Automatic Controversy Detection Based on Heterogeneous Signed Attributed Network and Deep Dual-Layer Self-Supervised Community Analysis

In this study, we propose a computational approach that applies text mining and deep learning to conduct controversy detection on social media platforms. Unlike previous research, our method integrates multidimensional and heterogeneous information from social media into a heterogeneous signed attributed network, encompassing various users' attributes, semantic information, and structural heterogeneity. We introduce a deep dual-layer self-supervised algorithm for community detection and analyze controversy within this network. A novel controversy metric is devised by considering three dimensions of controversy: community distinctions, betweenness centrality, and user representations. A comparison between our method and other classical controversy measures such as Random Walk, Biased Random Walk (BRW), BCC, EC, GMCK, MBLB, and community-based methods reveals that our model consistently produces more stable and accurate controversy scores. Additionally, we calculated the level of controversy and computed <i>p</i>-values for the detected communities on our crawled dataset Weibo, including #Microblog (3792), #Comment (45,741), #Retweet (36,126), and #User (61,327). Overall, our model had a comprehensive and nuanced understanding of controversy on social media platforms. To facilitate its use, we have developed a user-friendly web server.