| Zeit: | 5. November 2025, 09:45 Uhr |
|---|---|
| Universitätstraße 32 Room 101, Campus Vaihingen of the University of Stuttgart | |
| Download als iCal: |
|
We are pleased to announce our upcoming ELLIS Unit Stuttgart Distinguished Lecture Series talk by Florence d'Alché-Buc (Télécom Paris)! Looking forward to seeing you all there! No registration necessary.
The talk will be followed by an informal reception during which finger food and drinks will be provided.
Professor d'Alché-Buc will also be available for meetings. If you are interested in scheduling a meeting, please email ellis-office@uni-stuttgart.de
Title: Leveraging Optimal Transport for Graph-Valued Regression
Abstract: Driven by applications in molecular identification, we address a regression problem where the target variable is a graph. The structured, discrete nature of graphs—along with their permutation invariance and variable size—poses significant challenges for loss function design. We briefly emphasize two working directions, namely surrogate regression in Hilbert Spaces and end-to-end approaches. To implement the latter, we leverage Optimal Transport theory, introducing three loss functions based on Gromov-Wasserstein distances: the asymmetric Fused Gromov-Wasserstein (FGW) loss, the Partially-Masked FGW (PMFGW) loss, and its Sinkhorn-approximated variant. Each is crafted to overcome specific limitations of the previous one. We develop end-to-end solutions by minimizing these losses using both nonparametric (kernel-based) and parametric (transformer-based) models. Finally, we present numerical results and outline directions for future work.
Bio: Florence d'Alché-Buc is professor at Télécom Paris (Institute Polytechnique de Paris), in the Data, Signal and Image Department. She used to be the head of the Labex (Excellence Laboratories) in Science and Technology of Information of University Paris-Saclay between 2017 and 2019 and was program co-chair of NeurIPS in 2019. Ellis Fellow, she currently coordinates the Ellis PhD & Postdoc program and participates to the Ellis board. Her researches span diverse areas of Machine Learning, including structured and functional regression, kernel methods, optimal transport with a strong interest in interpretable and robust shallow and deep learning.
