Detecting Spam on Twitter via Message-Passing Based on Retweet-Relation - Technologies and Applications of Artificial Intelligence - page 58

2.3 Individual Detection Mechanisms
Stringhini et al. [16] proposed some features which are following and follower
ratio, the frequency of user that posted tweets including URL, message similarity
between tweets, friend choice (whether the user used a list of names to pick its
friends or not), the number of messages sent and friend number. Shekar et al. [14]
used specific keywords to filter the pharmaceutical spam on Twitter. Thomas et
al. [18] detected link spam on Twitter messages via a real-time system. Lee and
Kim [10] developed a WarningBird system to discover correlated URL redirect
chains. Once the malicious links be detected, it will be reported and added to
the blacklist. So spammer could not use this link to acquire victims.
Most of these works detected Twitter spam accounts individually, instead of
focusing on observing collusive spammers. They needed to renew effective fea-
tures at regular intervals so that can quickly adjust features to spam accounts'
new behaviors. Therefore, detecting spammers individually is not the best mech-
anism.
2.4 Collusive Spammers Detection Mechanisms
To find out more robust features, there are some researchers began to look at
the issue from another side. Yardi et al. [25] designed ten new detection features
including graph-based features (local clustering coe cient, betweenness central-
ity, and bi-directional links ratio), neighbor-based features (average neighbors'
followers, average neighbors' tweets, and followings to median neighbors' follow-
ers), automation-based features (the usage frequency related to Twitter API)
and timing-based feature (the speed an account follows others). Song et al. [15]
devoted to the relations between spam senders and receivers such as the short-
est paths, minimum cut and random walk. The reason why they choose those
features was spammers usually can not establish robust relationships with their
neighbors.
Most of the methods relied on accounts' following and follower relations. If
Twitter changes their detecting strategies, spammers may change their actions.
Hence, we try to find out the feature of spam accounts' campaigns that can not
be simply transformed.
3 Message-Passing Graph Analyzer
3.1 Concept of Proposed Methodology
Many researches [9,4,6] focused on analyzing retweeting behavior of users on
Twitter. The concept of spam collusion was proposed by Ghosh et al. [7] and
Yang et al. [24] in recent year. But both of Ghosh et al. and Yang et al. did
not consider the spammers' automatically retweet behavior. Therefore, we use
retweet relations and the time evolution features.
Since the spam flooding, Twitter defined rules to stop spammers growth rate.
The first policy is to limit users to aggressively follow others. That means an