Machine learning models are deterministic. Given the same input, a single model will always produce the same output. In practice, this approach does not provide information regarding the uncertainty of the prediction. When incorporating a model prediction into a forensic workflow, it is helpful to know how stable the prediction would be for an ensemble of models, each trained on a different data set. All training data sets should be casework relevant. Quantification of prediction uncertainty can be achieved by applying the principles of subjective logic to the predictions from an ensemble of models applied to a single sample. The resulting belief masses and uncertainty are especially useful in binary problems expressed as likelihoods conditioned on two competing propositions. This approach will be demonstrated for a forensic fire debris classification problem using multiple chemometric models and the challenges of the approach will be discussed within the context of forensic classification and attribution problems.

Quantifying uncertainty in forensic applications of machine learning

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.