Detecting Flagged Comments by Analyzing User Behavior Features in Online Communities

Comments violating community guidelines have long been a challenge in online discussion communities, often disrupting the user experience and overall community health. This study provides a quantitative analysis of such flagged comments within a large online Chinese discussion platform, examining factors like comment deletion rates, sentiment, discussion context influence, user voting behavior, and timing of comment publication. Our dataset comprises 6,900,119 historical comments from 17,226 users, collected and meticulously cleaned for analysis. Feature analysis reveals that users with higher comment deletion rates are more prone to trolling behavior. Additionally, flagged first or root comments are found to increase the likelihood of subsequent flagged comments within a discussion. Flagged comments also tend to attract more negative votes and appear earlier in the discussion. Leveraging these behavioral features, we built a high-accuracy predictive model that achieved an AUC of 99.2% in identifying flagged comments.