Fuzzy Rule-based Outlier Detector

The problem of detecting outliers in data is a widely discussed issue. The sources of outliers vary and can come from system errors, or human mistakes. Due to the constantly increasing number of data for analysis, an effective tool for detecting outliers should be proposed. Therefore, in this study we present a method based on the use of the fuzzy three-sigma rule to detect outliers. The novelty of the described method is the use of the properties of fuzzy sets to replace the properties of the analyzed data with common statistical semantics. Due to the untypical approach consisting in an independent analysis of each dimension of the data set, a universal method was obtained, which operates regardless of the specificity of the analyzed data. Moreover, the appropriate aggregation of the membership degrees to the descriptors describing the type and strength of deviation from the norm makes it possible to look at the analyzed data from various angles. The high performance of the proposed novel approach was confirmed in numerical experiments.