A multimodal framework for modeling and predicting information propagation in online social networks

Abstract In recent years, popular Online Social Networks (OSNs) including Twitter, Facebook, Weibo etc. have shown exponential growth which leads to huge amounts of information being posted online in various formats such as textual and multimedia messages. The reposting of such information is supported by all popular OSNs and acts as key mechanism for driving information propagation by amplifying the information reach for OSN users. The feature of reposting information has received a lot of attention in recent years, such as modeling user preference to predict whether a user will repost the information or not. The existing studies on information propagation lack in providing complete coverage of various feature modalities which are necessary to accurately model and predict information diffusions patterns. The likelihood that a certain piece of information will be reposted further or not? significantly depends on the semantic nature of content, user to user engagement processes, and the influence between users within their social circles. The task of modeling and predicting information propagation becomes more challenging due to the interdependence between various heterogeneous multimodal features including content semantics, social interactions, temporal activity patterns, and some additional contextual factors. This study proposes a multimodal information propagation framework that incorporates attributes from diverse feature modalities which are responsible for modeling and predicting retweet behavior on Twitter. This proposed multimodal framework not only offers complete coverage of the diverse features accountable for retweet behavior but also presents these features in a unified fashion to support autonomous learning of the hidden patterns for predicting tweet popularity. To empirically validate the proposed framework a comprehensive experimental study is designed where two different Deep Learning (DL) model architectures are evaluated using a newly constructed Twitter dataset. The experimental results revealed that the proposed framework is comprehensive which covers diverse feature modalities while achieving sufficient performance figures to accurately predict retweet behavior on Twitter.