Semi-supervised internet water army detection based on graph embedding

Sina Weibo is one of the most popular online social networks, which provides users with a convenient and fast way to communicate. However, its openness and large user base are often exploited by water armies to spread disinformation. These water armies seriously affect the security and reliability of social networks. Existing approaches on the water army detection mainly classify users by their account information, behavior and other features. However, the increasing anti-detection capability of water armies has led to insufficient differentiation of detection features. In addition, there is a lack of labeled data for model training. In this paper, we propose a semi-supervised approach combining network structure features and user attribute features for identifying the Internet water army. This approach uses the graph embedding algorithm to obtain network structure features of users, which together with the defined user attribute features constitute the detection feature set. Users are classified by Tri-Training, a semi-supervised learning algorithm, which leverages the synergistic advantages of multiple classifiers and reduces the need for a large number of labeled data. Experiments on real-world data illustrate that our approach can identify the Internet water army effectively, and it is more suitable for real scenarios with less labeled data. The accuracy can reach 95.15%.