Large-scale multi-label ensemble learning on Spark

摘要

Multi-label learning is a challenging problem which has received growing attention in the research community over the last years. Hence, there is a growing demand of effective and scalable multi-label learning methods for larger datasets both in terms of number of instances and numbers of output labels. The use of ensemble classifiers is a popular approach for improving multi-label model accuracy, especially for datasets with high-dimensional label spaces. However, the increasing computational complexity of the algorithms in such ever-growing high-dimensional label spaces, requires new approaches to manage data effectively and efficiently in distributed computing environments. Spark is a framework based on MapReduce, a distributed programming model that offers a robust paradigm to handle large-scale datasets in a cluster of nodes. This paper focuses on multi-label ensembles and proposes a number of implementations through the use of parallel and distributed computing using Spark. Additionally, five different implementations are proposed and the impact on the performance of the ensemble is analyzed. The experimental study shows the benefits of using distributed implementations over the traditional single-node single-thread execution, in terms of performance over multiple metrics as well as significant speedup tested on 29 benchmark datasets.