DETECTING COMMUNITY STRUCTURE IN DOMAIN-FEATURES DATA

Abstract

This thesis investigates the enhancement of clustering methodologies on general domains, motivated by questions arising in molecular dynamics (MD) simulations. Traditional clustering methods face limitations such as parameter dependence and inability to reveal complex structures, which Partitioned Local Depth (PaLD) addresses by employing a parameter-free framework that emphasizes local relationships to discern nuanced community structures. The integration of domain-specific features into PaLD extends its applicability, enabling more contextually relevant analyses across various data-centric fields, including time-series, geospatial, and network data. The thesis advocates for future research to refine PaLD, particularly for large-scale MD datasets, underscoring the method’s potential to significantly advance data science and biomolecular research.