Anomaly detection in temporal data mining

Temporal data mining is interested in analyzing big amount of temporal data to automatically discover interesting patterns or relationships which gives useful information. Getting the valid, useful information is the main motivation of temporal data mining. The major tasks of temporal data mining are; indexing, clustering, classification, prediction, summarization, anomaly detection and segmentation. In temporal data, anomaly detection or novelty detection is the identification of interesting patterns. Several anomaly detection algorithms have been proposed in the literature. In this thesis, anomaly detection methods, HOT-SAX, WAT, PAV and MPAV are investigated and accuracy of these methods is tested with real and synthetic time series data sets. Also, symbolic family of temporal data representation techniques is examined and effectiveness of these representation techniques is tested in terms of anomaly detection. Time series data sets with tagged anomaly points were used for comparing anomaly detection and representation methods. R statistical programming language was used for application. In general HOT-SAX algorithm performed better than other anomaly detection algorithms and Trend-based symbolic representation technique had better accuracy than other representation techniques.