Article Preview
TopIntroduction
Outlier detection is a field of data mining with a wide range of applications in several domains varying from intrusion detection (Angelo et al., 2018), process monitoring (Wang and Mao, 2019), detection of fake iris images (2018), security threats (Sharma and Gupta, 2021) and identification of credit card frauds (Carcillo et al, 2019). Outlier points might be of several types that may occur in different magnitude, intensity and frequency. Here, the magnitude refers to the extent of variation from the normal behavior and frequency of anomalies can be measured by the percentage of outliers present in the data. Identifying these outlier points is a tough task, data scientists utilise both supervised and unsupervised machine learning models for this purpose. Supervised learning algorithms work based on the assumption that the test data has similar characteristics as the training data. However, their performance is poor if they encounter a set of data points which is outside the range of training dataset. In unsupervised algorithms, identification of outliers is based on the characteristic difference of the abnormal data points from the rest of the data. Yet, unavailability of ground truth makes it difficult to evaluate these unsupervised algorithms (Dong et al, 2020; Omer and Lior, 2018). Thus, choosing an algorithm that accurately identifies the outlier objects is an essential need for all kinds of applications. Appropriate selection of the evaluation parameters that pick out a good outlier detection algorithm considering all the features of anomalies present in the data is another mandatory requirement.
A number of evaluation parameters are available for analysing the accuracy of outlier detection algorithms, varying from precision and recall (Domingues et al., 2018), Receiver Operator Characteristic (ROC) curve (Goldstein and Uchida, 2016), outlier detection rate and false alarm rate (Shahid et al., 2013). Even so, most of the algorithms that exhibit high detection rate usually express high false alarm rate also. Nevertheless, good outlier detection algorithms ought to have high detection rate as well as low false alarm rate. Hence, the greatest challenge is to design an outlier detection system with high detection rate and low false alarm rate. Several single stage outlier detection algorithms are available in the literature (Domingues et al., 2018; Shahid et al., 2013). These algorithms show lower average performance in the presence of different types of outliers. This leads to the finding of outlier ensembles that exhibit higher performance than individual detection methods. Other than this, deep learning techniques (Vinayakumar et al, 2017) and reinforcement learning (Ghosh et al., 2017) can also be used to improve the detection rate.
Aggarwal and Sathe (2015) put forward the concept of outlier ensemble with suggestions to design an ensemble-based system for outlier detection. Designing an ensemble-based outlier detection system with high detection rate and low false alarm rate is difficult (Aggarwal and Sathe, 2015), because an ensemble design demand understanding the structure of the data, form and frequency of outliers present in the dataset. The accurate evaluation of algorithms in the absence of labelled datasets is another challenge in such algorithm design.The taxonomy of outlier detection is given in Figure 1.