TSGAN: Temporal Social Graph Attention Network for Aggressive Behavior Forecasting

Swapnil Mane, Suman Kundu, Rajesh SharmaAAAI 2025 · DOI

Aggressive behavior on social media is a critical precursor to real-world violence and psychological distress. While significant research has focused on detection—identifying hostile content after it is posted—the task of forecasting future aggression remains a formidable challenge. Predicting which users will lash out next requires understanding complex social dynamics and individual behavioral patterns over time.

In research presented at AAAI-25, Swapnil Mane, Suman Kundu, and Rajesh Sharma introduced TSGAN (Temporal Social Graph Attention Network). This novel architecture moves beyond simple text analysis to model how a user's social circle and personal history interact to drive future hostility.

The Challenge of Aggression Forecasting

Forecasting aggression is inherently more difficult than detection due to three primary factors:

  1. Temporal Sparsity: Aggressive acts are relatively rare events in a user's overall timeline, making it difficult for standard time-series models to identify predictive patterns.
  2. Social Duality: Behavior is driven by both social influence (peer pressure and environmental toxicity) and temporal history (individual habits). Balancing these two forces is essential for accuracy.
  3. Scalability: Real-world social networks involve tens of thousands of active participants, often exceeding the capacity of traditional Graph Neural Networks (GNNs).
Conceptual framework of social influence and temporal dynamics in aggressive behavior.

Architecture: The TSGAN Framework

TSGAN is a sequence-to-sequence (seq2seq) architecture designed to handle the complexities of social media data. At its core is the Adaptive Socio-Temporal Attention Module (ASTAM), which dynamically integrates social and temporal signals.

Overview of the TSGAN architecture, featuring the Adaptive Socio-Temporal Attention Module (ASTAM).

Core Components of ASTAM:

  • Social Influence Attention (SIA): Unlike standard attention mechanisms that treat all neighbors equally, SIA prioritizes active peers. It recognizes that an inactive friend has negligible immediate influence on a user's current behavior.
  • Temporal Dynamics Attention (TDA): This module employs an exponential decay factor to prioritize recent behavior while maintaining the ability to "recall" relevant distant past events if they resurface contextually.
  • Integration Attention Gateway (IAG): A learnable gating mechanism that balances SIA and TDA. For highly social users, the gate prioritizes peer influence; for more isolated users, it leans on individual historical patterns.

To capture the broader social hierarchy, the model utilizes Graph Contrastive Learning (BGRL) to generate static embeddings for the global network context.

Experimental Setup and Data Labeling

The researchers evaluated TSGAN on two large-scale datasets:

  • X (formerly Twitter): 30,307 users and over 500,000 edges, tracking interactions over six months.
  • SMPD (Flickr): 38,312 nodes and 312,063 edges, used to test the model's ability to forecast user popularity as a secondary metric.

To generate high-quality ground truth labels for training, the authors developed a hybrid system. A fine-tuned RoBERTa model handles clear-cut cases of aggression, while LLaMa-2 acts as a high-level reasoner for ambiguous, low-confidence tweets. This approach achieved a 92.87% F1 score in aggression detection, providing a robust foundation for the forecasting task.

Key Results and Insights

TSGAN consistently outperformed traditional deep learning (GRU, LSTM) and graph-based (STGCN, DCRNN) baselines across multiple time horizons:

  • Performance Gains: The model achieved improvements of 18% in hourly, 24.8% in daily, and 3.2% in weekly aggression forecasting over the best-performing baselines on the X dataset.
  • Activity-Centric Modeling: A key insight was that a user's active social circle is a far stronger predictor than total connectivity. By filtering out "dead" nodes, the model improved both accuracy and computational efficiency.
  • Resilience to Noise: TSGAN demonstrated remarkable robustness, maintaining superior performance even when 60% of training labels were intentionally corrupted. This suggests the model captures structural social patterns rather than merely memorizing keywords.

Conclusion

TSGAN represents a significant step forward in proactive social media moderation. By effectively balancing social influence with temporal dynamics through the ASTAM module, it provides a scalable and accurate solution for forecasting aggressive behavior. Future work aims to scale the architecture to millions of nodes and incorporate multi-modal data, such as images and memes, to further refine its predictive capabilities.

Summary updated

BibTeX
@inproceedings{mane2025tsgan,
  author = {Swapnil Mane and Suman Kundu and Rajesh Sharma},
  title = {{TSGAN: Temporal Social Graph Attention Network for Aggressive Behavior Forecasting}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35045},
}
Figure