Abstract:

Contributions on outlier detection in time series data have seldom taken into account the data cyclical nature and numerical attributes values. An algorithm to find periodic patterns under different granul arities was proposed, which could be  used to detect outliers in time series data with numerical attributes. This method could avoid a false alarm, and experimental results showed that it could not only correctly identify multi-ranule periodic patterns but also effectively detect outliers in data. Compared to outlier detection methods without periodic patterns, the results showed  that it could  reduce falsealarms.

Key words: periodicity analysis; time series ; granularity; outlier detection