Recursive Data Mining for Author and Role Identification

Like paintings and verbal dialogues, written documents exhibit the author’s distinctive style and identification of the author of an anonymous document is an important and challenging task in computer security. Even more challenging is identification of a style of a group of diverse individuals acting in similar circumstances, like authors writing in certain literary period or people writing in a certain social role. The last application is important for analyzing hidden group communicating over the internet in which neither identities nor roles of the members are known. Other applications of the identification of such styles include fraud detection, author attribution and user profiling. The task of finding distinctive features of an artifact has much broader scientific implications that range from art and scriptures to network security. In this paper, we focus on capturing patterns in electronic documents. The approach involves discovering patterns at varying degrees of abstraction, in a hierarchical fashion. The discovered patterns capture the stylistic characteristics of either the author, or a group of authors, or even of the specific role that the author plays in relation to others. These patterns are used as features to build efficient classifiers. Due to the nature of the pattern discovery process, we call our approach Recursive Data Mining. The patterns discovered allow for certain degree of approximation, which is necessary for capturing non-trivial patterns on realistic datasets. Experiments on the Enron and SEA datasets, the former categorizes members into organizational roles and the latter categorizes a set of computer sessions, are conducted to substantiate our methodology. The results show that a classifier that uses the dominant patterns discovered by Recursive Data Mining performs better than the same classifier without features based on RDM patterns, in role detection and author identification.