Indian Journal of Science and Technology
Year: 2018, Volume: 11, Issue: 38, Pages: 1-13
Ghada M. A. Soliman1*, Tarek H. M. Abou-El-Enien1 , E. Emary2 and Motaz M. H. Khorshid1
1 Department of OR and DS, Faculty of Computers and Information, Cairo University, Giza, Egypt; gh.tolan@ fci-cu.edu.eg, [email protected], [email protected]
2 Faculty of Computer Studies, Arab Open University, Cairo, Egypt; [email protected]
*Author for correspondence
Ghada M. A. Soliman,
Department of OR and DS, Faculty of Computers and Information, Cairo University, Giza, Egypt; gh.tolan@ fci-cu.edu.eg
Objectives: A novel hybrid multi-objective swarm system is proposed in order to find the optimal feature subset that achieves data description with minor redundancy as well as keeps classification performance. Methods/Statistical Analysis: The advantages of filter and wrapper approaches characteristics’ are merged simultaneously via different phases of optimization. The proposed system based on Modified Moth Flame Optimization algorithm is assessed against the original algorithm in different experimentations with a single objective based on using MFO and with a multi-objective MFO. The proposed system has been tested over 21 data sets derived from UCI data repository under set of statistical assessment indicators. Findings: The experimental results proved the capability of the hybrid multi-objective MFO-2 to adaptively search the feature space to find optimal feature subset with highest mutual information and achieves maximum classification accuracy and tolerates the problems that are common on both wrapper-based feature selection as well as filter based ones. In addition to, the results obtained from using Random Forests ensemble classifier within the proposed system proved that the performance of hybrid system based on the modified MFO-2 is superior to the hybrid system based on original MFO algorithm in 62% from the data sets, while in case of using the K-Nearest Neighbor classifier we can highlight that the proposed hybrid multi-objective MFO-2 outperforms the other single objective MFO-2 as well as the hybrid multi-objective MFO algorithm in classification performance but also has a comparable ratio of features selected, which confirms that the hybrid multi-objective MFO-2 can select the optimal feature combination with comparable size. Application/Improvements: The novel hybrid multi-objective MFO-2 proved the capability to adaptively search the feature space and its ability to avoid premature convergence caused by falling in local minima.
Keywords: A Hybrid Model, Feature Selection, Filter and Wrapper Approaches, Modified Moth Flame Optimization (MFO2), Multi-Objective
Subscribe now for latest articles and news.