Speaker: Dalia Kriksciuniene, Assoc.prof.dr. (ERCIM researcher at Masaryk University, Brno, Czech)
Title: Anomaly detection methods and applications
Numerous statistical, econometrical and intelligent methods are researched in the scientific works and industrial applications for finding repeating patterns and approximated distribution rules of data series. The main problem experienced by the scientific research is lack of regularities and rather chaotic nature of real data. That hinders performance of the forecasting methods based on finding regularities in data distribution, limits their application to analytics of historical data repositories, and reveals high unreliability of the models to react to occurring external impacts for real-time data streams.
The current research aims to explore application of computational analysis methods for detecting behaviour anomalies. The methods and the principles of anomaly detection are analysed from the theoretical perspective and evaluated for their ability to serve for analysis, forecasting and designing optimal strategies for the enterprises operation in various application domains. They include analysis of anomalies of the financial markets, caused by changes of investor activeness due to different calendar effects, impacts of various media announcements and news, fluctuations of market situations, including crises and bubbles. The anomaly detection is important in the areas of facility management based on sensor data system, detection of network intrusions and performance failures.
The results of currently published research cover several methods for detecting behaviour anomalies including information efficiency evaluation methods (Shannon's entropy, Hurst exponent), event impacts explored by the methods of interrupted time series, and evaluation of binary clustering algorithms.
The research results are evaluated by discussing the potential power of explored methods for detection behaviour anomalies in the expanding research area of Big Data.