Differential evolution algorithm with cosine similarity-based individual reduction and symmetric uncertainty-based attribute recovery for feature selection

In recent years, with the increase in data scaling, effectively handling high-dimensional datasets has become a focal point of attention. Feature selection (FS), a method for dealing with datasets containing features, has emerged as a crucial technique in fields such as machine learning and data mining with the objective of selecting features that contain richer information while eliminating redundancy. Due to their remarkable performance in global search, evolutionary computation techniques hold substantial potential in the application of FS. However, many existing FS methods overlook the relationships between features. In the context of classification problems, this study presents a novel wrapper FS algorithm based on the differential evolution algorithm. The proposed method reduces redundant features among individuals based on cosine similarity and selectively and recovers certain features in the crossover phase according to the uncertainty similarity. In addition, a probabilistic-based initialization method is designed. The proposed algorithm significantly outperforms five other algorithms in terms of classification error rates over 18 experimental datasets. The experimental results demonstrate a significant enhancement in the performance of the proposed algorithm attributed to these two components. • A new initialization strategy analyzes feature relationships in SCDE’s early stage. • This strategy reduces redundant features using probability-based feature selection. • A novel crossover removes duplicates and restores unselected features with uncertainty. • SCDE is tested on 18 complex and comprehensive datasets for performance validation. • SCDE outperforms other advanced methods based on graphical and tabular results.