Optimal transport problemS for graphS learning
Title: Optimal transport problemS for graphS learning
Abstract: Originally introduced in order to compare probability distributions, the Optimal Transport (OT) problem has recently received a growing interest in the graph community thanks to its ability of finding the correspondences between the nodes of graphs with disparate structures. In this talk, we consider the problem of comparing and manipulating multiple structured objects such as undirected graphs with features using Optimal Transport. This setting is particularly challenging as the structure, the attributes or the number of nodes of each graph observed may vary.
After introducing the concept of Wasserstein and Gromov-Wasserstein distances we explain how the OT framework can be used on graph datasets in order to develop methods for both supervised and unsupervised learning on graphs. In short we will tackle the following problems: How OT can be used for the classification or the clustering of many graphs ? How OT can be used to find a notion of "average" of many graphs ? Can we use OT to simplify a complicated graph in a meaningful way ? For example can we use the OT framework to construct a interpretable "dictionary" of many graphs ?
Keywords: optimal transport, graph, (online) graphS dictionary learning, graphS clustering/classification.
Web page: https://tvayer.github.io
Talk online : on https://lpensl.my.webex.com/
Room : Webconf3