Paper at ICML 2025

The following paper was accepted at this year International Conference on Machine Learning (ICML)

  • J. Naram, F. Hellström, Z. Wang, R. Jörnsten, and G. Durisi, “Theoretical performance guarantees for partial domain adaptation via partial optimal transport,” in Proc. Int. Conf. Machine Learning (ICML), Vancouver, Canada, July 2025. [],

This work addresses the problem of partial domain adaptation, where the target label space is a subset of the source label space, by deriving generalization bounds using partial optimal transport. These bounds justify the use of partial Wasserstein distance for domain alignment and lead to theoretically motivated and explicit expressions for the empirical source loss weights to be used to avoid the problem of negative transfer. Finally, the bounds suggesta new algorithm, WARMPOT, which demonstrates competitive performance compared to state of the art.