Feature correlation fusion and feature selection under adaptive neighborhood group approximation space

In real-world scenarios, features often exhibit dynamic interdependence and interaction. While neighborhood rough sets in feature selection have been extensively studied, approaches focusing on searching over feature groups have received less attention. Drawing inspiration from this premise, a feature fusion method grounded in minimum redundancy is proposed to integrate fragmented features. Maximizing Jeffrey divergence constructs a metric function, facilitating the inscription of knowledge granules on the feature groups. This distance function effectively coordinates the importance of feature groups by mapping samples into an adaptive approximation space. Subsequently, traditional uncertainty measures are extended to the neighborhood granules formed by the feature groups. An objective function based on these metrics of neighborhood uncertainty measures is designed to ascertain the importance of feature groups, presenting a novel feature selection algorithm based on this function. Empirical evaluations of the proposed algorithms are conducted using various datasets sourced from the University of California, Irvine (UCI), providing a comprehensive assessment of the efficacy and performance. The experimental results demonstrate the effectiveness of the algorithm.