Choquet-based fuzzy rough sets (Q2671762)

This paper introduces a generalization of fuzzy rough sets based on Ordered Weighted Average (OWA) operators; namely, fuzzy rough sets based on the well-known Choquet integral (Choquet-based fuzzy rough sets). This generalization preserves some desirable theoretical properties (relation monotonicity, set monotonicity, duality). It also allows an integration of outlier detection algorithms, enhancing robustness of machine learning algorithms based on fuzzy rough sets. OWA operators can be understood as Choquet integrals w.r.t. symmetric measures (i.e., measures fulfilling \(\mu(A) = \mu(B)\) whenever \(|A| = |B|\)). Therefore, the authors look for examples of non-symmetric measures for their Choquet-based fuzzy rough sets. These measures can incorporate an outlier score or unreliability/inaccuracy of observations. The authors provide several examples of such non-symmetric measures (fuzzy removal, weighted ordered weighted average). These measures have a clear intuitive motivation in terms of vague quantification. Based on the theoretical results, the authors provide an experimental evaluation of the proposed methods for the task of classification.