Normalized residual-based outlier detection

Outlier detection is an important issue in data mining and knowledge discovery. The aim is to find the patterns that deviate too much from others. In this paper, a universal outlier detection method based on normalized residual is proposed. Different from previous methods, the residual of a pattern is calculated corresponding to its nearest normal patterns, so that the interaction between outliers is eliminated. To implement this, the method first estimates the center of normal patterns and derives the initial set of them, and then iteratively calculates the residual of the nearest pattern outside the set. Those with small residuals will be added to the set of normal patterns, and others are picked out as outliers. An effective distance weighting is also introduced to the calculation of the normalized residual. Simulation results show that the proposed method is efficient in detecting outliers and can hold a high detection probability even when 30% outliers appear in the dataset.