Rumor Detection of Sina Weibo Based on SDSMOTE and Feature Selection

Internet provides great convenience for our life and becomes an important channel to get information. However, a large amount of false information, called rumors, come with it. In terms of automatically detecting rumors, two main contributions of this paper are as follows: (1) To reduce the impact of unbalanced data on classification, we proposed an improvement SMOTE algorithm to resample data. (2) We proposed six new features based on Sina microblogs, including Words with Guidance (WG), Words with Menace (WM), Suspected Topic (ST), Recognition of Information (RI), Degree of Attention to Users (DAU) and Credit Rating (CR), which are related to user-based features, content-based features, propagation-based features and microblog-based features. By building subsets with new features and using machine learning algorithms including Xgboost etc. We tested the effect of rumor detection on a real data set. Experiments showed that our rumor detection method was significantly improved compared with the most advanced method of the same type, with precision, recall and F1 at 0.827, 0.837 and 0.825 respectively, and AUC at 0.895.