Location
https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php
Document Type
Event
Start Date
24-11-2025 4:00 PM
Description
Classical Optimal Transport (OT) is particularly sensitive to outliers. The existing robust variant, ROBOT, mitigates this through hard truncation, but its rigidity often compromises stability. We propose WROT-r, a unified r-power framework for weighted robust OT that combines rigorous hard-clipping and smooth cost compression through a single parameter r. WROT-r offers a continuous robustness spectrum, enabling adaptive control over how strongly transport costs are down-weighted for outliers. Experiments on synthetic mean estimation and resilient GANs show clear patterns: larger r performs best under weak contamination by preserving more inliers, while smaller r (≈1.5) is more effective under moderate and strong contamination. The extreme r→1 limit (ROBOT) remains consistently unstable. Overall, WROT-r improves robustness across a broad range of noise conditions.
Included in
GRM-0254 Unified Robust Optimal Transport For Outlier-resilient Learning
https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php
Classical Optimal Transport (OT) is particularly sensitive to outliers. The existing robust variant, ROBOT, mitigates this through hard truncation, but its rigidity often compromises stability. We propose WROT-r, a unified r-power framework for weighted robust OT that combines rigorous hard-clipping and smooth cost compression through a single parameter r. WROT-r offers a continuous robustness spectrum, enabling adaptive control over how strongly transport costs are down-weighted for outliers. Experiments on synthetic mean estimation and resilient GANs show clear patterns: larger r performs best under weak contamination by preserving more inliers, while smaller r (≈1.5) is more effective under moderate and strong contamination. The extreme r→1 limit (ROBOT) remains consistently unstable. Overall, WROT-r improves robustness across a broad range of noise conditions.