Problem : Regularized Optimization for Inducing Block-diagonal Matrices

Zhouchen Lin,Huan Xu,Shuicheng Yan
2015-01-01
Abstract:Many machine learning problems are parameterized by matrices with grouping effect. In the ideal cases, such matrices will naturally have a block-diagonal structure after permutation of rows and columns. However, real data are often produced with noise or perturbations. To learn robust parameter matrices, it is tempting to pose constraints or regularization terms in the optimization formulations to encourage for block-diagonal structures. The question posed here is whether there exists a generic way to induce block-diagonal structures for matrix-variable optimization problems.
What problem does this paper attempt to address?