Detection of Users Suspected of Pretending to Be Other Users in a Community Site by Using Messages Submitted to Non-Target Categories

Detection of Users Suspected of Pretending to Be Other Users in a Community Site by Using Messages Submitted to Non-Target Categories ⋆ Naoki Ishikawa a, Ryo Nishimura a, Yasuhiko Watanabe a, Masaki Murata b, and Yoshihiro Okada a aRyukoku University, Dep. of Media Informatics, Seta, Otsu, Shiga, 520-2194, Japan t10m096@mail.ryukoku.ac.jp, r nishimura@afc.ryukoku.ac.jp, {watanabe, okada}@rins.ryukoku.ac.jp bTottori University, Koyama-Minami, Tottori, 680-8552, Japan murata@ike.tottori-u.ac.jp Abstract. Some users abuse the anonymity and disrupt communications in a community site. Authorship identification based on analyzing stylistic features of messages is effective in detecting these inadequate users. However, in this method, the scope of target users was often limited because the criteria for selecting learning examples were strict. To relax the criteria and extend the scope of target users, we propose a method of detecting users suspected of pretending to be other users in a certain category of a community site by using their messages submitted to other (non-target) categories. Also, we show the accuracy of user identification when we relax the criteria of selecting learning examples and extend the scope of target users. Keywords: user identification, extension of scope of target users, messages in other cate- gories, stylistic feature, community site 1 Introduction In these days, many people use community sites, such as Q&A sites and social network services, where users share their information and knowledge. One of the essential factors in community sites is anonymous submission. It is important to submit messages anonymously to a community site. This is because anonymity gives users chances to submit messages without regard to shame and reputation. However, some users abuse the anonymity and disrupt communications in a com- munity site. For example, some users pretend to be other users by using multiple user accounts and attempt to manipulate communications in the community site. Manipulated communications dis- courage other submitters, keep users from retrieving good communication records, and decrease the credibility of the community site. As a result, it is important to detect users suspected of pre- tending to be other users to manipulate communications in a community site. In this case, identity tracing based on user accounts is not effective because these suspicious users often attempt to hide their true identity to avoid detection. A possible solution is authorship identification based on analyzing stylistic features of messages (Craig, 1999) (de Vel et al., 2001) (Koppel et al., 2002) (Argamon et al., 2003) (Zheng et al., 2006). However, most of previous studies selected their learning examples carefully and, consequently, limited the scope of target users. For example, if target users are set to submitters in a certain category of a community site, learning examples are generally selected from their messages submitted to the category, not other categories. Actually, in ⋆ This research has been supported partly by the Grant-in-Aid for Scientific Research under Grant No.20500106 and 22650021. Copyright 2010 by Naoki Ishikawa, Ryo Nishimura, Yasuhiko Watanabe, Masaki Murata, and Yoshihiro Okada PACLIC 24 Proceedings 637 (Ishikawa et al., 2010), we developed learning examples by using target user’s messages submitted …

Content truncated to 500 words