An Evaluation Of Authorship Attribution Using Random Forests
Faculty of Health, Engineering and Science
School of Computer and Security Science
Electronic text (e-text) stylometry aims at identifying the writing style of authors of electronic texts, such as electronic documents, blog posts, tweets, etc. Identifying such styles is quite attractive for identifying authors of disputed e-text, identifying their profile attributes (e.g. gender, age group, etc), or even enhancing services such as search engines and recommender systems. Despite the success of Random Forests, its performance has not been evaluated on Author Attribtion problems. In this paper, we present an evaluation of Random Forests in the problem domain of Authorship Attribution. Additionally, we have taken advantage of Random Forests' robustness against noisy features by extracting a diverse set of features from evaluated e-texts. Interestingly, the resultant model achieved the highest classification accuracy in all problems, except one where it misclassified only a single instance.