Surrogate-Assisted and Filter-Based Multiobjective Evolutionary Feature Selection for Deep Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Feature selection (FS) is crucial for deep learning (DL) prediction models but presents significant challenges.
• Existing methods like embedded, filter, and wrapper approaches have limitations, including computational cost and potential precision loss.

Purpose of the Study:

• To propose novel attribute subset evaluation FS methods for deep learning.
• To address the computational expense of wrapper methods and enhance filter methods for DL applications.

Main Methods:

• Development of new wrapper, filter, and hybrid FS methods utilizing multiobjective and many-objective evolutionary algorithms.
• Implementation of a surrogate-assisted approach to mitigate the high computational cost of wrapper-type objective functions.
• Filter-type objective functions based on correlation and an adaptation of the ReliefF algorithm.

Main Results:

• The proposed FS techniques were applied to time series forecasting for air quality and indoor temperature.
• Promising results were achieved, demonstrating improved performance compared to existing FS techniques in the literature.

Conclusions:

• The novel FS methods offer an effective solution for deep learning prediction models.
• The surrogate-assisted approach and adapted ReliefF algorithm contribute to efficient and accurate feature selection.