New methods in outlier detection

Outlier detection has been studied extensively in data mining. However, as the emergence of huge data sets in real-life applications nowadays, outlier detection faces a series of new challenges. Many traditional outlier detection techniques do not work well in such an environment. Therefore, developing up-to-date outlier detection methods becomes urgent tasks. In this thesis, we propose several new methods for detecting different kinds of outliers in high-dimensional data sets from two different perspectives, namely, detecting the outlying aspects of a data object and detecting outlying data objects of a data set. Specifically, for detecting the outlying aspects of a data object, we propose the problems of mining outlying aspects and mining contrast subspaces; for detecting outlying data objects of a data set, we propose the problems of mining contextual outliers and mining Markov blanket based outliers. We develop efficient and scalable algorithms to tackle the computational challenges. We also conduct comprehensive empirical studies on real and synthetic data sets to verify the effectiveness of the proposed outlier detection techniques and the efficiency of our algorithms.