The efficiency of conformal predictors for anomaly detection

This thesis explores the application of conformal prediction to the anomaly detection domain. Anomaly detection is a large area of research in machine learning and many interesting techniques have been developed to detect ‘abnormal’ behaviour of objects that do not conform to typical behaviour. Recently conformal predictors (CP) have emerged which allow the detection of the non-conformal behaviour of objects using some measures of non-conformity. Conformal predictors have the advantage of delivering provably valid confidence measures under the exchangeability assumption that is usually weaker than those traditionally used. The suitability of existing performance criteria for conformal predictors applied to anomaly detection problems is explored. A difficulty in some anomaly detection domains is collecting sufficient examples of anomalies. Two new performance criteria average pvalue (APV) and logarithmic average p-value (LAPV) are proposed that do not require labelled anomalies unlike previous criteria. These new criteria allow the discovery of appropriate non-conformity measure for anomaly detection under any setting. Experiments are conducted with real world data on ship vessel trajectories. A dimensionality reduction package is used and a comparison of a kernel density based nonconformity measure with a k-nearest neighbours non-conformity measure is presented and the results are discussed. In previous applications of applying conformal prediction to anomaly detection, typically one global class of ‘normal’ is used to encompass all previous data. However with vessel trajectories there exists an information hierarchy between objects. In this thesis a multi-class hierarchy framework for the anomaly detection of trajectories is proposed. Experiments are conducted comparing the multi-class hierarchy approach to the traditional global class under various conditions and the results are discussed. A study of aggregating p-values from various classes in the hierarchy is also presented. This framework can also