Thesis defense of Cédric Vincent-Cuaz

Published on March 27, 2023 Updated on March 27, 2023

on the March 29, 2023

Inria Sophia Antipolis

Cédric Vincent-Cuaz, 3IA Côte d'Azur PhD student, invites you to his thesis defence on March 29th 2023.

Thesis title: Optimal Transport for Graph Representation Learning

A key challenge in Machine Learning (ML) is to design models able to learn efficiently from collections of graphs, potentially endowed with node features. Graph Representation Learning (GRL) aims to encode these two sources of heterogeneity into a vectorial graph embedding easing downstream tasks. In this field, Optimal Transport (OT) has been successful in providing meaningful comparison between graphs seen as discrete probability distributions. This thesis focuses on GRL through the lens of OT.

First presented contributions made amenable Dictionary Learning (DL) to graph dataset analysis. However, the conservation of mass at the core of OT, makes the learned representations considerably sensitive to the relative importance of the nodes assigned to them. To overcome these limitations, we introduced a new OT-based divergence between graphs, leading to a paradigm shift from the quantification of graph isomorphism to subgraph isomorphism. Moreover, the corresponding estimation of a barycenter naturally induces novel DL, where graphs are embedded as reweighed subgraphs of a single graph atom. Both DL methods consistently improve the performance of the existing literature on various unsupervised GRL tasks, such as clustering, denoising, and completion.

Finally we focused on supervised GRL, by establishing a first link between OT on graphs, that captures their global properties, and Graph Neural Networks (GNN) known to produce discriminative node embeddings. To this end, we proposed a novel relational pooling layer that places distances to some learnable graph templates at the core of the graph representation. Interestingly, this approach outperforms diverse kernels and GNN approaches in terms of expressivity and generalization abilities.

Supervisors: Rémi Flamary, Marco Corneli

Jury members: Gabriel Peyré, Florence d'Alché-buc, Laetitia Chapel, Marco Gori, Hongteng Xu

In-person meeting: Inria Sophia-Antipolis, Khan amphitheatre. March 29 at 2 pm.

Online meeting: The thesis defense will be broadcasted as a live on the Youtube channel -