Content-Driven Detection of Cyberbullying on the Instagram Social Network

Haoti Zhong, Hao Li, Anna Cinzia Squicciarini, Sarah Michele Rajtmajer, Christopher Griffin, David J. Miller, Cornelia CarageaIJCAI 2016

Content-Driven Detection of Cyberbullying on the Instagram Social Network

Introduction

As social media has become increasingly visual, the landscape of cyberbullying has shifted from text-based forums to platforms like Instagram. Traditional automated detection methods, which rely primarily on natural language processing (NLP), often fail to capture the full context of harassment on these platforms. This research presents a multimodal approach that integrates image analysis, caption metadata, and comment text to identify and even anticipate cyberbullying with high accuracy.

The Challenge of Visual Context

On Instagram, the image is the primary catalyst for interaction, yet it is frequently ignored by detection algorithms. The authors identify two major challenges in the current landscape:

  1. Contextual Bullying: A comment may appear benign in isolation but becomes abusive when paired with a specific image or provocative caption.
  2. Early Warning: The potential to predict if a post is "bully-prone" before the abuse even begins, allowing for proactive moderation.

Methodology

The study employs a supervised learning framework using a Support Vector Machine (SVM) with an RBF kernel. The novelty of the approach lies in its comprehensive feature engineering:

  • Textual Features: Beyond standard Bag of Words (BoW) and Word2Vec, the researchers developed an Offensiveness (OFF) score. This score uses a parser to identify grammatical dependencies between second-person pronouns and offensive words.
  • Visual Features: The team utilized a pre-trained Convolutional Neural Network (CNN) to extract high-level image features. They also tested traditional computer vision metrics like SIFT (keypoint descriptors) and Color Histograms.
  • Caption Analysis: Using Latent Dirichlet Allocation (LDA), the researchers extracted 50 distinct topics from image captions, treating the caption as a vital piece of context that often "sets the stage" for the comments section.

Key Results and Multimodal Synergy

The researchers found that combining these disparate data sources significantly improves detection performance. When identifying whether a comment contains bullying (post-hoc detection), the multimodal model achieved a remarkable 93.21% accuracy.

Precision-Recall curves for different feature sets, demonstrating that the Concatenate Classifier significantly outperforms individual text or caption features.

In the more difficult task of predicting if an image will be bullied—using only the image and caption before any comments are posted—the model achieved 68.55% accuracy. A significant finding was that LDA-generated caption topics were a surprisingly powerful predictor, often outperforming complex image-only features in the anticipation task.

A comparison of various classification strategies, including Deep Learning and Stacked Classifiers, showing the performance trade-offs in detection accuracy.

Insights and Implications

One of the most significant insights is that captions are often stronger predictors of bullying than the images themselves. An innocent photo can become a target if the caption includes provocative language or references to sensitive topics. Additionally, the authors noted a "Celebrity Effect," where standard user metadata (like follower counts) was less predictive due to the unique social dynamics of high-traffic accounts.

Conclusion

By bridging computer vision and social science, this work demonstrates that multimodal analysis is essential for effective moderation on visual social networks. While anticipating human malice remains a "noisy" problem, the 68% accuracy in prediction offers a strong foundation for real-time intervention tools that can help mitigate abuse as it happens.

Summary updated

BibTeX
@inproceedings{zhong2016content,
  author = {Haoti Zhong and Hao Li and Anna Cinzia Squicciarini and Sarah Michele Rajtmajer and Christopher Griffin and David J. Miller and Cornelia Caragea},
  title = {{Content-Driven Detection of Cyberbullying on the Instagram Social Network}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure