Handwriting identification using random forests and score‐based likelihood ratios

Abstract

The student author, whose presentation of the scholarship herein was approved by the program of study committee, is solely responsible for the content of this dissertation/thesis. The Graduate College will ensure this dissertation/thesis is globally accessible and will not permit alterations after a degree is conferred.

Ensemble learning for score likelihood ratios under the common source problem

“…There has been a general call to strengthen the scientific basis and statistical foundations in criminal justice and to push for more objective means of comparison in forensic analysis [27], and handwriting analysis has been previously identified as an area to be strengthened [10]. Previous work has shown the potential of the SLR approach in handwriting analysis [21, 25], and our work contributes in the same direction. The simulation study suggests that an ensemble approach can enhance traditional SLRs.…”

“…An alternative to the LR relies on using a SLR, often involving popular machine learning algorithms. Adapting the notation presented in [21, 25, 43], the SLR can be generically defined as follows: $$$$ \mathrm{SLR}\left({u}_x,{u}_y\right)=\frac{g\left(\Delta \left({u}_x{u}_y\right)\mid {H}_p\right)}{g\left(\Delta \left({u}_x{u}_y\right)\mid {H}_d\right)}, $$$$ where $$$ \Delta \left(\right) $$$ is a (dis)similarity metric that allows the comparison of items $$$ {E}_x $$$ and $$$ {E}_y $$$ via their observed features $$$ {u}_x $$$ and $$$ {u}_y $$$, respectively, and the conditional density functions $$$ g\left(\cdot \mid {H}_p\right) $$$ and $$$ g\left(\cdot \mid {H}_d\right) $$$ allows to assess the likelihood of the score obtained under the alternative propositions. The numerator (denominator) in Equation () can be interpreted as the likelihood of the score under $$$ {H}_p $$$ ($$$ {H}_d $$$).…”

“…Pairs from the same source (KM) can be considered as positive cases ($$$ y=1 $$$), and pairs from different sources (KNM) as negative cases $$$ \left(y=-1\right) $$$. The Random Forest classifier [6] has become a popular model in forensic science to construct a similarity metric [25, 40, 41]. Once the Random Forest classifier is trained, it can be used to map from the features space to a univariate score between zero and one, $$$ \hat{rf}:{\left({\mathrm{\mathbb{R}}}^{+}\right)}^{P+1}\mapsto \left[0,1\right] $$$.…”

“…Although the SLR framework has shown promise in different areas of forensics (handwriting: [8,13,21,25], glass: [40,41], fingerprints: [20,28,35], speaker recognition: [19], ink: [36], MDMA tablets: [3,4], digital: [17], cameras: [42]), concerns have been raised regarding their behavior and use in forensic settings [32,34], and their evaluation has been the subject of extensive research in the literature [18,33]. Ishihara and Carne [22] summarize the benefits and shortcomings of using score-based methods; using a lower dimensional metric reduces the need for complex models, and simpler estimation procedures are required to estimate univariate conditional scores.…”

A Probabilistic Formalisation of Contextual Bias: from Forensic Analysis to Systemic Bias in the Criminal Justice System

“…Algorithms are increasingly used in forensic science to inform and complement examiners' conclusions. While law enforcement primarily uses algorithms in three domains: DNA analysis, latent print analysis and facial recognition (US Government Accountability Office, 2020); algorithms have also been developed in a variety of pattern evidence disciplines including firearms and toolmarks (Rice et al, 2020; Tai & Eddy, 2018), handwriting (Crawford et al, 2021; Johnson & Ommen, 2022) and shoeprints (Kong et al, 2019; Park & Carriquiry, 2022). Swofford and Champod (2021) note that a range of levels of algorithm implementation is possible: At the lowest level, the algorithm output is used after the analyst has formed an initial opinion as optional supplemental assistance; at the highest level, the algorithm is solely responsible for the resulting conclusion.…”