Flexible and Limited Supervision for Anomaly Detection

Traditionally, anomaly detection was approached from an unsupervised perspective. This was the case because it is often difficult or impossible to acquire labels for anomalous instances. Hence, algorithms for these tasks were built upon assumptions such as that anomalies are rare and are different (e.g., far away or lie in a low-density region) from normal examples. However, there is a growing realization that some labeled data is needed to solve anomaly detection problems in practice. On the hand labeled information can help correct when the assumptions encoded in anomaly detectors are violated. For example, there may be rare but normal behavior such as maintenance operations which should not be flagged as anomalous. On the other hand labeled data is needed to configure (i.e., set the hyperparameters) and evaluate anomaly detectors. The goal of this PhD is to develop novel approaches to training and evaluating anomaly detection algorithms based on limited and flexible supervision. By flexible supervision, we mean that the feedback may be in the form of labels for one class, labels for groups of instances or constraints.