A comparative evaluation of nature-inspired algorithms for feature selection problems.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Mining

Background:

• Feature selection is crucial for mitigating challenges in large-scale data, such as irrelevance, noise, and redundancy.
• The curse of dimensionality significantly impacts machine learning model performance and efficiency.
• Nature-inspired algorithms offer novel approaches to complex optimization problems like feature selection.

Purpose of the Study:

• To implement and compare six nature-inspired algorithms for feature selection using a K-nearest neighbour wrapper.
• To evaluate algorithm performance based on accuracy, feature count, fitness, convergence, and computational cost.
• To identify the most effective algorithms for robust feature selection in real-world datasets.

Main Methods:

• Utilized a K-nearest neighbour (KNN) wrapper approach for feature selection.
• Employed six nature-inspired algorithms: Human Learning Optimization (HLO), Poor and Rich Optimization (PRO), Grey Wolf Optimizer (GWO), Harmony Search (HS), and two others.
• Evaluated algorithm performance on six diverse, real-world datasets.

Main Results:

• Human Learning Optimization (HLO), Poor and Rich Optimization (PRO), and Grey Wolf Optimizer (GWO) demonstrated significant efficacy across multiple performance metrics.
• HLO achieved the highest mean fitness, followed closely by PRO and Harmony Search (HS).
• PRO showed particular promise for effective feature selection without sacrificing classification accuracy.

Conclusions:

• Human-inspired algorithms, especially PRO, are highly effective for robust feature selection.
• The study validates the potential of nature-inspired techniques in addressing dimensionality challenges in machine learning.
• Optimized feature subsets can be achieved while maintaining high classification performance.