Developing a network-centric approach for anomalous behavior detection on youtube

Abstract As the second most visited website globally, YouTube serves as a central platform for video sharing, entertainment, and information dissemination. However, its expansive and highly active user base also facilitates problematic behavior, particularly among commenters. This study presents a methodology driven by social network analysis to detect and examine anomalous commenter behaviors, with a specific focus on commenter mobs that collaborate to artificially manipulate engagement metrics on videos. Additionally, the study seeks to characterize YouTube channels based on the prevalence of such behaviors, uncovering patterns of coordination among channels. The analysis utilizes a dataset comprising 47 YouTube channels, 26,901 videos, 1,377,902 commenters, and 2,496,558 comments, including 20 channels involved in disseminating misleading information about the U.S. Military and 27 additional channels, which serve as a control group to provide a baseline for normal behavior, helping to distinguish between anomalous and non-anomalous patterns more clearly. The methodology compares principal component analysis (PCA) with Graph2vec and uniform manifold approximation and projection (UMAP), in conjunction with K-means and hierarchical clustering, to identify and categorize anomalous behaviors across channels. Through comprehensive qualitative and quantitative analyses, the study identifies the themes of the videos where these anomalous behaviors occurred in comment sections. These findings provide valuable insights into the dynamics of online discourse and the mechanisms by which coordinated groups influence content and engagement on YouTube.