An efficient parallel implementation for training supervised optimum-path forest classifiers

摘要

In this work, we propose and analyze parallel training algorithms for the Optimum-Path Forest (OPF) classifier. We start with a naive parallelization approach where, following traditional sequential training that considers the supervised OPF, a priority queue is used to store the best samples at each learning iteration. The proposed approach replaces the priority queue with an array and a linear search aiming at using a parallel-friendly data structure. We show that this approach leads to less competition among threads, thus yielding a more temporal and spatial locality. Additionally, we show how the use of vectorization in distance calculations affects the overall speedup and also provide directions on the situations one can benefit from that. The experiments are carried out on five public datasets with a different number of samples and features on architectures with distinct levels of parallelism. On average, the proposed approach provides speedups of up to 11.8 x and 26 x in a 24-core Intel and 64-core AMD processors, respectively. (C) 2019 Elsevier B.V. All rights reserved.