A Relative Deviation Detection for Time Series Data Based on Equality

Outlier detection is an important problem in various fields. A lot of algorithms have been proposed, meanwhile a lot of definitions. Unsatisfying point is that definitions seem vague, which makes the problem an ad hoc one. We analyze the nature of outliers, and give a supplementary definition. Based on it, we develop an efficient relative deviation degree (RDD) algorithm to identify outliers, which converts outlier problem to pattern and degree problem. We give two type of application in time series data – line type and curve type and introduce a longest k -turn subsequence problem. We treat the structure of pattern as a kind of order, which is a novel view. We also present experimental results on synthetic and real datasets showing the efficiency of our technique.