Abstract: Aiming at the problem that the classifiers have significant bias due to the imbalance of data set for microblog retweet prediction, this paper proposes a microblog retweet prediction method based on multi- task supervised contrastive learning. This method consisted of two tasks: retweet prediction and stance detection. Both tasks extracted relevant features and consisted of two branches: a contrastive learning branch for feature representation learning and a cross- entropy driven branch for classifier learning, where the learning was progressively transited from feature representation learning to classifier learning. The experimental results show that the proposed method can effectively achieve the correct prediction for the samples of "retweet" category, and improve the performance of microblog retweet prediction.