🧬 68-Dimensional Social Graph: Computational Social Science on 10M Persian Telegram Messages
·
2 min read
Overview
A large-scale computational social science project that constructs a 68-dimensional hypergraph from 10 million real Persian chat messages, modeling the complete social, psychological, and behavioral dynamics of an online community.
Key Metrics
| Metric | Value |
|---|---|
| Messages Analyzed | 10,000,000 |
| Users (Nodes) | 7,636 |
| Relationships (Edges) | 310,732 |
| Node Dimensions | 68 |
| Edge Dimensions | 20 |
| Total Data Points | ~6.7 million |
Technical Architecture
- GPU Server: NVIDIA A100 (40GB VRAM) + 128GB RAM + 32 vCPUs
- NLP Engine: ParsBERT v3 (fine-tuned on Persian chats)
- Graph Database: Neo4j (Cypher queries) + Parquet (analytics) + PyTorch Geometric (GNN)
- ML Stack: XGBoost, Random Forest, Gradient Boosting, Ensemble Voting, K-Means
- Pipeline: Prefect orchestration, Polars data processing, CUDA acceleration
9 Analytical Layers
| Layer | Dimensions | Description |
|---|---|---|
| 🎭 Personality | 10 | Big Five traits, gender, age, name authenticity |
| 🧠 Psychology | 13 | Depression (PHQ-9), anxiety (GAD-7), narcissism, bipolar, PTSD, self-harm risk |
| 🤝 Relational | 12 | Deception detection, diplomacy scoring, social masking, genuine vs. fake closeness |
| 🔗 Network | 9 | PageRank, Betweenness, Community Detection (Louvain), Broker Score |
| ⏱️ Temporal | 6 | Chronotype, mood volatility, recovery time, leave prediction |
| 📝 Content | 8 | Lexical richness, sarcasm detection, BERTopic, metaphor usage |
| 🛡️ Cybersecurity | 3 | Bot detection (emoji signature), sockpuppet identification, social engineering risk |
| ❤️ Digital Health | 3 | Screen addiction, sleep deprivation, wellbeing index |
| 👑 Power & Influence | 4 | Opinion leaders, gatekeepers, trendsetters, echo chamber contributors |
Key Discoveries
- “Silence Before the Storm”: 5/5 major conflict events were preceded by 10–15 consecutive silence (bradycardia) patterns — a statistically significant early warning signal.
- Emojis Outperform BERT: Emoji-based clustering achieves 4.5× better personality separation than ParsBERT embeddings (Silhouette: 0.265 vs 0.058).
- Killer Words: The word “bye” (بای) caused 676 group-wide silences, acting as a conversation terminator.
- Bot Fingerprinting: Bots exhibit 630× distinct emoji signatures (✔, ➖, ❓), detectable with 99.2% accuracy without text analysis.
- Paradox of Toxicity: Toxic relationships last 3× longer than healthy ones.
- Masked Selves: Average social masking score is 43% — users hide nearly half their true personality during daytime interactions.
100 Research Analyses
The hypergraph enables 100 distinct analyses across 9 layers, with potential for publication in venues such as:
- Nature Human Behaviour
- PNAS / Science Advances
- Nature Communications
- Journal of Personality and Social Psychology
- Computational Linguistics
- EPJ Data Science
Technologies Used
Python PyTorch ParsBERT HuggingFace Transformers Neo4j Cypher PyTorch Geometric XGBoost Scikit-learn Polars NetworkX Prefect Docker CUDA Plotly BERTopic Leiden Algorithm
Publication Potential
100 analyses | 4.6M+ data points
“When Emojis Outperform BERT: A 68-Dimensional Framework for Personality, Mental Health, and Relationship Dynamics in Digital Communities”
Authors
Abbas Hamidavi
(he/him)
CS Student & AI Researcher
Open Science advocate
Building tools to probe, analyze, and secure LLM architectures
Exploring the intersection of Red Teaming and Interpretability.
AI Safety Researcher at Islamic Azad University, Mashhad, running an independent research programme in Machine Psychology.
ORCID: https://orcid.org/0009-0001-7342-7844