GRALA: modeling social information for microblog sentiment analysis from the view of balancing sparsity and smoothness of social contexts

As online social networks provide people with a popular and convenient platform to express opinions, microblog sentiment analysis has become a hot research area. Currently, social relations between microblogs are extracted and integrated into content-based microblog sentiment analysis models. However, these methods take all links between microblogs into consideration, ignoring that some links are noisy and some neighbor microblogs are ambiguous semantically, resulting in the insufficiency of sentiment analysis. To deal with this problem, we propose a method that introduces sparse modeling of social contexts and tries to balance the sparsity and smoothness of social contexts to integrate social information prior knowledge into the content-based models in this paper. Specifically, a microblog graph is built based on sentiment consistency and emotional contagion, then we not only formalize the smoothness assumption of this graph but also introduce a graph-structured sparse regularization term to automatically select microblogs with high quality among all neighbors to facilitate sentiment analysis. Extensive experiments on two real-world public datasets show that our method achieves better performance than other baseline methods, indicating the effectiveness of sparse modeling of graph structure for microblog sentiment analysis.