Explainable Contextual Anomaly Detection using Quantile Regression Forests

Chandola et al (2009) subdivided anomalies into three types: point anomalies (an object is considered anomalous when compared against the rest of objects), contextual anomalies (an object is anomalous in a specific context), and collective anomalies (a collection of objects is anomalous with respect to the entire dataset). The analysis of anomalies has a wide range of applications, such as in network security (Ahmed et al, 2016a), bioinformatics (Spinosa and Carvalho, 2005), fraud detection (Ahmed et al, 2016b), and fault detection and isolation (Hwang et al, 2009). Anomaly analysis consists of two equally important tasks: anomaly detection and anomaly explanation. A wealth of'shallow' machine learning based methods, i.e., not based on deep learning, have been proposed to detect anomalies (Chandola et al, 2009). More recently, many deep learning based anomaly detection methods have also been developed (Pang et al, 2021). However, deep learning based anomaly detection methods are notoriously known as not being interpretable, in the sense that generally both the model itself is non-transparent and the resulting anomaly scores are challenging to interpret without the use of a post-hoc explainer.