Design Methodology for Rough Neuro-Fuzzy Classification with Missing Data

One of important methods designed to classify objects with missing feature values are rough neuro-fuzzy classifiers (RNFC). Similarly to neuro-fuzzy systems, they are specific network structures, which can be trained by optimization methods based on gradient descent. However, to the best of our knowledge, there are no publications concerning such way of RNFC designing. In the paper the problems with gradient learning of RNFC are denoted and the suitable solutions are proposed. The influence of missing values level on the learning process and classification quality is examined. The RNFC is compared with the k-NN classifier which is adapted to missing values problem by a "wide imputation" method. All experiments use 10-fold cross validation.