Title: Incremental SVD for some numerical aspects of multiblock redundancy analysis and big data streams
Authors: Alba Martinez-Ruiz - Universidad Diego Portales (Chile) [presenting]
Carlo Lauro - University of Naples - "Federico II" (Italy)
Abstract: Based on the incremental SVD algorithm, a new procedure is proposed to solve the decomposition problem found by multiblock redundancy analysis when analyzing streaming data, i.e. data that is generated continuously. The redundancy procedure involves the SVD of a square and symmetric matrix of $q\times q$, where $q$ is the number of variables in the endogenous block of variables. If $q$ is large, the factorization is a time- and resource-consuming task, much more if the matrix is continuously updated in real-time. A good strategy is analyzing the data in small sets that are continually updated. To preserve the column-wise formulation of the incremental SVD algorithm, we derived the column-wise variant of the redundancy method and implemented an incremental approach for the procedure. Numerical experiments are reported to illustrate the accuracy and performance of the incremental solution for analyzing streaming multiblock data. In addition, we report results for examining how the incremental SVD algorithm approximates singular vectors when varying a forgetting factor and the number of significant singular vectors kept at each iteration. The results provide evidence about the suitability of the new approach for the analysis of large streaming data.