Feature selection constitutes a critical step in data mining workflows. The primary objective of feature selection methodologies lies in identifying an optimal feature subset from the original dataset that maintains high predictive power while preserving essential informational content. This process effectively reduces data dimensionality and improves the performance of downstream machine learning algorithms. Nevertheless, intricate interdependencies within high-dimensional datasets pose substantial challenges to feature selection tasks. In this study, we propose an iterative feature selection framework leveraging symmetric uncertainty to precisely quantify nonlinear feature relationships. Our methodology implements a three-phase approach: (1) initial feature-class correlation assessment using symmetric uncertainty, (2) redundancy quantification through normalized conditional mutual information, and (3) interaction analysis between candidate features and selected subsets via multivariate mutual information. The proposed Symmetric Uncertainty-based Iterative Feature Selection (SUIFS) method was rigorously evaluated against benchmark algorithms across multiple publicly available datasets. Experimental results demonstrate that SUIFS-generated feature subsets achieve superior classification accuracy and enhanced clustering performance compared to conventional approaches.