Discriminative feature selection helps enhance the performance, efficiency, and interpretability of models, rendering it one of the crucial steps in machine learning and pattern recognition. However, when samples from different classes are close to each other in the data, current discriminative feature selection methods often overlook many features with strong predictive power, resulting in suboptimal feature selection outcomes. To address this issue, we propose a novel discriminative feature selection method that utilizes grouping relative comparison from the perspective of relative distance metrics. In this approach, an anchor is initially selected to form a triplet by comparing the similarity gaps between pairs of data. To reduce triplet calculations and integrate overall information, the triplets consist of groups of samples rather than individual ones. Additionally, the determination of loss values shifts from hard judgment to soft judgment to account for data compactness within classes, thereby enhancing feature discrimination. Furthermore, sparse learning is employed in the proposed approach to constrain discriminative features and determine feature weights. Comprehensive experiments are conducted on various benchmark datasets, including face images, biomedical data, and speech letter recognition data, to validate the effectiveness of the proposed method. The source code of the proposed algorithm is available at https://github.com/xhchangsha/FS .