Detection of content-level collusive activities in online social networks

Abstract

Twitter, as a micro-blogging service has been increasingly used to express opinions, promote brands and share news. Often, the popularity of a tweet is denoted by how other users of the platform are reacting to it. Retweets are a very important aspect of the endorsement of tweets a high number of retweets gain a lot of attention. This has led to the creation of unfair methods for gaining a high number of retweets as a natural way of gaining retweets is very time-consuming { one such shortcut is to approach the black market services and gain retweets for their own tweets by retweeting other customers' tweets. Thus, the users intrinsically become a part of a collusive ecosystem controlled by these services. Previously, we had studied collusion from an unsupervised and semi-supervised setup. Since we achieved great success in our methods by using both network-level as well as attribute-level features, we wanted to see how such a combination would perform in a supervised setup. We do so by generating a vector representation/embedding of the user that is rich with attribute-level as well as network-level features and using it for classifi cation. A collusive retweeting activity can be affected by a variety of activities of the user - following, being followed, liking the content, etc. Thus, we make a rich representation of what constitutes a collusive user - depending on various factors that affect their collusive nature. We propose the application of a concept called Multi-view Embedding that considers the collusive nature of a user from multiple views/perspectives and uses it to create rich embeddings of the user that can be used to predict the collusive retweeting activity of that particular user.