A1291
Title: Global optimization approach for multiblock methods: Applications to large-scale data
Authors: Alba Martinez-Ruiz - Universidad Andres Bello (Chile) [presenting]
Abstract: Global regularized generalized canonical correlation analysis (GRGCCA) is a multiblock method that finds a set of components for a set of blocks of variables, maximizing a function of the sum of the covariances between components. Unlike the traditional sequential approach, all components for each block are obtained simultaneously by solving a single optimization problem. This procedure avoids stage-by-stage component extraction, the consecutive solution of as many optimization problems as components to be extracted, and the deflation of the matrices and residuals. However, the search algorithm slows down when analyzing large datasets. Several approaches to scale the search algorithm for big high-dimensional matrices are explored. One strategy involves implementing randomized singular value decomposition, a low-rank approximation method. MPI-based parallel implementations are investigated to evaluate speedups on synthetic datasets. The results show that parallel implementations achieved speedups of up to 75x compared to the serial version when executed on 16 nodes. These implementations are particularly useful for biomedical research.
