Feature subset selection (FSS) for classification is inherently a bi-objective optimization problem, where the task is to obtain a feature subset which yields the maximum possible area under the receiver operator characteristic curve (AUC) with minimum cardinality of the feature subset. In todays world, a humungous amount of data is generated in all activities of humans. To mine such voluminous data, which is often high-dimensional, there is a need to develop parallel and scalable frameworks. In the first-of-its-kind study, we propose and develop an iterative MapReduce-based framework for bi-objective evolutionary algorithms (EAs) based wrappers under Apache spark with the migration strategy. In order to accomplish this, we parallelized the non-dominated sorting based algorithms namely non dominated sorting algorithm (NSGA-II), and non-dominated sorting particle swarm optimization (NSPSO), also the decomposition-based algorithm, namely the multi-objective evolutionary algorithm based on decomposition (MOEA-D), and named them P-NSGA-II-IS, P-NSPSO-IS, P-MOEA-D-IS, respectively. We proposed a modified MOEA-D by incorporating the non-dominated sorting principle while parallelizing it. Throughout the study, AUC is computed by logistic regression (LR). We test the effectiveness of the proposed methodology on various datasets. It is noteworthy that the P-NSGA-II turns out to be statistically significant by being in the top 2 positions on most datasets. We also reported the empirical attainment plots, speed up analysis, and mean AUC obtained by the most repeated feature subset and the least cardinal feature subset with the highest AUC, and diversity analysis using hypervolume.
View on arXiv