User Level Multi-Feed Weighted Topic Embeddings for Studying Network Interaction in Twitter

Over half a billion tweets on a wide range of topics are posted daily by hundreds of millions of Twitter users. Insights of user behavior and network interactions can be applied to practical applications like targeted advertising, viral marketing, political campaigns, etc. In this paper, we propose a Multi-Feed Weighted Topic Embeddings (MFWTE) model to study user network interaction and topic diffusion patterns on Twitter. Our method extracts topic embeddings from multiple views of a Twitter user feed and weights them according to their content authoring roles, where the authored tweets, replied tweets, retweeted tweets, and favorited tweets are the views we separate for constructing the embeddings. We test the proposed method using two different topic modeling algorithms: (i) Latent Dirichlet Allocation (ii) Twitter-Latent Dirichlet Allocation. The users in our study are divided into multiple hierarchies based on their activity composition regarding individual topics, and the effectiveness of MFWTE is evaluated in the multi-hierarchical setting. The performance of our method on friendship recommendation and retweet behavior prediction task is evaluated using various ranked retrieval measures. The results indicate that our MFWTE method for topic modeling of Twitter users improves over various previous baselines. We conclude our work by applying the proposed model, MFWTE to discover various information diffusion patterns on Twitter