Multi-variant differential evolution algorithm for feature selection | JoVE Visualize

Area of Science:

Computational Intelligence
Machine Learning
Optimization Algorithms

Background:

Feature selection is crucial for improving classification accuracy and reducing computational complexity by removing irrelevant or redundant data.
Existing optimization methods often require careful parameter tuning, limiting their efficiency in real-world applications.

Purpose of the Study:

To introduce a new population-based stochastic search technique, the multi-variant differential evolution (MVDE) algorithm.
To develop a novel feature selection algorithm integrating MVDE with artificial neural networks (ANNs) for simultaneous optimization of feature sets, ANN structure, and weights.
To evaluate the performance of the proposed MVDE-based feature selection method against established optimization techniques on benchmark real-world problems.

Main Methods:

The proposed MVDE algorithm incorporates a self-adaptive scaling factor based on cosine and logistic distributions, combined with binary mapping and an adaptive crossover operator.
A multi-mutation crossover process is employed, managing both greedy and less-greedy variants through sequentially evolutionary phases.
The MVDE method is applied to feature selection in conjunction with an artificial neural network, optimizing feature subsets, ANN architecture, and weights concurrently.

Main Results:

The MVDE algorithm demonstrated competitive performance compared to four popular optimization methods on fifteen real-world problems from the UCI repository.
The integrated MVDE-based feature selection algorithm successfully identified optimal feature combinations, leading to accelerated classification accuracy.
The proposed method effectively optimized both the structure and weights of the Artificial Neural Network (ANN) simultaneously.

Conclusions:

The multi-variant differential evolution (MVDE) algorithm offers a robust and efficient approach to optimization problems.
The novel feature selection algorithm leveraging MVDE and ANNs significantly enhances classification performance and optimizes feature subsets.
The experimental results validate the effectiveness of the proposed approach for improving classification accuracy and achieving optimal feature selection.