PaperlayerView exact revision on arXiv

Source-linked AI summary

Community Interaction and Conflict on the Web

Srijan Kumar, William L. Hamilton, Jure Leskovec, Dan Jurafsky

arXiv:1803.03697v1cs.SIcs.CLcs.HC

TL;DR

Web-community conflicts are poorly understood at the level of their initiation, participant behavior, and effects on users. Using cross-links and activity controls across 36,000 Reddit communities, the paper identifies negative mobilizations, analyzes their dynamics and consequences, tests mitigation-related patterns, and predicts mobilization risk. Conflicts are concentrated among a tiny set of communities, have adverse long-term effects, and can be predicted with an AUC of 0.76.

  • Problem

    Research has provided limited evidence about how intercommunity conflicts start online, how users behave during them, and how these conflicts affect participants, constraining prediction and mitigation.

  • Method

    The study analyzes 137,113 cross-links among 36,000 Reddit communities and uses activity-based null models and sentiment labels to identify negative mobilizations.

  • Results

    Negative mobilizations are concentrated among a small set of communities, adversely affect target-community participation, and are predicted by a socially-primed LSTM with an AUC of 0.76.

  • Takeaways & Limitations

    Increased engagement between attackers and defenders is associated with improved outcomes, while the prediction model could support moderator early-warning systems.

  • Takeaways & Limitations

    The analysis studies pairs of pseudonymous communities on a single platform, so interactions involving more than two communities or real-identity settings may differ.

Abstract

from arXiv · show

Users organize themselves into communities on web platforms. These communities can interact with one another, often leading to conflicts and toxic interactions. However, little is known about the mechanisms of interactions between communities and how they impact users. Here we study intercommunity interactions across 36,000 communities on Reddit, examining cases where users of one community are mobilized by negative sentiment to comment in another community. We show that such conflicts tend to be initiated by a handful of communities---less than 1% of communities start 74% of conflicts. While conflicts tend to be initiated by highly active community members, they are carried out by significantly less active members. We find that conflicts are marked by formation of echo chambers, where users primarily talk to other users from their own community. In the long-term, conflicts have adverse effects and reduce the overall activity of users in the targeted communities. Our analysis of user interactions also suggests strategies for mitigating the negative impact of conflicts---such as increasing direct engagement between attackers and defenders. Further, we accurately predict whether a conflict will occur by creating a novel LSTM model that combines graph embeddings, user, community, and text features. This model can be used toreate early-warning systems for community moderators to prevent conflicts. Altogether, this work presents a data-driven view of community interactions and conflict, and paves the way towards healthier online communities.

1 INTRODUCTION

This study develops a data-driven account of how Reddit communities interact and conflict, addressing limited evidence about how web conflicts begin, unfold, and affect users. It identifies negative mobilizations, characterizes their participants and consequences, evaluates mitigation strategies, and predicts mobilizations for possible moderator early-warning systems.

  • Motivation: Intercommunity conflict research on the web has largely overlooked how conflicts begin, how users behave during them, and how participants are affected.Prior work commonly focused on community detection, within-community interactions, or users’ allocation of time across communities.
  • Approach: The study identifies negative mobilizations when a negatively framed cross-link from a source community mobilizes users to comment in a target community.The method uses cross-links, a null model of expected source-member activity, and crowd-sourced sentiment labels.
  • Findings: 74% of negative mobilizations are initiated by 1% of source communities, while highly active members initiate them and significantly less active users carry them out.Conflict-initiating communities are concentrated in dense regions of a learned Reddit social map, and attackers tend to target a small set of defenders collectively.
  • Findings: Negative mobilizations produce long-term “colonization,” as defenders participate less in target communities while attackers become more active.Reduced echo-chamber effects and greater defender use of anger words toward attackers are associated with lower colonization and higher future defender participation.
  • Implications: More direct engagement and fiercer defense may mitigate conflicts more effectively than ignoring or isolating attacking users.The proposed mitigation follows from observed associations between attacker–defender interaction, defender language, colonization, and later participation.
  • Prediction: The socially-primed LSTM predicts whether a cross-link will cause mobilization with an AUC of 0.76, versus 0.67 for an expert-crafted-feature baseline.The model combines social-network advances with user, community, and text information and could support moderator early-warning systems.

