B0314
Title: Profile processes for approximate Bayesian computational model selection in forensic identification-of-source problems
Authors: Christopher Saunders - South Dakota State Univerisity (United States) [presenting]
Janean Hanka - South Dakota State University (United States)
Danica Ommen - Iowa State University (United States)
JoAnn Buscaglia - FBI Laboratory (United States)
Abstract: The forensic identification-of-source problem usually considers two competing propositions for how the evidence (traces) and their corresponding quantifications of measured physical and chemical properties arose. The first proposition is typically associated with the prosecution model that a specified source is the actual source of the traces; the second proposition is associated with the defence model that a source in some relevant background population is the actual source of the traces. With the advent of modern analytical methods leading to moderate to high-dimensional analytical quantifications of the evidence, this type of data analysis has become technically difficult in the sense that there is no natural likelihood structure on which to build the Bayesian infrastructure for model selection. This has led to a focus on using learned metrics that facilitate the use of so-called score-based likelihood ratios (SLRs) and the use of approximate Bayesian computation (ABC) for model selection. The focus is on developing U-processes that can be used as approximate generative distributions in ABC methods for model selection which will lead to a class of Bayes factor-like objects that can be used in both pattern recognition problems and as quantifications of evidential value in forensic identification of source problems related to measurements of physical and chemical properties of component materials commonly used in improvised explosive devices.
