Retweet Prediction Using a DNN Model With Community-Based Attention

One of the modes of information dissemination on Twitter micro-blogging platform is via "Retweet". Ideas, comments and information posted by an author can be relayed further across spheres. Prediction of retweet is an important field, especially due to the recent spotlight on the role of social media in dissemination of ideologies and the creation of echo chambers. In this work, we propose the use of attention mechanism on a retweet graph-embedding to focus on community networks of Twitter users. Along with a generated profile vector for tweet author and the current user, as well as a generated word embedding from tweet content, we also generate community embeddings from a retweet graph containing social context among the author, user and user's neighbourhood. The community embeddings form the input to the attention layer, the output of which is reduced by a CNN layer to become an influence vector. All of these form a concatenated input into a fully connected network, capped by a softmax layer to produce a binary classification for the retweet prediction. We then compare our proposed model against several baseline models, evaluate against a balanced dataset and specific feature sets, both implicit and explicit. Our results show that the use of attention mechanism to incorporate social graph based on retweet behaviour is promising and provides a positive yield for recall and f1-scores compared to baseline prediction approaches.