Feature selection in a dynamic learning system often encounters challenges from sample variation. Incremental selection techniques using rough set theory (RST) address this challenge, to a certain degree, but two issues remain: (1) high redundancy and (2) low efficiency in processing high-dimensional datasets. To address these issues and further improve feature selection, we develop triple nested equivalence class (TNEC) RST for a group incremental approach to feature selection. In particular, we construct a group TNEC to enable universe-reduction learning to help consistently filter out inconsequential samples and incrementally update dependencies. Furthermore, with TNEC-based incremental partitioning, we develop a novel dependency computing technique to reduce redundant features and avoid repeated learning. We conducted experimental verification using 18 dynamic datasets, including ultra-high-dimensional datasets. Tests on five incremental datasets from each of the 18 original datasets validated that the group TNEC approach significantly outperformed the state-of-the-art methods in terms of classification efficiency, selection accuracy, feature significance, and suitability for ultra-high-dimensional data. In both incremental and static feature selection cases, the TNEC approach extended the ability of evolutionary and other types of heuristic learning algorithms in handling sample variations.