Niklas Melton, a doctoral candidate in computer science, will defend their dissertation titled “Incremental Cluster Validity Indices and Their Role in Interpreting Lifelong Learning Systems.” Their advisor, Dr. Donald Wunsch is an Director in Kummer Institute Center for Artificial Intelligence and Autonomous Systems and Mary Finley Mo Professor in Computer Engineering. The dissertation abstract is provided below.

Clustering and supervised learning are often treated as distinct paradigms, yet both fundamentally rely on the structure of data in feature space. This dissertation investigates the interplay between cluster validity indices (CVIs) and supervised learning, with a particular focus on real-time and lifelong learning settings where data arrive incrementally and cannot be revisited. Across four studies, this work develops new methodologies for online cluster validation, demonstrates how these tools can enhance interpretability through supervised learning, and applies them to address foundational challenges in evaluating performance degradation in continual learning systems. The first contribution extends the family of incremental cluster validity indices (iCVIs), enabling a broad set of widely used validation metrics to operate in streaming environments. Through extensive experimentation on synthetic and real-world datasets, this study reveals systematic differences in how iCVIs respond to under- and over-partitioning, establishing both their practical utility and their limitations in online clustering scenarios. Building on this foundation, the second study introduces the Meta iCVI, an ensemble framework that leverages supervised learning to combine multiple iCVIs into a single, interpretable indicator of partition quality. This approach achieves high accuracy in labeling clustering outcomes and demonstrates how supervised models can enhance the usability of unsupervised validation signals. The third contribution shifts focus to supervised lifelong learning and critically examines the limitations of accuracy-based metrics for diagnosing catastrophic forgetting. It shows that performance degradation can arise not only from forgetting but also from intrinsic data properties, specifically class overlap in feature space. This phenomenon, termed memory overshadowing, is shown to impose a fundamental performance bound independent of model memory, thereby exposing a key flaw in widely used evaluation practices. Motivated by this insight, the fourth study reintroduces cluster-based reasoning into supervised evaluation by proposing new metrics derived from iCVIs. The Overlap Index quantifies feature-space overlap directly, while the Overshadowing and Forgetting Index disentangles performance degradation into contributions from true forgetting and data-induced interference. Together, these tools provide a principled and incremental framework for diagnosing learning dynamics in both online and batch settings. Collectively, this dissertation establishes a novel connection between cluster validation and supervised learning evaluation. By extending CVIs to incremental settings, integrating them with supervised meta-models, and repurposing them to analyze lifelong learning behavior, this work provides both theoretical insight and practical tools for interpreting model performance in complex, non-stationary environments.