Jointly Learning Sentiment, Keyword and Opinion Leader in Social Reviews

Social review sites offer a wealth of information beyond sentiment polarity. For instance, on IMDb users leave valuable reviews on different aspects of a movie (e.g. Actors, visual effects). This inspires researchers to fully discover information from social review texts for the sake of modelling users behavior. Previous studies have spent a large amount of effort to identify sentiment scores from reviews. Yet questions like "What are the key plot in this movie?" and "Who is the valuable user I should follow?", the answers of which comprehensively support user decision making process, can not be answered in those works. To jointly learn from sentiment, text and user in social reviews, we draw inspiration that only a small portion of reviewers can generate useful information, and propose a sparse overlapping user lasso model to tackle these challenges. In addition, we show how to efficiently solve the resulting optimization challenges using the alternating directions method of multipliers (ADMM), a framework which divides our objective into sub-tasks that are easy to fulfill. By experimenting several experiments on 3 real world social review datasets, we demonstrate that our method consistently outperforms other state-of-the-art models in sentiment classification tasks, meanwhile generating accurate results on keywords discovering and opinion leader identification task.