Abstract
Online antisocial behavior, such as cyberbullying and harassment, threatens free discussion and has severe mental health consequences. While automated methods exist to identify hostile comments, they typically work retrospectively. This paper proposes a proactive approach: forecasting future hostilities. We decompose this into two tasks: (1) predicting if a hostile comment will appear given a sequence of clean comments, and (2) predicting if an initial hostile comment will lead to an escalation. Using a new corpus of 30K annotated Instagram comments, our approach achieves an AUC of 0.82 for predicting hostility presence ten hours in advance and an AUC of 0.91 for distinguishing between high and low levels of future hostility.
Introduction: From Detection to Forecasting
Most current automated moderation tools operate in a "reactive" loop. They identify harassment or hate speech only after it has been posted and viewed, meaning victims are already exposed to toxic content before intervention occurs. By the time a system flags a comment, the interaction has often already escalated.
This research shifts the focus toward proactive forecasting. By treating hostility as a predictable event rather than a random occurrence, moderators can potentially intervene before a conflict begins. The study defines two primary forecasting challenges:
- Hostility Presence Forecasting: Predicting whether a currently "clean" thread will attract a hostile comment in the future.
- Hostility Intensity Forecasting: Predicting whether a single hostile comment will devolve into a high-volume "conflict spiral."
Methodology and Feature Engineering
The researchers utilized a Logistic Regression model with L2 regularization. To ensure interpretability and robustness, they focused on sophisticated feature engineering rather than complex neural architectures. The feature set included:
- Linguistic Features: Standard unigrams and Word2vec embeddings, including n-gram Word2vec to account for the typos and abbreviations prevalent in social media text.
- Lexical Signals: Features derived from established databases like "Hatebase" and offensive word lists.
- Social and Contextual Features: The model analyzed user activity (ratio of unique users to total comments) and trend features (the trajectory of hostility probability within a specific thread).
- Historical Context: A key innovation was tracking the recent behavior of the post's author and current commenters in other threads to determine if the general "vibe" of their recent interactions predicted future conflict.
Dataset Characteristics
The study introduced a manually annotated corpus of 30,987 Instagram comments across 1,134 posts. The sampling strategy targeted the teenager demographic, as they are particularly vulnerable to cyberbullying. To handle the natural class imbalance of social media (where most comments are benign), the researchers used keyword-based snowball sampling to identify potentially hostile threads, balancing them with non-hostile posts from similar populations.
As shown in the data distribution above, the dataset captures a wide range of engagement levels. While the median post receives 14 comments, the distribution is long-tailed. Notably, the median time until the first hostile comment appears is only one day, though the mean is 13 days, suggesting that while some conflicts ignite quickly, others simmer over longer periods.
Key Insights and Results
The model demonstrated strong predictive power across both tasks:
- Task 1 (Presence): The system achieved an AUC of 0.82 with a lead time of 10 hours. This indicates that a future fight can be reliably predicted nearly half a day before it begins.
- Task 2 (Intensity): Predicting whether a post would receive ten or more hostile comments was even more successful, reaching an AUC of 0.91. This suggests that once a conflict starts, it follows a highly predictable social pattern.
The "Outsider" and "Target" Effects
One of the most interesting findings was the role of user history. Features derived from the history of current commenters actually reduced accuracy for predicting the start of a conflict. This is because the person who initiates hostility is often an "outsider" who has not yet participated in the thread.
Conversely, the history of the post's author was highly predictive. Certain users consistently receive more hostility based on their past interactions or the nature of their content, making the author's profile a strong signal for potential future harassment.
Limitations and Future Directions
While the results are promising, the researchers noted several areas for growth. The dataset's reliance on keyword seeding may miss subtle, coded, or sarcastic harassment that avoids profanity. Additionally, as Instagram is a visual platform, future models should incorporate image analysis to see if specific visual content triggers hostile responses.
Finally, the transition from forecasting to prevention remains an open question. Future research must explore how "soft" interventions—such as slowing comment frequency or alerting authors—can effectively disrupt the predicted escalation of hostility.
Summary updated