Abstract: The rapid broadcasting of hot events by new media makes the phenomenon of public opinion reversal occur from time to time. It is important to identify the influence factors of public opinion reversal, which can help predict the reversal of public opinion at the beginning of public events. Successful prediction of public opinion reversal can help emergency management departments predict and guide the development trend of public opinion in time. It also helps enhance the credibility of the media and maintain the healthy development of the online ecological environment. We crawled the hot microblog posts of the topics about 38 hot events on Sina Weibo from 2017 to 2020. Based on the previous research, we proposed the public opinion reversal prediction models that consisted of 30 features from four aspects of events, users, information, and dissemination. The XGBoost technique was used to calculate the importance of different features in the public opinion reversal prediction model. The Logistic Regression, Decision Tree, Random Forest, XGBoost and Gaussian Naive Bayes techniques were used to construct the public opinion reversal prediction models. The models were trained and evaluated in the experiment to find out the optimal prediction model. The experimental result showed that the information balance, the type of the user who exposes the event, and the event type had the most significant impact on the prediction performance of the public opinion reversal model. The prediction models based on the Random Forest and XGBoost techniques achieved the best performance among all the five public opinion reversal models. Suggestions were also made on the discrimination and governance of public opinion reversal from three aspects, i.e., media, the public and social media platform.

Key words: Public opinion reversal, Public opinion prediction, Hot events, Public opinion governance, Microblog analysis, Machine learning