Predicting Large-Scale Internet Censorship - A Machine Learning Approach

Abstract

Social media have an increasingly penetrating effect on our daily lives and entire society. Reviewing on social media research conducted in the past, one important aspect, content deletion due to Internet censorship, has received little direct attention in light of the ongoing media censorship in China. Exposing this aspect of censorship allows citizens to better understand the mechanism of Internet censorship, to help them make informed decisions on how to efficiently participate in society events and in the larger context to maintain a free and open Internet. Our research aims to facilitate a better understanding of social media censorship, and to provide means to automatically detect and predict future content deletion. In this research, a machine learning approach is introduced and applied for this effort. Our research results have revealed vital correlations between the occurrence of real-world political events and online censorship activities as well as public opinion and sentiment expressed; a framework is proposed to predict which microblog will be more likely to be deleted under Internet censorship; and first results are produced. Furthermore, we evaluate model performance by incorporating public sentiment as an aggregate feature in model construction and test the feasibility. As a result, we achieve 95.6% AUC score using naive Bayes algorithm with social features. To our knowledge, this is the first analysis results ever reported in such task.