Classification Analysis with a Twist: Identification of the Writership of Handwritten Documents by Pairwise Application of Linear Discriminant Analysis

• Activity Number: 607
• Type: Contributed
• Date/Time: Thursday, August 6, 2009 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): John J. Miller*+ and Amanda B. Hepler and Donald T. Gantz and Daniel B. Carr and Clifton Sutton and Robert Patterson
• Companies: George Mason University and George Mason University and George Mason University and George Mason University and George Mason University and George Mason University
• Keywords: classification ; biometric identification ; forensics ; handwriting

Abstract

We consider a classification problem with the following complications: many classes (hundreds or thousands); many variables (again hundreds or thousands); few training observations per class (perhaps as few as four). The objective is not necessarily to classify every test observation correctly but to generate scores which are then combined over natural groups that occur in the test data. The object is to correctly classify these groups of test observations. Our biometric technique is applied in forensics to successfully identify writership of handwritten documents. In handwriting, classes, observations, variables, and observation groups correspond to writers, characters or graphemes, features, and documents respectively. Our method achieves very high accuracy using stepwise linear discriminant analysis on pairs of classes and combining pairwise results to yield the desired scores.