A0630
Title: A Bayesian hierarchical mixture model for classifying compositional data
Authors: Catherine Holland - University of Glasgow (United Kingdom)
Tereza Neocleous - University of Glasgow (United Kingdom) [presenting]
Oliver Stoner - University of Glasgow (United Kingdom)
Abstract: Compositional data refers to multivariate sets of non-negative components, where the primary interest is in the size of the components relative to their total and relative to each other. Such data can be expressed directly as proportions summing to one or measured in absolute terms. Standard statistical techniques are often unsuitable due to the constrained nature of the data, complex correlations, and the presence of zeros, particularly structural zeros, which represent true absence rather than values below detection limits. In forensic glass analysis, compositional data methods have emerged as powerful tools to account for dependencies between elemental components and aid in quantifying the strength of the glass evidence found at crime scenes. The focus is on a forensic elemental glass database that contains a significant number of structural zeros. In these instances, traditional log-ratio transformations, the main technique for compositional data, are undefined. A flexible integrated clustering approach is proposed within a Bayesian hierarchical model for compositional data with structural zeros and a multilevel structure. The model is motivated by the interest in classifying glass fragments by type. Performance was evaluated using five-fold cross-validation, demonstrating superior classification accuracy over less flexible methods. The method offers computational efficiency, supporting the method's practicality for forensic and other real-world applications.
