Predicting Topic Participation by Jointly Learning User Intrinsic and Extrinsic Preference

Understanding the preferences of social media participants plays a crucial role in many business applications. A specific aspect of interest is to predict which topics a particular user is more likely to be involved in. Existing efforts on such topic participation forecasting mainly focus on learning from historical user-generated texts to infer their preferred topics, or leverage on an information propagation theory on networks to predict the topics of potential interest. However, jointly utilizing both the sources of data to provide a holistic prediction for such a task has not been exploited. We present a novel joint learning framework that takes advantage of both users’
intrinsic
(learned from their past social media postings) and
extrinsic
preference (learned from the social network influence) and that embeds them into low-dimensional space vectors. To facilitate effective learning, we encode latent continuous embedded vectors into binary ones via locality-sensitive hashing. Furthermore, to explain the predictions made by our “black-box” model, we investigate the importance of each training sample on the topic prediction performance of a testing instance to demonstrate its interpretability. The experiments conducted on datasets collected from several popular social media platforms demonstrate the effectiveness of our proposed method when compared with existing baselines.