Parallel Implementation of Nearest Feature Line and Rectified Nearest Feature Line Segment Classifiers Using OpenMP

Abstract

Parallelizing computationally expensive classification algorithms, such as the Rectified Nearest Feature Line Segment (RNFLS), remains a task for expert programmers due to the complexity involved in rewriting the application and the required knowledge of the available hardware and tools. A simple parallel implementation of the Nearest Feature Line (NFL) and the RNFLS algorithms using OpenMP over multicore architectures is presented. Both non-parametric classifiers are used in the classification of datasets that contain few samples. The training and testing evaluation technique is used, with a 70–30 ratio respectively and seven datasets from the UCI repository, to verify the speedup and the accuracy of the classifier. Results and experiments derived from the parallel execution of both algorithms, calculating the average of 20 repetitions on a small architecture of 6 physical cores and 12 ones with multi-threading, show that accelerations of up to 21 times can be achieved with NFL and up to 13 times with RNFLS.