinkhorn algorithm. We also propose robust versions of AE and AW using rank-based criteria rather than cost values. We show in our experiments that the AE and AW distances perform well in 3D shape comparison and graph matching, at a fraction of the computational cost of popular GW approximations.