Integrating Social Homophily and Network Lasso for Personalized Sentiment Classification

People often have different opinions toward the same topic, posing challenges for traditional sentiment analysis systems that rely on statistical models estimated with population-level data. Personalization is a solution to the problem, but it suffers from data scarcity issue, as the labeled date for a specific user is usually only available in a small quantity. In this paper, we exploit the theory of social homophily to overcome this difficulty by assuming that social neighbors are likely to express similar opinions with respect to the same topic or document. We present a personalized sentiment analysis system that leverages the social relation information to balance between personalization and overfitting. Our approach is based on network lasso that automatically clusters socially proximate users together, so that users in the same group share the same personalized model. We develop a distributed optimization algorithm that makes our system scale to large networks as well. The experiments on Yelp review datasets show that our approach consistently outperforms competitive methods.