2 DATA AND DEFINITIONS

The study identifies Reddit intercommunity mobilizations from cross-links, matched comparisons, and sentiment classification. It defines participating users and distinguishes negative from neutral mobilizations using observed commenting increases and source-post sentiment.

  • Data: 40 months of Reddit data yield 137,113 cross-links among 36,000 communities for analyzing intercommunity interactions.The dataset covers January 2014 through April 2017.
  • Cross-links: Cross-links connect a source post to a target post and can mobilize source members to comment in the target thread.The analysis treats these links as the initiation step and target-thread commenting as the interaction step.
  • Defining mobilizations: Mobilizations are cross-links producing more source-member comments in target threads than matched posts under a null model.The null model controls for target-post popularity and compares commenting within a 12-hour window before and after linking.
  • Defining mobilizations: 8.8× versus 1.6× matched-thread increases define mobilizations as above-baseline commenting, yielding 22,075 cases, about 16% of cross-links.The operational threshold is an after-to-before increase greater than 1.6×.
  • Classifying sentiment: Negative mobilizations are identified by classifying source-post sentiment as negative versus neutral, using crowd labels and a Random Forest classifier.Two labels were collected for 1,020 pairs with inter-rater agreement above 0.95; positive and neutral labels were merged.
  • Classifying sentiment: The classifier achieves 0.80 accuracy, labeling 8% of source posts negative; the resulting set contains 1,809 negative and 20,266 neutral mobilizations.The classification uses 10-fold cross-validation.

3 INITIATION OF MOBILIZATIONS

Negative mobilizations concentrate in a small number of communities and involve highly similar communities. Initiating posts come from highly active source members, whereas mobilized attackers and defenders are less active and show stronger prior anger expression, with discussions bifurcating into community-based echo chambers.

  • Initiating communities: 38% and 74% of negative mobilizations are initiated by less than 0.1% and 1% of source communities, respectively.These communities function as hubs of negatively mobilizing users.
  • Initiating communities: Negative mobilizations generally occur between highly similar source and target communities, measured using tf-idf post similarity.The analysis evaluates similarity across the interacting community pair.
  • Users involved: Cross-link creators for negative mobilizations are 10% more active in their source communities than matched users and are more active than unsuccessful mobilizers.Both comparisons are statistically significant at p < 0.001.
  • Users involved: Defenders are less active in target communities than matched users, with 0.27 versus 0.19 of past comments, and use 2.2× more anger words.The anger-word rates are 0.32 versus 0.145, with p < 0.001 for both comparisons.
  • Interaction structure: Negative-mobilization discussions form echo chambers: attackers have higher A-PageRank and defenders higher D-PageRank, indicating primarily within-community interaction.The reported PageRank separation marks a bifurcation between attackers and defenders.

4 USER INTERACTION DURING MOBILIZATIONS

Negative mobilizations create polarized reply networks in which attackers and defenders mainly interact within their own communities. Attackers may gang up on a small number of defenders, while direct cross-group interaction is comparatively rare.

  • Attackers interact 2× more with attackers, while defenders interact 20× more with defenders than with the opposing group.These differences are statistically significant (p < 0.001).
  • Negative mobilizations produce an echo-chamber bifurcation, with attackers and defenders primarily talking to members of their own communities.The interaction structure is quantified using A-PageRank and D-PageRank.
  • 83% of defenders have zero A-PageRank, while 1.14% have scores at least ten times the mean across attackers and defenders.The skew indicates that a small set of defenders participates in most attacker–defender interactions.
  • Attackers gang up on some defending users, while other defenders do not engage with attackers.When cross-group exchanges occur, the comments are highly angry.

5 IMPACT OF MOBILIZATIONS

Negative mobilizations can reshape participation in target communities and provide a basis for comparing defensive strategies. Successful defenses are associated with direct, heated attacker–defender engagement that disrupts the usual echo chambers.

