Abstract:

Relational data appears in many applications, and are usually represented as graphs. An important property of such graphs is their community structure, that is, the partition of vertices such that a large proportion of the edges appear within the parts, also known as clusters. However, this is often a simplified representation of the data, and representations such as hypergraphs are required to capture more complex relations.

The modularity function is at the heart of several leading graph clustering algorithms. In this talk, we propose a generalization of the modularity function for relational data represented as hypergraphs. We also present the theory required to develop hypergraph clustering algorithms. We conclude by presenting some preliminary results obtained with simple algorithms, and propose areas for future work.

Speaker Bio:

F. T. holds a B.Sc. degree in applied mathematics and computer science from the University of Ottawa, a M.Sc. in telecommunications from INRS and a PhD. in electrical engineering from McGill University. He has been employed by CSE 1996 during which he was involved in the creation of the data science team as well as the research group now known as TIMC. He also holds an adjunct professorial position at the Department of Mathematics and Statistics of the University of Ottawa since 1997. His current interests include relational-data mining and deep learning.