Joint uncertainty model and metric for robust feature selection: A bi-level distribution consideration and feature evaluation approach

Excavating discriminative features in uncertain data environment is essential and challenging. Randomness, fuzziness, and inconsistency are the main characteristics of uncertain data. For the uncertainty in data, fuzzy rough set (FRS) is considered as a powerful analytical model. However, the classical FRS model is susceptible to noisy information. Moreover, fewer existing works utilize robust evaluation metrics to explore complex feature correlations in high-dimensional spaces, which leads to the loss of some important discriminative information and the selection results are not robust. Therefore, this paper proposes a joint uncertainty model and metric for robust feature selection method with bi-level distribution consideration and feature evaluation (JBRFS). The robust FRS model and the robust uncertainty metric are firstly constructed by twofold consideration of data distribution characteristics (i.e., the density distribution of heterogeneous samples and the statistical distribution of different features). Then, the bi-level feature evaluation (i.e., one-order individual feature relevance and two-order pairwise feature correlation) is devised by jointing the robust uncertainty model and metric. Finally, a novel feature selection strategy of max-fuzzy relative dependency, max-significance and interaction is proposed. The experimental results show that JBRFS can select features with discriminative ability, and it is verified to have significant advantages in most cases by nine comparison algorithms and the ablation experiment.