Date of Award:


Document Type:


Degree Name:

Doctor of Philosophy (PhD)


Computer Science

Committee Chair(s)

Shuhan Yuan


Shuhan Yuan


Soukaina Filali Boubrahimi


Curtis Dyreson


John Edwards


Luis Gordillo


In the digital transformation era, safeguarding online systems against anomalies – unusual patterns indicating potential threats or malfunctions – has become crucial. This dissertation embarks on enhancing the accuracy, explainability, and ethical integrity of anomaly detection systems. By integrating advanced machine learning techniques, it improves anomaly detection performance and incorporates fairness and explainability at its core.

The research tackles performance enhancement in anomaly detection by leveraging few-shot learning, demonstrating how systems can effectively identify anomalies with minimal training data. This approach overcomes data scarcity challenges. Reinforcement learning is employed to iteratively refine models, enhancing decision-making processes. Transfer learning enables the application of insights across domains, improving system versatility. The integration of Generative Pre-trained Transformer (GPT) models marks a significant advancement, offering enhanced precision in anomaly identification through sophisticated language modeling techniques.

Exploring explainability and root cause analysis, the dissertation introduces advanced frameworks that shed light on the mechanisms behind anomaly detections across different data types. InterpretableSAD enhances sequential log data analysis, pinpointing specific anomalous events to clarify detection processes. RootCLAM addresses tabular data anomalies through causal inference, identifying root causes and suggesting actionable mitigation strategies. The narrative extends to time series data with RecAD and AERCA; RecAD offers algorithmic recourse by proposing minimal-cost actions for correcting anomalies, while AERCA utilizes an autoencoder-based framework to unravel Granger causality, illuminating causative factors behind anomalies. These frameworks empower users with the knowledge and tools to understand and act upon findings, facilitating the identification of irregularities across diverse data landscapes.

Ethical integrity remains paramount, addressed through the CFAD framework, which ensures counterfactual fairness by embedding ethical principles directly into anomaly detection processes. This guarantees equitable treatment across scenarios, advocating for technologies that serve all users equitably and challenge inherent biases.

Extensive evaluations on various datasets demonstrate the proposed models' effectiveness in addressing anomaly detection challenges. This dissertation contributes to advancing techniques that are not only accurate but also interpretable and fair, promoting the responsible use of anomaly detection in real-world applications. This dissertation lays a solid foundation for further exploration into advanced anomaly detection techniques, promising to guide the development of even more robust, transparent, and equitable systems in the digital age.