Modelo de procesamiento paralelo en arquitecturas heterogéneas para regresiones lineales multivariables

Abstract

The generation of multiple linear regression models demands an exhaustive selection of the return variables that allow obtaining a high level of precision in the prediction tasks. This selection process represents a high algorithmic and computational challenge, due to the fact that it is necessary to obtain and evaluate each of the possible models in order to efficiently select the most accurate one. In this work, a parallel processing model was created to parameterize multivariable linear regression models, using heterogeneous architectures: HMMMR (Heterogeneous Model for Massive Multilinear Regressions). HMMMR was designed to exploit the benefits of parallel computing capabilities of GPUs through the use of data structures and optimized matrix operations to perform batch calculations. The main objective of HMMMR is to make a selection of a subset of predictors that present better results in a linear regression for a given target variable. The implementation of HMMMR shows superiority in the regression calculation time since a more efficient use of the batch processing capacity of the GPU is made. For the datasets evaluated (29332215 and 46626033 regressions with data levels and precipitations of reservoirs located in Colombia) the implementation of HMMMR was up to 9.8 and 5.06 times faster than the implementation in a homogeneous platform. Availability: https://github.com/carojasq/HMMMR.