Optimal transport ($\OT$) theory provides a useful set of tools to compare probability distributions. As a consequence, the field of $\OT$ is gaining traction and interest within the machine learning community. A few deficiencies usually associated with $\OT$ include its high computational complexity when comparing discrete measures, which is quadratic when approximating it through entropic regularization; or supercubic when solving it exactly. For some applications, the fact that $\OT$ distances are not usually negative definite also means that they cannot be used with usual Hilbertian tools. In this work, we consider a particular family of ground metrics, namely tree metrics, which yield negative definite $\OT$ metrics that can be computed in linear time. By averaging over randomly sampled tree metrics, we obtain a tree-sliced-Wasserstein distance. We illustrate that the proposed tree-sliced-Wasserstein distances compare favorably with other baselines on various benchmark datasets.

本文针对最优输送（Optimal transport）理论的计算和统计问题，提出了一种新的基于树度量的方法：树切片 Wasserstein 距离。通过使用随机树评估度量，该方法可以在低或高维空间中自适应地计算 Wasserstein 距离的平均值，并通过正定核与其他基线进行比较。

树切片变种的Wasserstein距离