An efficient binary Gradient-based optimizer for feature selection.

Area of Science:

• Machine Learning
• Data Mining
• Optimization

Background:

• Feature selection (FS) is a critical optimization challenge in machine learning and data mining.
• Gradient-based optimizer (GBO) is a population-based metaheuristic inspired by Newton's method, utilizing gradient search rule (GSR) and local escape operator (LEO).

Purpose of the Study:

• To introduce a binary variant of the Gradient-based Optimizer (BGBO) specifically for feature selection tasks.
• To evaluate the efficacy of eight independent GBO variants and two families of transfer functions (S-shaped and V-shaped) for mapping continuous search spaces to discrete feature subsets.

Main Methods:

• Development of a binary Gradient-based Optimizer (BGBO) algorithm.
• Application of eight transfer functions (S-shaped and V-shaped) to adapt the GBO for discrete feature selection.
• Performance evaluation using 18 UCI datasets and 10 high-dimensional datasets.

Main Results:

• The proposed BGBO algorithms demonstrated strong performance in feature selection.
• One specific BGBO variant achieved the best comprehensive performance among the developed algorithms.
• The BGBO approach exhibited superior performance compared to other well-established metaheuristic algorithms for feature selection.

Conclusions:

• The binary Gradient-based Optimizer (BGBO) is an effective metaheuristic for addressing feature selection problems.
• BGBO offers a competitive and high-performing alternative to existing feature selection methods.