SC²S Colloquium - April 28, 2017
|Date:||April 28, 2017|
|Time:||1:00 pm, s.t.|
Michael Lettrich: Parallel Incomplete Cholesky Decomposition for Data Mining with Sparse Grids
Approaches how to use Sparse Grids for higher dimensional data mining scenarios have been around for several years by now. An approach, which allows multi-class classification, clustering and dimensionality reduction uses Sparse Grid based density estimation and relies on a variational formulation. Discretization of this approach results in solving a large linear system of equations with a often sparse system matrix. As this matrix only relies on the current grid, it can be decomposed in a preprocessing step and used on arbitrary datasets. Previous work studied using a full Cholesky Decomposition which allows modifying existing decompositions without the need to recompute the entire decomposition. During the project we studied the effects on accuracy and performance when replacing the full Cholesky Decomposition with an iterative incomplete Cholesky Decomposition which in most cases promise a significant boost in performance while sacrificing only an insignificant amount of accuracy.