🏙️ Mashhad Social Approach Study — AI-Powered Behavioral Analysis

Mon, 20 Jul 2026 00:00:00 +0000

Overview

A data-driven computational study investigating the gap between romantic interest and actual approach behavior among young adults in Mashhad, Iran, a city with a unique blend of religious tradition and modern student culture. Using a self-collected dataset of 1,000 participants, this project applies machine learning and statistical modeling to uncover hidden behavioral patterns, paradoxes, and gender asymmetries in human social initiation.

🔍 Focus

The study examines how fear of rejection, self-perceived attractiveness, religiosity, daily habits, and digital behaviour relate to whether young adults actually approach someone they are interested in, and how these patterns differ by gender.

🧪 Methodology

Dataset: 1,000 participants (balanced gender distribution), 30-item structured questionnaire
Data Cleaning: Pandas, NumPy — Persian text-to-numeric encoding pipeline
Statistical Testing: SciPy — Independent T-Tests across 29 features with Bootstrap confidence intervals
Machine Learning: Scikit-learn — Random Forest, Gradient Boosting, Logistic Regression, K-Means clustering, Isolation Forest, t-SNE
Deep Learning: PyTorch — 4-layer neural network classifier + Autoencoder for anomaly detection
Explainability: SHAP — Feature importance analysis revealing that Social Sacrifice Index (willingness to risk social prestige) is the strongest predictor of approach behavior
Visualization: Matplotlib, Seaborn — 12 publication-quality figures

📊 Results & Impact

Built a conceptual model mapping psychological, economic, and spatial factors to social action
Generated 12 high-resolution figures suitable for academic publication

🛠️ Technologies Used

Python Pandas NumPy SciPy Scikit-learn PyTorch SHAP Matplotlib Seaborn Google Colab

💡 Personal Contributions

Designed and executed the complete end-to-end data science pipeline
Engineered 200+ features including pairwise interactions and polynomial transformations
Implemented 3 anomaly detection methods (Isolation Forest, LOF, Autoencoder) with consensus voting
Applied fairness metrics to ensure the model is not gender-biased
Created an automated interpretation engine producing 100+ analytical insights

Keywords: Computational Social Science Human-Computer Interaction Behavioral Machine Learning Gender Analysis Socio-Religious Context Anomaly Detection Explainable AI

🧬 68-Dimensional Social Graph: Computational Social Science on 10M Persian Telegram Messages

Mon, 01 Jan 2024 00:00:00 +0000

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”

PyTorch | Abbas Hamidavi | عباس حمیداوی | AI Researcher

🏙️ Mashhad Social Approach Study — AI-Powered Behavioral Analysis

Overview

🔍 Focus

🧪 Methodology

📊 Results & Impact

🛠️ Technologies Used

💡 Personal Contributions

🧬 68-Dimensional Social Graph: Computational Social Science on 10M Persian Telegram Messages

Overview

Key Metrics

Technical Architecture

9 Analytical Layers

Key Discoveries

100 Research Analyses

Technologies Used

Publication Potential