Handwriting analysis has long been a largely subjective field of study, relying on visual inspections from trained examiners to determine if questioned documents come from the same source. In recent years, however, efforts have been made to develop methods and software which quantify the similarity between writing samples more objectively. Researchers funded by CSAFE developed and tested a new statistical method for handwriting recognition, using a score-based likelihood
ratio (SLR) system to determine the evidential value.
Madeline Quinn Johnson
Danica M. Ommen
Statistical Analysis and Data Mining
03 December 2021
Apply the SLR system to various handwritten documents.
Evaluate the system’s performance with various approaches to the data.
CSAFE collected handwriting samples from 90 participants, using prompts of various lengths to get samples of different sizes. These writing samples were broken down into graphs, or writing segments with nodes and connecting edges, then grouped into clusters for comparison.
When comparing the gathered samples, Johnson and Ommen considered two possible scenarios:
Common Source Scenario:
two questioned documents with unknown writers are compared to determine whether they come from the same source.
Specific Source Scenario:
a questioned document is compared to a prepared sample from a known writer.
They then used Score-based Likelihood Ratios (SLRs) to approximate the weight of the evidence in both types of scenarios.
The researchers used three different approaches when generating the known non-matching comparisons for the specific source SLRs:
only uses comparisons between the questioned document (the trace) and a collection of writers different from the specific source (the background population).
only uses comparisons between writing from the specific source and the background population.
only uses comparisons between samples from different writers in the background population.
Once the SLRs for each scenario were generated, they used random forest algorithms to determine comparison scores, including a pre-trained random forest using all of the gathered data, and one trained according to the relevant SLR.
In common source scenarios, the trained random forest performed well with longer writing samples, but struggled with shorter ones.
The specific source SLRs performed better than the common source SLRs because they are tailored to the case at hand.
In all scenarios, it was more difficult for the SLR system to confirm samples with the same source than with different sources.
FOCUS ON THE FUTURE
The SLRs do not perform well with short documents, possibly due to a mismatch between the number of clusters used and the length of the document. Future work could determine the optimal number of clusters based on the document’s length.
Because the SLRs provide data on the strength of forensic handwriting evidence for an open-set of sources, this approach is an improvement on the previous clustering method developed by CSAFE, which used a closed set of known sources.