Presentation: 2025 ND EPSCoR Annual conference 

October 21, 2025, NDSU Memorial Union, Fargo, North Dakota

MACHINE LEARNING SUBJECTIVE OPINIONS: AN APPLICATION IN FIRE DEBRIS ANALYSIS

This study used three machine learning (ML) ensemble models: linear discriminant analysis (LDA), random forest (RF), and support vector machine (SVM), to calculate the evidentiary value “likelihood ratio” based on the computed opinions. One hundred models were trained for each ML method. For each iteration, n samples were chosen by bootstrapping 60,000 of the in-silico total ion spectral data. (n = 100, 200, 1,000, 2,000, 20,000, 60,000). This data set comprises fire debris samples with ignitable liquids (IL) and without IL. The trained models were validated on 1,117 laboratory-generated fire debris samples. The pre-treated data set consisted of mass-to-charge ratios (m/z) from 30 to 160 (131 ions in total). However, only 26 ions were selected. The probabilities generated from each ensemble model for validation data were fitted to a beta distribution. The fitting parameters alpha (α) and beta (β) were used to compute the belief, disbelief, and uncertainty masses. These values were then used to calculate the subjective opinion. Then, using the opinion, the likelihood ratio for each validation sample was calculated using the odds form of the Bayes equation. The performance of each machine learning model (100 ensemble models for each technique) was evaluated by the area under the curve (AUC) of the receiver operating characteristic plot. Out of the three machine learning models, the random forest ensemble model had the highest area under the curve, whereas the AUC of LDA was the lowest. The AUC increased in all the models when the data set size was increased.