Projecting Away the Class Imbalance Problem in Author Attribution

Abstract

Author identification algorithms attempt to ascribe document to author, with an eye towards diverse application areas including: forensic evidence, authenticating communications, and intelligence gathering. We view author identification as a single label classification problem, where 2000 authors would imply 2000 possible categories to assign to a post. Experiments with a naive Bayes classifier on a blog author identification task demonstrate a remarkable tendency to over-predict the most prolific authors. Literature search confirms that the class imbalance phenomenon is a challenge for author identification as well as other machine learning tasks. We develop a vector projection method to remove this hazard, and achieve a 63% improvement in accuracy over the baseline on the same task. Our method adds no additional asymptotic computational complexity to naive Bayes, and has no free parameters to set. The projection technique will likely prove useful for other natural language tasks exhibiting class imbalance.