We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Statistics > Machine Learning

Title: Outlier-Robust Optimal Transport: Duality, Structure, and Statistical Analysis

Abstract: The Wasserstein distance, rooted in optimal transport (OT) theory, is a popular discrepancy measure between probability distributions with various applications to statistics and machine learning. Despite their rich structure and demonstrated utility, Wasserstein distances are sensitive to outliers in the considered distributions, which hinders applicability in practice. We propose a new outlier-robust Wasserstein distance $\mathsf{W}_p^\varepsilon$ which allows for $\varepsilon$ outlier mass to be removed from each contaminated distribution. Under standard moment assumptions, $\mathsf{W}_p^\varepsilon$ is shown to be minimax optimal for robust estimation under the Huber $\varepsilon$-contamination model. Our formulation of this robust distance amounts to a highly regular optimization problem that lends itself better for analysis compared to previously considered frameworks. Leveraging this, we conduct a thorough theoretical study of $\mathsf{W}_p^\varepsilon$, encompassing robustness guarantees, characterization of optimal perturbations, regularity, duality, and statistical estimation. In particular, by decoupling the optimization variables, we arrive at a simple dual form for $\mathsf{W}_p^\varepsilon$ that can be implemented via an elementary modification to standard, duality-based OT solvers. We illustrate the virtues of our framework via applications to generative modeling with contaminated datasets.
Comments: updated to match AISTATS publication
Subjects: Machine Learning (stat.ML); Machine Learning (cs.LG)
Cite as: arXiv:2111.01361 [stat.ML]
  (or arXiv:2111.01361v3 [stat.ML] for this version)

Submission history

From: Sloan Nietert [view email]
[v1] Tue, 2 Nov 2021 04:05:45 GMT (4697kb,D)
[v2] Fri, 5 Nov 2021 18:31:46 GMT (4697kb,D)
[v3] Sun, 27 Feb 2022 01:02:32 GMT (3060kb,D)

Link back to: arXiv, form interface, contact.