A Bayesian Hierarchical Mixture Model with Applications in Forensic Handwriting Analysis


When forensic examiners compare handwritten evidence they often focus on small details of writing. Likewise, our statistical approach to the comparison of such documents begins by decomposing writing into small meaningful connected pieces of ink, often corresponding to letters. We treat these small pieces of connected ink as graphical structures with nodes and edges and group them into types based on similarity of their structures. The frequency at which graph types appear in writing, along with measurements taken on the small graphs serve as data for a Bayesian hierarchical mixture model. We will discuss a feature selection method used to decide which graph types to include in the model, and theoretical properties of the selection method. Through a latent variable, the model is able to separate print and cursive writings. The groupings are used to introduce a mixture aspect to the model and we assign different priors to similar groups of writers. We will present results of a writer discrimination analysis based on a closed set of writers.

This is join work with Amy Crawford.