Speaker: Elías Hernandis Prieto.

Abstract: Forensic evidence evaluation using the likelihood ratio framework requires knowledge about the probability distribution of the data. For evaluating samples of glass remains, this translates to obtaining the joint probability distribution of the relative concentrations of different isotopes found in glass samples, which are obtained by LA-ICP-MS. This multivariate distribution is highly correlated and there is little data available to learn from. Linear-Gaussian Bayesian networks are a mathematical tool to model conditional independence between components of a continuous distribution, a first step towards learning the distribution of the data in glass evidence. They also present a powerful tool to visualize these independences, which is especially useful for experts without a background in AI or ML for whom we demonstrate a new interactive web-based tool. We briefly discuss how different structure learning algorithms can be applied to real-world data provided by the German Bundeskriminalamt and the Florida International University, and evaluate the performance of the obtained networks using the Bayesian Information Criterion score, as well as their ability to generalize using cross validation and test-train data separation. Overall, we find that the algorithms considered perform equivalently well for our data, and that the structures learned generalize correctly. These results are a first step to help forensic scientist perform a more robust statistical evaluation and we hope they can be published soon.