Classical matrix factorization techniques such as the LU, QR, EVD, and SVD or their variants have been used with great success in various data analytic applications. These include information retrieval, text mining, bioinformatics, computer graphics, computer vision, product recommendations, etc. In the past few years, internet applications have thrown some new challenges to the numerical linear algebraists — unprecedented features and problems inherent in internet data have rendered traditional matrix factorization techniques less effective. Some of the new issues that arise in internet data mining include: (i) prohibitively large data size — internet data sets are often much too large to allow multiple random accesses; (ii) massively incomplete data — a significant proportion of the data may be missing; (iii) novel structures in data — most importantly, datasets whose underlying structure cannot be adequately unraveled by LU, QR, EVD, or SVD have become increasingly common, not just in internet applications but also in other scientific and engineering fields. In the last few years, several new matrix factorizations have been proposed to deal with these issues. Some notables ones include (1) Nonnegative Matrix Factorization; (2) Maximum Margin Matrix Factorization; (3) Matrix Subspace Factorization; (4) Sparse Overcomplete Factorization (Compressed Sensing). The key difference between these and the classical matrix factorizations is that they are not 'rank-revealing' in the traditional sense but instead they 'reveal' other properties of the structure under consideration. This minisymposium will focus on the development of these and other novel matrix computational tools with a view towards internet data mining applications.
This minisymposium is to be held as a part of the 6th International Congress on Industrial and Applied Mathematics (ICIAM 2007), which takes place quadrennially and is the biggest event in applied mathematics. ICIAM 2007 will be held at the Swiss Federal Institute of Technology (ETH) in Zürich, Switzerland from July 16–20, 2007.
Anirban Dasgupta, Gene Golub, Lek-Heng Lim, Michael Mahoney
The ICIAM program is now online. Our minisymposium will take place on Thursday, July 19.
First Session: Thursday, July 19, 11:15am–1:15pm | ||
Art Owen | Stanford University | An Empirical Meta-Analysis for Combining Dependent Hypothesis Tests |
Nathan Srebro | University of Chicago | Maximum Margin Matrix Factorization |
Petros Drineas | Rensselaer Polytechnic Institute | Sampling-based Algorithms for General Regression Problems |
Joel Tropp | California Institute of Technology | Kernel learning with Bregman matrix divergences |
Second Session: Thursday, July 19, 3:45pm–5:45pm |
||
Ravi Kannan | Yale University | Random Sampling in Large Matrices |
Ming Gu | University of California at Berkeley | New Fast Algorithms for Semi-Definite Programming |
Yuan Yao | Stanford University | Combinatorial Laplacian and Rank Aggregation |
Michael Saunders & Sou-Cheng Choi | Stanford University | PageRank by Basis Pursuit |
For further information on this meeting, please email Lek-Heng Lim at lekheng(at)stanford.edu