Author : Vinod S.Wadne, Alka P.Beldar and Suvarna V.Nalawad

Outlier detection is an important problem which has been studied within various application domains and research areas. Most of the previous methods assume that data examples are exactly categorized as either normal class or negative class. However, in many applications data are imperfectly labeled due to various error and noise. These kinds of data can cause system to give output wrong; because the label is either damaged by noise or wrongly labeled so that a normal data behaves like outlier. These kinds of data make outlier detection difficult as compared to clearly separated data. To handle uncertain data one classifier is used i.e. SVDD (model based outlier detection). The propose system work in two steps. In first step we calculate likelihood values or confidence score for each data example of training data, which define the degree of membership towards a positive or normal class. These generated likelihood values for training data are passed to the SVDD classifier to detect outlier. In this phase, the contribution of the examples with the least confidence score on the construction of the decision boundary has been reduced.

Keywords: Imperfectly labeled data, SVDD classifier

Article published in International Journal of Current  Engineering  and Technology, Vol.4,No.2 (April- 2014)