Robust and efficient density based outlier detection methods for streaming data

Detection of outliers is critical for reliable results and efficient performance in all scientific studies. Most outlier detection methods proposed in the literature work in batch mode learning, where all samples are loaded into memory. However, due to the enormous increase in data size and the demand for instant analysis, there is a growing need for methods that operate in incremental mode. Moreover, detecting outliers from high-speed large-volume data streams is more difficult and complex than batch-mode learning. There are some recent studies focusing on incremental learning mode to detect outliers from data streams with varying distributions. Although they show promising results, there are still some major limitations such as poor detection in high dimensional data, inability to detect a long sequence of outliers (small cluster of outliers), use of improper general purpose metrics, the need for a large number of hyperparameters and the difficulty of tuning them, either labels outliers or assigns outlier scores. The aim of this thesis is to address these problems and to offer solutions. For this purpose, firstly, a new robust unsupervised outlier detection (RiLOF) method is presented which implements a newly developed metric named MoNNAD using the median of the local absolute deviation of the Loal Outlier Factor (LOF) values of the samples. Secondly, a novel incremental Local Density and Cluster-Based Outlier Factor (iLDCBOF) method, which unifies incremental LOF (iLOF) and incremental density-based spatial clustering of applications with noise methods and employs a newly developed concept called CkNN, which can automatically adapt hyperparameters for different data streams, is proposed. Thirdly, a new incremental Multi-Class Outlier Detection (iMCOD) method, which integrates the incremental Support Vector Machine (iSVM) and iLOF in a unified framework, capable of simultaneously performing both multi-class outlier detection and classification, is developed. The proposed methods are comprehensively analyzed qualitatively and numerically and benchmarked with state-of-the-art outlier detection methods in real-world data sets.