When our life is becoming more and more inseparable with the social network, our judgment also becomes easier to be affected by the online community. In the virtual Internet, when people see the opinion they are agree with is widely welcomed, they will actively participate in. On the other hand, if they find that there is no one or only few people care about a certain opinion, even if they are agree with this opinion, they will keep silent. Nowadays, a large amount of paid posters exist in social network are hired to affect the others opinions. These accounts, so-called water army, are often hired by PR companies to post specific content in social network and get reward.

As the former work mainly detect water army from individual account perspective, this work investigates int...[ Read more ]

When our life is becoming more and more inseparable with the social network, our judgment also becomes easier to be affected by the online community. In the virtual Internet, when people see the opinion they are agree with is widely welcomed, they will actively participate in. On the other hand, if they find that there is no one or only few people care about a certain opinion, even if they are agree with this opinion, they will keep silent. Nowadays, a large amount of paid posters exist in social network are hired to affect the others opinions. These accounts, so-called water army, are often hired by PR companies to post specific content in social network and get reward.

As the former work mainly detect water army from individual account perspective, this work investigates into collective detection approach. We clustered the posts collected from a water army activity into groups and label them based on their nature: normal user groups, water army groups and traditional advertise groups. Five group features, such as tweet interval of time (TIT), group member diversity (GMD) and three group member features, such as group reputation value (GRV) are used as attributes of machine learning process for classifying. Our strategy succeeds at detecting much of the water army groups while only a small percentage of normal user groups are misclassified. Approximately 70% of water army groups and 98% of normal user groups were correctly classified. Our results also highlight the most important attributes for water army detection on Sina micro-blog. In the end, we also use the same approach to detect traditional advertise groups from normal user groups and the result clearly indicates the difference between water army groups and traditional advertise group.