Optimal transport (OT) is widely used to align distributions and transfer labels between unpaired source and target data. But when the composition of classes shifts between the two domains, the method can systematically under-predict rare classes, even when the overall alignment looks reasonable.
The new paper, posted on arXiv, argues that this problem stems from a confusion between two distinct factors: the marginal capacity of each class and the geometric structure of the data. By disentangling these, the authors aim to show how much of the under-prediction is due to class imbalance rather than poor alignment.
Because the abstract is truncated, the full proposed remedy is not available in the source. Still, the framing offers a useful diagnostic lens for practitioners using OT-based label transfer in settings with skewed class distributions.