New Methods for Deviation-Based Outlier Detection in Large Database

Outlier (also called deviation or exception) detection is an important function in data mining. In identifying outliers, the deviation-based approach has many advantages and draws much attention. Although a linear algorithm for sequential deviation detection is proposed, it is not stable and always loses many deviation points. In this paper, we present three algorithms on detecting deviations. The first algorithm is time proportional to the square of the dataset length, and the second is time proportional to the square of the number of distinct data values. These two algorithms lead to same result, while the latter is much more efficient than the former. In the third algorithm, a deviation factor is defined to help finding deviation points. Although leading to approximation results, it is the most efficient of the three, especially to large datasets with lots of distinct values.