TIEOD: Three-way concept-based information entropy for outlier detection

Abstract

Outlier detection is an attractive research area in data mining, which is intended to find out the few data objects that are abnormal to the normal data set. Formal concept analysis is an efficacious mathematical tool to perform data analysis and processing. Three-way concepts contain both information of co-having and co-not-having, and reflect the correlation among objects (attributes). Information entropy reflects the degree of uncertainty of the system. Information entropy-based outlier detection methods have been widely studied and have shown excellent performance, but most current information entropy-based methods contain parameters, which leads to detection results are sensitive to parameters settings and taking longer detection times. Aiming at this deficiency, this paper constructs a three-way concept-based information entropy outlier detection method. Firstly, the information entropy of the formal context is defined by utilizing three-way granular concepts, and then the relative entropy of each object is defined. According to it, the relative cardinality-based outlier degree of each object is given, and then the outlier factor of the object is defined by combining with the relative entropy. Then the three-way concept information entropy-based outlier factor is presented and the associated algorithm is proposed. Finally, the effectiveness and efficiency of the proposed algorithm is verified on a public dataset.