A1356
Title: Bayesian nonparametric approaches for functional clustering
Authors: Wenyu Gao - University of North Carolina at Charlotte (United States) [presenting]
Abstract: Functional clustering of high-dimensional data is essential for uncovering latent structure in complex signals such as fMRI, where the number of underlying clusters is often unknown. Traditional functional clustering approaches, such as centroid-based, density-based, and parametric model-based methods, provide straightforward implementations but typically require pre-specifying the number of clusters and offer limited flexibility in capturing complex dependencies. In contrast, Bayesian nonparametric (BNP) methods, particularly Dirichlet process mixture (DPM) models, allow the number of clusters to be inferred directly from the data while naturally quantifying uncertainty in cluster assignments. Building on this framework, weighted Dirichlet process mixture (WDPM) models incorporate auxiliary or subject-level information through weights, enabling more informed clustering of functional signals. WDPM models are especially well-suited to high-dimensional neuroimaging studies such as fMRI in autism spectrum disorder (ASD), as they can accommodate high dimensionality, spatial correlation, and heterogeneity across subjects. This BNP framework offers a principled, data-driven alternative to conventional methods, providing enhanced flexibility, interpretability, and adaptability for functional clustering across diverse application domains.
