Estratégias de paralelização para o algoritmo Feature Space Partition

Abstract

This dissertation presents the parallelization of the supervised learning algorithm Feature Space Partition (FSP), originally implemented sequentially in the Python language. Although FSP has proven to be efficient in small datasets, it has presented a high execution time when used in largers amount of data. The parallelization process investigates more than one parallelization approach, including parallel programming on CPU and GPU, maintaining the original sequence of steps of the sequential version of the algorithm, in order to evaluate which is the most appropriate strategy for paralleli- zation, but without losing its accuracy. One of the premises of the project is to try to use parallelization resources and libraries already existing in the vast and heterogeneous environment of the Python language. At the end of the experiment, the best scenario exe- cuted was the multiprocess CPU parallelization, which presented a 36.44% improvement in performance, when compared to its initial implementation, without decreasing its accuracy.