Microblog Sentiment Analysis has become a popular research topic extensively examined in the literature. However, microblogging messages are usually short, unstructured, contain less information and much noise, creating a significant challenge for the application of traditional content-based methods. In this study, we propose a novel method, MSA-USSR, where user similarity information and interaction-based social relations information are combined to build sentiment relationships between microblogging data. We employ these microblog–microblog sentiment relations to train the sentiment polarity classifier. The experimental results on two Sina-Weibo datasets show that our model has a better sentiment classification accuracy and F1-score than the Support Vector Machine method and the state-of-the-art supervised model called SANT. In addition, it was proven that the improvement in accuracy brought by interaction-based social relations information is greater than the user similarity information, but MSA-USSR achieved the best performance when incorporating both user similarity information and users’ interaction-based social relations.