  • 5.1 Quantifying impact: After negative mobilizations, attackers post more frequently in the target community while defenders post less frequently (p < 0.001).The authors describe this outcome as “colonization.”
  • 5.1 Quantifying impact: Neutral mobilizations increase mobilized users’ activity in the target community without significantly changing defenders’ activity.The authors call this pattern “immigration.”
  • 5.2 What makes a defense successful?: The paper’s success definition focuses on post-mobilization defender activity rather than the quality of future comments.Alternative definitions involving future comment quality are left for future work.
  • 5.2 What makes a defense successful?: Successful defenses involve more direct replies between defenders and attackers, preventing attackers from ganging up and breaking echo chambers.Success is defined by defenders becoming more active after mobilization ends.
  • 5.2 What makes a defense successful?: Successful defenses have higher defender A-PageRank (0.036 vs. 0.031) and attacker D-PageRank (0.052 vs. 0.028).The defender A-PageRank difference has p < 0.001, as does the attacker D-PageRank difference.
  • 5.2 What makes a defense successful?: Successful defenses show a strong increase in defenders’ replies to attackers (correlation coefficient = 0.97), without a statistically significant increase in defender numbers.Defenders also use more anger words toward attackers than attackers use toward defenders (0.017 vs. 0.015; p < 0.05).

6 PREDICTION OF MOBILIZATIONS

The paper predicts whether cross-links mobilize users by combining textual information with user–community interaction structure in a socially-primed LSTM framework. The combined approach achieves its best reported performance with an AUC of 0.76.

  • 6.3 Socially-primed LSTM model: The socially-primed LSTM combines post text with user and source/target community information for mobilization prediction.Traditional LSTMs use text only; social priming adds user and community information.
  • 6.2 Text, user, and community embeddings: User and community embeddings are learned from a bipartite multigraph recording user posts in communities.Community embeddings become similar when similar users post in them, while user embeddings reflect shared community participation.
  • 6.1 Prediction task: The prediction task is to determine whether a cross-linking post will mobilize users, using 80% training, 10% validation, and 10% held-out test data.The task uses the full set of cross-links rather than restricting examples to negative mobilizations.
  • 6.4 Prediction results: The socially-primed LSTM reaches an AUC of 0.72, compared with 0.67 for the Random Forest baseline and 0.66 for text-only LSTM.A Random Forest ensemble combining hand-engineered features, embeddings, and average LSTM hidden states reaches AUC 0.76.
  • 6.4 Prediction results: The models are intended to support early-warning systems that help moderators predict potential negative mobilizations when cross-link posts are created.The paper presents this as a way to help curb adverse effects.

7 FURTHER RELATED WORK

The paper situates community-level conflict among broader research on online controversies and antisocial behavior. Unlike much prior work, it focuses on interactions between community pairs and their effects on users.

  • Prior online research has examined controversies, trolling, sockpuppetry, harassment, cyberbullying, vandalism, and hate speech.
  • This work differs from those studies by analyzing negative interactions between community–community dyads and their impact.
  • The paper notes that prior research on correction measures has studied approaches such as interventions that reduce hate speech on Reddit.

8 DISCUSSION AND CONCLUSIONS

The paper identifies intercommunity mobilizations at scale, characterizes their adverse effects and correlates of mitigation, and predicts mobilizations using a socially primed LSTM. These findings provide a data-driven methodology for quantifying and managing disruptive interactions, while leaving several platform and interaction-scope limitations.

  • 8 DISCUSSION AND CONCLUSIONS: Cross-links between posts and controls for baseline user activity identify explicit cases of intercommunity mobilization.This methodology provides a large-scale view of interactions and conflict on the web.
  • 8 DISCUSSION AND CONCLUSIONS: Negative mobilizations can produce long-term adverse effects through “colonization,” where ill-behaved users come to dominate target communities.Increased engagement between members of interacting communities correlates with improved outcomes.
  • 8 DISCUSSION AND CONCLUSIONS: A socially-primed LSTM combining user, community, and text embeddings predicts whether mobilizations occur with an AUC of 0.76.The model combines multiple feature types to predict mobilization occurrence.
  • 8 DISCUSSION AND CONCLUSIONS: The analysis studies pairs of communities on one pseudonymous platform, so interactions involving more communities or real identities remain outside its scope.The impact on conversations elsewhere in the community, links to antisocial behavior, and implicit sentiment also remain open questions.
  • 8 DISCUSSION AND CONCLUSIONS: The methodology supports developing policies, strategies, and tools for promoting positive interactions on multi-community platforms.The paper particularly highlights the growing importance of managing highly negative and disruptive communities.
Loading 1803.03697v1…