<p>Feature selection plays a vital role in data mining and machine learning. However, real-world data are often complex and contain numerous redundant features, making the construction of an optimal feature subset a highly challenging problem. To address this issue, this paper proposes a feature selection framework based on relative fuzzy rough sets (RFRS) and granularity consistency integrated with combination mutual information. First, a fuzzy similarity relation based on k-nearest neighbor relative distance is constructed to capture local neighborhood relationships in fuzzy similarity measurement. Building on this, a series of relative fuzzy combination mutual information measures are established to characterize feature redundancy and quantify the contribution of features to decision uncertainty. Then, granularity consistency is integrated with combined entropy to form a new feature ranking criterion, leading to a novel feature evaluation metric. Finally, a forward search feature selection algorithm (FSCMG) is designed, driven by maximum relevance, minimum redundancy, and granularity consistency gain. Comparative experiments with eight existing algorithms on 14 public datasets demonstrate the effectiveness and superiority of the proposed approach.</p>