Feature selection in Angelov-Yager binary classification fuzzy systems in the process of stream data processing

Abstract: An important problem in machine learning is keeping trained models up to date using data streams, as existing solutions are not always capable of updating and processing data incremen-tally. One of the existing solutions with incremental learning support is first-order Angelov-Yager-type fuzzy binary classi-fication systems. The disadvantage of this system is that in in-ference mode the system operates on a full set of features, even if not all features are relevant. This paper proposes a compre-hensive method for calculating feature importance for the spec-ified fuzzy system and presents an experimental study results of feature selection for processing data streams on datasets, the-matically dedicated to spam detection, phishing sites, and net-work connection attacks. A statistically significant difference in accuracy and number of rules was found in favor of using the proposed method for calculating feature importance.

Keywords: binary classification, data streams, feature selection, angelov-yager fuzzy systems