TSGAN: Temporal Social Graph Attention Network for Aggressive Behaviour Forecasting

In an era where online aggression increasingly spills into real-world harm—from cyberbullying to offline violence—the ability to predict aggressive behavior before it escalates has become a societal imperative. TSGAN (Temporal Social Graph Attention Network), an AI model that doesn’t just detect aggression but forecasts it, giving platforms and policymakers crucial lead time to intervene.

The Innovation: What Makes TSGAN Unique?

1. First-of-its-Kind Forecasting

   • TSGAN is the first model designed to predict individual user aggression on social networks, addressing a critical gap in proactive content moderation.
   • Prior work focused on detection or static diffusion; TSGAN answers the question:
     “Who will turn aggressive tomorrow?”

2. Core Architecture

   • ASTAM (Attention-based Social Temporal Aggregation Module):
     An attention module that models social influence (active peer interactions) and temporal decay (how past aggression fades or resurges).
     Think of it as a “dual-lens camera” capturing both who you interact with and when you interacted.
   • Global Network Context Embedding (GNCE):
     Uses contrastive learning to embed social hierarchies (e.g., influencers vs. passive users), ensuring predictions respect real-world power dynamics.

3. Hybrid Aggression Detection

   • Combines a fine-tuned RoBERTa transformer with LLaMa-2’s reasoning to detect aggressive content at 92.87% F1-score—critical for accurate forecasting.

Why TSGAN Works: Results That Matter

• 24.8% Better Daily Predictions:
  On X (Twitter), TSGAN outperformed models like STGCN and Transformers in forecasting hourly/daily/weekly aggression.
• Scalability:
  Processes 30K+ users efficiently by focusing on active subgraphs—a game-changer for platforms like Facebook or TikTok.
• Robust to Noise:
  Even with 60% incorrect aggression labels, TSGAN’s predictions stayed reliable, mimicking real-world detector imperfections.
• Beyond Aggression:
  Achieved SOTA on Flickr for user popularity prediction, proving its adaptability to diverse social tasks.

The Bigger Picture: Why This Matters

• Proactive Moderation:
  Platforms can flag high-risk users before they post harmful content, enabling targeted interventions (e.g., warnings, support resources).
• Policy Impact:
  TSGAN’s forecasts could inform regulations on platform accountability, helping lawmakers prioritize “prevention over reaction.”
• A New Lens for Social Science:
  By quantifying how aggression propagates, TSGAN opens doors to studying behavioral contagion at scale.

Behind the Scenes: Technical Highlights

• Dynamic Attention:
  Unlike static graph models, TSGAN’s attention mechanisms adapt to who’s active and what’s relevant in real time.
• Decay-Aware Design:
  Recognizes that a user’s aggression isn’t just about how much they posted, but when and with whom.
• Open Challenges:
  While trained on Twitter/Flickr, TSGAN’s architecture is platform-agnostic—ready to tackle emerging networks like Mastodon or Bluesky.

The Road Ahead

TSGAN isn’t just a technical milestone—it’s a step toward healthier online ecosystems. Future work includes:

• Multimodal Forecasting:
  Adding text/images to improve prediction granularity.
• Global Deployment:
  Partnering with platforms to stress-test TSGAN in diverse cultural contexts.

In the fight against online harm, TSGAN shifts the paradigm from reacting to anticipating—a vision where AI doesn’t just moderate content but helps prevent it.

Did you find this page helpful? Consider sharing it 🙌