when: April 4, 2014 from 11 am to noon
where: room B21
speaker: Shyam Sundar
title: The total variation on hypergraphs – Learning on hypergraphs revisited
abstract: In this talk, I present a new learning framework on hypergraphs which, unlike the existing methods, fully uses the hypergraph structure. The key element in our learning framework (for both semi-supervised learning and clustering) is a family of regularizers based on the total variation on hypergraphs. These novel regularizers exactly preserve the hypergraph cut functional and thereby yield unbiased formulations for the learning problems on
hypergraphs. I will also show how we efficiently solve, based on a primal-dual algorithm, the convex optimization problems occurring in the resulting learning problems. Experiments are presented showing better scalability and quality of our approach over the state-of-the-art.
