When: October 4, 2013 from 11AM to noon.
Where: room B21
Speaker: Thomas Ricatte
Title: A Spectral framework for Un-directed Hypergraphs
Abstract: Undirected graphs are a natural way to formalize pairwise homophilic relationships between objects. Spectral methods based on graph Laplacians and graph Kernels allow us to leverage these relations in many efficient applications (semi-supervised learning, clustering, …). We propose a natural extension of this framework to represent interactions between groups of collaborating objects through the notion of undirected hypergraphs. We extend the notion of graph Laplacian and graph Kernel for this newly defined class and discuss the related properties (connectivity, resistance distance, links with signed graphs). We also see that the newly defined space of hypergraph Laplacians and hypergraph Kernels is a natural convex relaxation of the space of graph Kernels. Thus in many cases, combining graph kernels will lead to a direct interpretation in terms of undirected hypergraph.
