ADAPTIVE OUTLIER DETECTION IN STREAMING TIME SERIES

Abstract

"Outlier" is a scientific term to describe things or phenomena that lie outside normal expectation or behaviour. In data mining outlier detection is a type of data analysis technique that seeks to determine and report such data objects which are grossly different from or inconsistent with the remaining set of data. The technique is used for data cleansing, spotting emerging trends and recognizing unusually good or bad performers. Typical applications are financial data analysis, intrusion detection, event detection in sensor networks, biomedicine etc. The existing outlier detection schemes aim to detect the global outliers from the entire time series data and therefore fail to detect the local outliers. The detection of local outliers is helpful as they tell the degree of isolation of objects from their immediate neighbourhood. The existing schemes process outliers by working on the entire outlier time sequences. But in case of streaming time series data, this is not possible as the data keeps on arriving from the source. In the proposed work, we aim to develop an algorithm that detects outliers from streaming time series. The outliers are extracted as abnormally behaving subsequences in the data. The emphasis is on detecting the local outliers in addition to global outliers. The notion of "outlierness" has also been introduced which is used to capture the extent of abnormal behaviour shown by the outliers. Further, the type is also defined. It refers to the deviation of outliers above or below the normal behaviour. The HOT SAX algorithm has been extended to detect the local outlier subsequences in the time series streams. The outlier distribution is generated on the basis of reference set, to develop a rule based adaptive model to classify outliers into local and global classes. The proposed work has been evaluated on real life datasets. The first dataset used is a daily vehicular traffic dataset, that is, Gotthard tunnel dataset- number of motorcycles in one direction (in year 2005). The other dataset used is ECG dataset. iii