|
|
Detailed schedule
Click on a link for more details
Show all the abstracts
Show all the abstracts
Thursday 11:00:00 Timetabling in education and sport Room 126 - Chair: G. Vanden Berghe
Thursday 11:00:00 Transportation management Room 130 - Chair: F. Semet
Thursday 11:00:00 Networks Room 138 - Chair: B. Fortz
Thursday 11:00:00 Nonconvex optimization 1 Room 035 - Chair: F. Bach
- On the best low multilinear rank approximation of higher-order tensors
Mariya Ishteva (Université catholique de Louvain, Department of Mathematical Engineering) Co-authors: PA. Absil, S. Van Huffel, L. De Lathauwer
- Regression on fixed-rank positive semidefinite matrices: a geometric approach
Gilles Meyer (University of Liège) Co-authors: Gilles Meyer, Silvère Bonnabel and Rodolphe Sepulchre Abstract: In this paper, we adopt a geometric viewpoint to tackle the problem of learning a regression model whose parameter is a fixed-rank positive semidefinite (PSD) matrix.
An important instance of that problem is the learning of a distance function parameterized by a fixed-rank PSD matrix. This task is a central issue for many machine learning applications where a data-specific distance has to be constructed, or where an existing distance needs to be improved based on additional side information.
Learning low-rank matrices is a typical solution to reduce the computational cost of subsequent algorithms. Indeed, the complexity generally decreases from O(d^3) to O(d*r^2) where the approximation rank r is generally much smaller than the problem size d.
Whereas efficient convex formulations exist in the full-rank case, the problem is no longer convex as soon as the rank constraint is introduced. Nevertheless, the set of rank-r PSD matrices has a rich Riemannian geometry that can be exploited for algorithmic purposes.
We discuss the choice of two particular geometries of fixed-rank PSD matrices and we derive the corresponding gradient descent algorithms. In contrast to previous contributions in the literature, the range space of the matrix is free to evolve during the optimization and the resulting algorithms enjoy important invariance properties.
We apply the two proposed algorithms to the distance learning problem. The good performance of the algorithms is illustrated on several well-known classification and clustering benchmarks.
- Generalized Power Method for Sparse Principal Component Analysis
Rodolphe Sepulchre (Université de Liège) Co-authors: Michel Journée, Peter Richtarik, Yurii Nesterov
- A pooling approach for the feed mixing problem
Jannes Verstichel (KaHo Sint-Lieven) Co-authors: G. Vanden Berghe, H. Callens
Thursday 14:00:00 Constraint programming models 1 Room 126 - Chair: Y. Deville
Thursday 14:00:00 Vehicle routing Room 130 - Chair: S. Limbourg
Thursday 14:00:00 Combinatorial optimization and IP applications Room 138 - Chair: Q. Louveaux
Thursday 14:00:00 Nonconvex Optimization 2 Room 035 - Chair: R. Sepulchre
Thursday 16:10:00 Constraint programming models 2 Room 126 - Chair: P. Schaus
Thursday 16:10:00 Performance modeling Room 130 - Chair: G. Janssens
Thursday 16:10:00 Scheduling Room 138 - Chair: K. Sorensen
Thursday 16:10:00 Planning under uncertainty Room 035 - Chair: R. Leus
Friday 09:00:00 Metaheuristics Room 126 - Chair: J. Teghem
Friday 09:25:00 Production and distribution (9:25) Room 130 - Chair: Y. Arda
Friday 09:00:00 Multiple criteria Room 138 - Chair: R. Bisdorff
Friday 09:25:00 Stochastic models (9:25) Room 035 - Chair: L. Esch
Friday 11:00:00 Constraint programming and Supply Chain Management Room 126 - Chair: Y. Deville
Friday 11:00:00 OR in health management Room 130 - Chair: P. De Causmaecker
Friday 11:00:00 Rankings and importance indices Room 138 - Chair: JL. Marichal
Friday 11:00:00 Queueing Room 035 - Chair: S. Wittevrongel
Friday 15:10:00 Optimization software Room 126 - Chair: E. Loute
Friday 15:10:00 Integrated operations planning Room 130 - Chair: B. Raa
Friday 15:10:00 Cycles in graphs Room 138 - Chair: F. Spieksma
|
|