High-order interaction feature selection for classification learning: A robust knowledge metric perspective

•A robust fuzzy granularity space for the statistical distribution characteristics of different features is constructed by using a soft fuzzy rough set.•A novel knowledge metric approach, i.e., robust fuzzy uncertainty measures (RFUMs), is devised which achieves metric fitness and robustness on different granular structures.•This is the first exploration of high-order interaction-dependent features in the fuzzy rough set.•A high-order interaction feature selection with RFUMs (HIFS-RFUMs) algorithm is designed to perceive and capture interactive features in data containing noise, fuzziness and uncertainty.•The experiment demonstrated that HIFS-RFUMs is effective. The robustness of the RFUMs metric and the effectiveness of HIFS-RFUMs in mining features with high-order interaction dependencies are verified by ablation experiments on the HIFS-FUMs and R2-RFUMs algorithms, respectively.