Modeling methods for deep fuzzy inference systems based on feature selection

Abstract

A bottom-up, layer-by-layer feature selection modeling method is proposed for deep fuzzy inference systems. By selecting features in the input space, redundant and interfering features in each layer's input space can be removed. The selected features have been appropriately ordered to avoid the randomness of features within the feature sliding window. Based on fuzzy rough set theory, the matching degree between the fuzzy granules associated with the conditional feature set and those associated with the decision feature is used as a likelihood function. The λ-AIC criterion with a penalty term have been redefined as an evaluation index for features. The obtained model is compared to four baseline models and an existing deep fuzzy rule classification system. The results show that the feature selection-based deep fuzzy inference system effectively improves data classification performance and efficiency. Compared to four rough set-based feature selection methods based on different measures, the method proposed is more convincing in terms of accuracy and the number of features selected. Compared to the modeling method where input space features are randomly shuffled, the feature sorting modeling method performs better and avoids the instability in model accuracy caused by randomness.