Abstract:

The existing researches on outlier data mainly focus on the outlier detection. In order to completely analyze the origin, classification, meaning, behavior characteristics and outlying trend of outlier data, some concepts such as the nearest outlying neighbor, the atomic outlier class and the outlying mutation class are defined and the approaches to outlier clustering and outlying trend analyses are proposed based on the existing outlier mining techniques as well as a series of concepts and their searching algorithms including the outlying reduction and the key attribute subspace. Furthermore, an integrated framework of characteristic description and extended knowledge discovery of outlier data set is constructed, whose validity in practical applications is finally verified by the outlier analysis of mobile communication operation data.

Key words: data mining, outlier analysis, key attribute subspace, knowledge discovery framework