Detecting Spam on Twitter via Message-Passing Based on Retweet-Relation - Technologies and Applications of Artificial Intelligence - page 56

Detecting Spam on Twitter via Message-Passing
Based on Retweet-Relation
Pei-Chi Chen 1 , Hahn-Ming Lee 1 , 2 , Hsiao-Rong Tyan 3 , Jain-Shing Wu 1 , 4 ,
and Te-En Wei 1
1 Dept. Computer Science and Information Engineering
National Taiwan University of Science and Technology, Taiwan
{ m10115081,hmlee,d9807501 } @mail.ntust.edu.tw
2 Institute of Information Science
Academia Sinica, Taiwan
3 Dept. of Information and Computer Engineering
Chung Yuan Christian University, Taiwan
[e-mail]
4 CyberTrust Technology Institute
Institute for Information Industry, Taiwan
[e-mail]
Abstract. Due to the popularity of Twitter, it attracts malicious users'
interests. Most of previous approaches relied on account-based features
such as message similarity between tweets, following-followers ratio, and
so on. Account-based features can be easily manipulated by spam ac-
counts. Spam collusion is a new way to escape the detection mechanisms.
Therefore, we need an advanced mechanism to identify the spam collu-
sion relations.
We exploit spam campaign which spreads spam tweets. We focus on
the tweet with the high retweet count. We create the message-passing
graph via the retweet relations, following relations, and retweet time,
then we extract the time evolution feature in the aspect of graph struc-
ture. The latent behavior indexing technique is used to extract critical
concepts for spam collusion recognition. We collect 5 million tweets from
May 14, 2014 to July 15, 2014 and the ground-truth has been labeled by
domain experts. Our approach can achieve 86% accuracy.
Keywords: spam detection, information propagation, social network.
1 Introduction
In recent years, Twitter has become one of the most popular social networking
sites. It allows users to read and send messages which are less than 140 characters,
known as “tweets”. With tremendous growing availability of Twitter, it also
attracts malicious users' attentions. A previous research [3] showed that nearly
50% of users on social networking sites click on the links which posted by their
friend accounts, even if they do not meet that friend in real life. This has revealed