Ensemble SLRs for Forensic Evidence Comparison

To strengthen the statistical foundations of forensic evidence interpretation, likelihood ratios and Bayes factors are advocated to quantify the value of evidence. Both methods rely on formulating a statistical model, which can be challenging for complex evidence. Machine learning score-based likelihood ratios have been proposed as an alternative in those cases. Under this framework, a (dis)similarity score and its distribution under alternative propositions are estimated using pairwise comparisons, but pairwise comparisons of all the evidential objects result in dependent scores. While machine learning methods may not require distributional assumptions, most assume independence. We introduce a sampling and ensembling approach to remedy this lack of independence. We generate sets where assumptions are met to develop multiple score-based likelihood ratios later aggregated into a final score to quantify the value of evidence.