Scalable feature subset selection for big data using parallel hybrid evolutionary algorithm based wrapper under...

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Sequential wrapper-based feature subset selection (FSS) algorithms struggle with scalability and performance on large datasets.
• Existing methods often fail to meet the demands of big data analytics.

Purpose of the Study:

• To develop scalable and efficient parallel and distributed hybrid evolutionary algorithms (EAs) for feature subset selection on big datasets.
• To enhance the search capabilities and prevent premature convergence in evolutionary algorithms.

Main Methods:

• Proposed two hybrid EAs: Parallel Binary Differential Evolution and Threshold Accepting (PB-DETA) and its variant Parallel Binary Threshold Accepting and Differential Evolution (PB-TADE) using Apache Spark.
• Parallelized state-of-the-art algorithms: adaptive DE (ADE) and permutation based DE (DE-FS ^PM).
• Utilized logistic regression (LR) for fitness evaluation, specifically the area under the receiver operator characteristic curve (AUC).

Main Results:

• PB-TADE achieved statistically significant results compared to other tested algorithms.
• All proposed parallel algorithms demonstrated the repeatability property.
• The parallel model achieved a speedup of 2.2-2.9x.
• Identified feature subsets with high AUC and minimal cardinality.

Conclusions:

• The developed parallel and distributed hybrid EAs offer a scalable solution for feature subset selection in big data scenarios.
• PB-TADE is a highly effective algorithm for optimizing feature selection, outperforming existing methods.
• The parallel approach significantly improves computational efficiency while maintaining accuracy.