CoRoFR: Community Detection of Feature Graph Improves Feature Selection Using Robust Fuzzy Rough Set

In machine learning, features often function as communities in many tasks, especially in medicine. However, existing feature selection methods struggle to mine feature collaborations, which can boost predictive performance. Moreover, they are noise‐sensitive, leading to suboptimal feature selection and accuracy degradation. To meet these challenges, a community detection‐based feature selection algorithm employing a novel robust fuzzy rough sets is put forward. First, a novel robust fuzzy rough sets model is devised, which successfully mitigates the impact of noisy labels and facilitates the construction of a feature graph. Second, drawing inspiration from community detection algorithms, strategies for assessing features within community clusters are formulated by taking into account both the internal architectures of individual feature communities and the interactions among different communities. Finally, experiments on publicly accessible datasets illustrate that the strategies are both efficient and robust across a range of domains.