Abstract: Considering the distance- and density-based outlier detection algorithms are often sensitive to a nearest neighbor parameter k, termed k-center offset outlier factor (COOF), a robust outlier detection criterion for the characterization of abnormal degree of each data object was proposed. Each data object is included in a region within its k nearest neighbors, and the center of region will migrate with the change of nearest neighbor parameter k. In general, the variation of center offset of k nearest neighbor region is greater for an outlier than a normal object. According to this observation, for each data object, COOF is defined as the accumulation of this kind of offset when increasing the nearest neighbor parameter from one to k. Finally, the outlier detection algorithm based on COOF was also presented. Through artificial data and real data experimental simulations show that COOF is insensitive to parameter k, and has more stable and accurate outlier detection performance compared to k nearest neighbor, local distance-based outlier factor and local outlier factor, which are the distance-based method and density-based method respectively.

Key words: outlier detection, k nearest neighbor, local outlier factor, center offset outlier factor