Research map / Graph learning / Knowledge and temporal graphs
TSGAN: Temporal Social Graph Attention Network for Aggressive Behavior Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TSGAN, a social-aware seq2seq graph attention model that forecasts user aggression intensity on X (Twitter) using social influence and decay-aware temporal attention, with a hybrid RoBERTa+LLaMa-2 aggression detector feeding its inputs.
Problem
Forecasting individual user aggressive behavior in dynamic online social networks, which prior temporal graph models and aggression detection/diffusion work do not address, and which existing graph forecasting models fail to scale to large networks.
Approach
An adaptive socio-temporal attention module (ASTAM) combines social influence attention over active peers (with BGRL global context embeddings), decay-aware temporal attention, a gated integration step, and cross-temporal attention linking historical and future steps in a seq2seq encoder-decoder.
Key results
- TSGAN reports up to 24.8% improvement over baselines in daily aggression prediction on X, and 18% and 3.2% gains for hourly and weekly horizons.
- The hybrid aggression detector (AG-BERT RoBERTa-large plus LLaMa-2 for low-confidence cases) reaches 92.87% F1, versus 91.85% for the prior Mane et al. (2023) model.
- TSGAN outperforms STGCN, the next best deep learning model, with a reported t-test p-value < 0.01 across intervals.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TSGAN instead of the strongest listed baseline, RNN | X aggression forecasting · X aggression, hourly, 1H → 1F | MAE: 0.911 ± 0.001 -> 0.029 ± 0.001 | Helped |
| use TSGAN instead of the strongest listed baseline, RNN | X aggression forecasting · X aggression, daily, 1H → 1F | MAE: 0.378 ± 0.006 -> 0.138 ± 0.003 | Helped |
| use TSGAN instead of the strongest listed baseline, STGCN | X aggression forecasting · X aggression, weekly, 1H → 1F | MAE: 0.239 ± 0.007 -> 0.218 ± 0.023 | Helped |
| use TSGAN instead of the strongest listed baseline, RNN | Flickr popularity forecasting · Flickr SMPD popularity, weekly, 1H → 1F | MAE: 0.367 ± 0.002 -> 0.114 ± 0.043 | Helped |
| replace LLaMA-2 with Falcon in the hybrid content detector (ablation) | AG-BERT + LLaMA-2 · Aggression content detection | F1: 0.9287 -> 0.9274 | Hurt |
| replace LLaMA-2 with Gemma in the hybrid content detector (ablation) | AG-BERT + LLaMA-2 · Aggression content detection | F1: 0.9287 -> 0.921 | Hurt |
| replace LLaMA-2 with LLaMA-3 in the hybrid content detector (ablation) | AG-BERT + LLaMA-2 · Aggression content detection | F1: 0.9287 -> 0.9249 | Hurt |
| remove SIA (ablation) | TSGAN · X aggression forecasting, hourly, daily, and weekly horizons | MAE | Hurt |
| remove TDA (ablation) | TSGAN · X aggression forecasting, hourly, daily, and weekly horizons | MAE | Hurt |
| remove IAG (ablation) | TSGAN · X aggression forecasting, hourly, daily, and weekly horizons | MAE | Hurt |
| remove CTA (ablation) | TSGAN · X aggression forecasting, hourly, daily, and weekly horizons | MAE | Hurt |
Surprising findings
- Traditional sequence models and some graph models such as GRU and LSTM produce near-constant high errors on the sparse aggression series, and DCRNN performs worst in many settings.
Limitations
- Baseline graph models (ASTGCN, AGCRN, Graph WaveNet, GMAN, MTGNN) could not be run at this network size, so comparisons are limited.
- Robustness evaluated only with synthetic uniform noise injected into AGI values, not real detector errors.
- Aggression labels depend on a detector trained on a single dataset and an LLM prompt, with no reported hardware or code release.
Experimental setup
Datasets: X aggression dataset (Mane et al. 2025) Social Media Prediction Dataset (SMPD, Flickr)
Benchmarks: X aggression forecasting dataset SMPD Flickr popularity prediction
Metrics: MAE RMSE Precision Recall F1
Baselines: ARIMA RNN GRU LSTM LSTNet STGCN DCRNN Akter et al. (2022) Ghosh et al. (2023) Rawat et al. (2023) Mane et al. (2023)
Builds on: Graph Attention Networks BGRL Transformer attention DCRNN STGCN
Closest papers by meaning
- Adversarial Socialbots Modeling Based on Structural Information Principles (AAAI 2024 · similarity 0.85)
- TGB-Seq Benchmark: Challenging Temporal GNNs with Complex Sequential Dynamics (ICLR 2025 · similarity 0.84)
- DAG-Aware Variational Autoencoder for Social Propagation Graph Generation (AAAI 2024 · similarity 0.83)
- THGNets: Constrained Temporal Hypergraphs and Graph Neural Networks in Hyperbolic Space for Information Diffusion Prediction (AAAI 2025 · similarity 0.82)
- DyG-Mamba: Continuous State Space Modeling on Dynamic Graphs (NeurIPS 2025 · similarity 0.82)
- Spear and Shield: Adversarial Attacks and Defense Methods for Model-Based Link Prediction on Continuous-Time Dynamic Graphs (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.