Prediction of Cyberbullying Incidents on the Instagram Social Network

Homa Hosseinmardi, Sabrina Arredondo Mattson, Rahat Ibn Rafiq, Richard Han, Qin Lv, Shivakant MishraarXiv 2015 · arXiv

Introduction

Cyberbullying has become a pervasive issue on social media, affecting over half of American teenagers and leading to severe psychological consequences such as depression and suicide ideation. While previous research has focused on text-heavy platforms like Twitter, Instagram presents a unique challenge as a media-based network where the context of an attack is often split between a visual post and a stream of comments. This research introduces a multi-modal framework designed to move beyond reactive detection toward proactive prediction of cyberbullying incidents.

An example of an Instagram media session illustrating the distinction between cyberaggression (red underlines) and defensive or neutral comments (blue underlines).

Methodology

The researchers distinguish between cyberaggression (isolated aggressive remarks) and cyberbullying (repeated patterns of aggression involving a power imbalance). To address this, they developed a multi-stage approach:

  1. Data Acquisition: Using snowball sampling via the Instagram API, the team collected data from 25,000 public profiles, resulting in over 3 million "media sessions" (an image plus its associated comments).
  2. Feature Engineering: Features were extracted across four dimensions: linguistic (using LIWC to categorize word usage), temporal (analyzing comment frequency), social graph (measuring likes and follower counts), and image content (categorizing images into themes like "Drugs" or "Selfies").
  3. Classification: Linear SVM and Logistic Regression models were used for detection, while a Maximum Entropy (MaxEnt) classifier was employed for the predictive task.

Dataset and Labeling

The study utilized a novel, manually labeled dataset of 2,218 Instagram media sessions. Crowdsourced contributors labeled these sessions by reviewing both the image and the comments to ensure contextual accuracy.

The distribution of human labels (k votes) for cyberaggression and cyberbullying across the media sessions, showing that bullying is a rarer, more specific phenomenon.

As shown in the labeling distribution, cyberbullying is significantly less frequent than general aggression, highlighting the importance of precise definitions in training data.

Key Insights

  • Linguistic Patterns: In bullying sessions, users are significantly more likely to use third-person pronouns (he, she, they) and less likely to use first-person singular (I). This suggests that bullies often "talk about" the victim to an audience rather than addressing them directly.
  • Visual Triggers: Certain image categories are strong indicators of potential conflict. For instance, images labeled as "Drugs" had a 75% likelihood of being associated with cyberbullying, whereas categories like "Food" had almost zero correlation.
  • Social Power Imbalance: Victims of cyberbullying tended to have higher follower counts but lower "likes" per post compared to non-victims, providing a measurable metric for social imbalance.

Results and Performance

The framework was tested on two primary tasks:

  • Detection (Reactive): Using text-based N-grams with a Linear SVM, the researchers achieved 71% precision and 79% recall in identifying existing bullying incidents.
  • Prediction (Proactive): By analyzing only the image content and user metadata at the moment of posting (before any comments arrived), the MaxEnt classifier predicted future bullying with 76% recall and 62% precision. In this predictive context, non-textual features were the primary drivers of accuracy.

Conclusion

This research demonstrates that cyberbullying can be predicted with reasonable accuracy at the time of content upload. By integrating visual and social metadata, social networks can move toward proactive prevention. Future work will focus on automating image labeling through computer vision and identifying the role of "defenders" who support victims in comment sections.

Summary updated

BibTeX
@misc{hosseinmardi2015prediction,
  author = {Homa Hosseinmardi and Sabrina Arredondo Mattson and Rahat Ibn Rafiq and Richard Han and Qin Lv and Shivakant Mishra},
  title = {{Prediction of Cyberbullying Incidents on the Instagram Social Network}},
  year = {2015},
  eprint = {1508.06257},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1508.06257},
}
Figure