Parallel Multiobjective Feature Selection for Binary Classificatio

Abstract

Feature selection has become a prominent step for many research studies as available data increases continuously with the advances in technology. The objective of feature selection is two-fold: minimizing the number of features and maximizing learning performance. Therefore, it requires a multi-objective optimization. In this study, we utilize the multi-core nature of a regular PC in the feature selection domain. For this purpose, we build three models that exploit the parallel processing capability of a modern CPU. We execute the feature selection task on a single processor in the first model as a baseline. In other models, we execute the feature selection task in four cores of the CPU, in parallel. Specifically, in the second model, we decrease the population size per processor and explore whether we can achieve comparable solution sets in less amount of time. The third model preserves the population size and explores a more extensive search space. We compare the results of these models in terms of accuracy, number of features and execution time. Experiment results show that parallel processing in the feature selection domain leads to faster execution and better feature subsets.