Social Links Enhanced Microblog Sentiment Analysis: Integrating Link Prediction and Sentiment Connection Weights

The emerging microblogging service provides a new channel for people to share opinions and sentiment. As a result, microblog sentiment analysis has become a cutting-edge and popular research field, which has many important applications. Existing methods mostly extract sophisticated features from microblog texts without considering that microblogs are networked data, which suffer from poor performance. To address this issue, we propose a new model that assumes microblogs are interconnected and that connected microblogs are more likely to share the same sentiment. We leverage two types of information to model the connections between microblogs: user information and friend information. Our assumption is supported by two sociological theories: sentiment consistency and emotional contagion. The connections between microblogs based on user and friend information are often sparse and noisy, which can limit the effectiveness of sentiment analysis. To mitigate this issue, we use link prediction to identify potential connections between microblogs and introduce a sentiment connection weights matrix to quantify the degree of sentiment difference between connected microblogs. We then integrate potential social links and sentiment connection weights into our content-based sentiment model using a Laplacian regularization term. To demonstrate the effectiveness, sufficient experiments are conducted on two real datasets to show that exploring potential links and introducing sentiment connection weights can improve the performance of microblog sentiment analysis significantly.