Abstract
Anomaly detection (AD) algorithms are widely used for data-driven decision support in domains where quantifying risk is critical, such as identifying fraudulent healthcare providers in public health insurance, consumer lending, and detecting aberrant patterns in human electroencephalography (EEG) records. However, AD in decision support is challenging due to the multitude of data modalities (e.g. timeseries, or structural data) and data scale, unavailability of ground truth labels for learning and evaluation, and difficulty in yielding human interpretable results for domain-specific problems. This thesis proposes to address the challenges and build intelligent detection systems with the following desirable properties: unsupervised, explainable, scalable, and equitable. Throughout, we propose novel AD algorithms that enable better decision support by addressing domain-specific key challenges such as including domain or expert knowledge, mitigating bias that may adversely affect minority groups, and handling aberrant behavior involving a group of actors. We present applications in public healthcare fraud, and health monitoring.