Handwriting identification using random forests and score-based likelihood ratios

Abstract

Handwriting analysis is conducted by forensic document examiners who are able to visually recognize characteristics of writing to evaluate the evidence of writership. Recently, there have been incentives to investigate how to quantify the similarity between two written documents to support the conclusions drawn by experts. We use an automatic algorithm within the “handwriter” package in R, to decompose a handwritten sample into small graphical units of writing. These graphs are sorted into 40 exemplar groups or clusters. We hypothesize that the frequency with which a person contributes graphs to each cluster is characteristic of their handwriting. Given two questioned handwritten documents, we can then use the vectors of cluster frequencies to quantify the similarity between the two documents. We extract features from the difference between the vectors and combine them using a random forest. The output from the random forest is used as the similarity score to compare documents. We estimate the distributions of the similarity scores computed from multiple pairs of documents known to have been written by the same and by different persons, and use these estimated densities to obtain score-based likelihood ratios (SLRs) that rely on different assumptions. We find that the SLRs are able to indicate whether the similarity observed between two documents is more or less likely depending on writership.