Are Proactive Interventions for Reddit Communities Feasible?

TL;DR

This study addresses the feasibility of proactive moderation for Reddit by analyzing the evolution of subreddits through fixed-length vocabulary and active-user vectors and by quantifying inter-subreddit dynamics with relative distance measures, notably Rank-Biased Overlap ($RBO$). It demonstrates that subreddit states do not stabilize over time and that evolution into problematic subreddits can be predicted using interpretable ML models trained on features spanning community dynamics, moderation, user behavior, structure, mentions, and language. The authors show strong predictive performance (up to $AUC \approx 0.95$ and $F1$ near 0.95 in some settings) and validate usefulness through a continuous-learning deployment that identifies numerous problematic subreddits well before administrative actions. These results suggest that machine-assisted, human-guided interventions can enable earlier, more nuanced moderation while mitigating the scalability challenges of manual oversight, with implications for similar platforms beyond Reddit.

Abstract

Reddit has found its communities playing a prominent role in originating and propagating problematic socio-political discourse. Reddit administrators have generally struggled to prevent or contain such discourse for several reasons including: (1) the inability for a handful of human administrators to track and react to millions of posts and comments per day and (2) fear of backlash as a consequence of administrative decisions to ban or quarantine hateful communities. Consequently, administrative actions (community bans and quarantines) are often taken only when problematic discourse within a community spills over into the real world with serious consequences. In this paper, we investigate the feasibility of deploying tools to proactively identify problematic communities on Reddit. Proactive identification strategies show promise for three reasons: (1) they have potential to reduce the manual efforts required to track communities for problematic content, (2) they give administrators a scientific rationale to back their decisions and interventions, and (3) they facilitate early and more nuanced interventions (than banning or quarantining) to mitigate problematic discourse.

Figures (1)

• Figure 1: Distribution of RBO distances of vocabulary/user vectors between consecutive months for subreddits in $\mathfrak{D}_{A}$, $\mathfrak{D}_{C}$, and $\mathfrak{D}_{I}$.