https://avabbas.github.io/academic-cv/tags/nlp/NLPHugoBlox Kit (https://hugoblox.com)en-usMon, 01 Jan 2024 00:00:00 +0000https://avabbas.github.io/academic-cv/media/icon_hu_5ac22c9ce43dded7.pnghttps://avabbas.github.io/academic-cv/tags/nlp/https://avabbas.github.io/academic-cv/projects/social-graph/Mon, 01 Jan 2024 00:00:00 +0000https://avabbas.github.io/academic-cv/projects/social-graph/<h2 id="overview">Overview</h2> <p>A large-scale computational social science project that constructs a <strong>68-dimensional hypergraph</strong> from 10 million real Persian chat messages, modeling the complete social, psychological, and behavioral dynamics of an online community.</p> <hr> <h2 id="key-metrics">Key Metrics</h2> <table> <thead> <tr> <th>Metric</th> <th>Value</th> </tr> </thead> <tbody> <tr> <td>Messages Analyzed</td> <td>10,000,000</td> </tr> <tr> <td>Users (Nodes)</td> <td>7,636</td> </tr> <tr> <td>Relationships (Edges)</td> <td>310,732</td> </tr> <tr> <td>Node Dimensions</td> <td>68</td> </tr> <tr> <td>Edge Dimensions</td> <td>20</td> </tr> <tr> <td>Total Data Points</td> <td>~6.7 million</td> </tr> </tbody> </table> <hr> <h2 id="technical-architecture">Technical Architecture</h2> <ul> <li><strong>GPU Server:</strong> NVIDIA A100 (40GB VRAM) + 128GB RAM + 32 vCPUs</li> <li><strong>NLP Engine:</strong> ParsBERT v3 (fine-tuned on Persian chats)</li> <li><strong>Graph Database:</strong> Neo4j (Cypher queries) + Parquet (analytics) + PyTorch Geometric (GNN)</li> <li><strong>ML Stack:</strong> XGBoost, Random Forest, Gradient Boosting, Ensemble Voting, K-Means</li> <li><strong>Pipeline:</strong> Prefect orchestration, Polars data processing, CUDA acceleration</li> </ul> <hr> <h2 id="9-analytical-layers">9 Analytical Layers</h2> <table> <thead> <tr> <th>Layer</th> <th>Dimensions</th> <th>Description</th> </tr> </thead> <tbody> <tr> <td>🎭 Personality</td> <td>10</td> <td>Big Five traits, gender, age, name authenticity</td> </tr> <tr> <td>🧠 Psychology</td> <td>13</td> <td>Depression (PHQ-9), anxiety (GAD-7), narcissism, bipolar, PTSD, self-harm risk</td> </tr> <tr> <td>🤝 Relational</td> <td>12</td> <td>Deception detection, diplomacy scoring, social masking, genuine vs. fake closeness</td> </tr> <tr> <td>🔗 Network</td> <td>9</td> <td>PageRank, Betweenness, Community Detection (Louvain), Broker Score</td> </tr> <tr> <td>⏱️ Temporal</td> <td>6</td> <td>Chronotype, mood volatility, recovery time, leave prediction</td> </tr> <tr> <td>📝 Content</td> <td>8</td> <td>Lexical richness, sarcasm detection, BERTopic, metaphor usage</td> </tr> <tr> <td>🛡️ Cybersecurity</td> <td>3</td> <td>Bot detection (emoji signature), sockpuppet identification, social engineering risk</td> </tr> <tr> <td>❤️ Digital Health</td> <td>3</td> <td>Screen addiction, sleep deprivation, wellbeing index</td> </tr> <tr> <td>👑 Power &amp; Influence</td> <td>4</td> <td>Opinion leaders, gatekeepers, trendsetters, echo chamber contributors</td> </tr> </tbody> </table> <hr> <h2 id="key-discoveries">Key Discoveries</h2> <ul> <li><strong>&ldquo;Silence Before the Storm&rdquo;:</strong> 5/5 major conflict events were preceded by 10–15 consecutive silence (bradycardia) patterns — a statistically significant early warning signal.</li> <li><strong>Emojis Outperform BERT:</strong> Emoji-based clustering achieves <strong>4.5× better</strong> personality separation than ParsBERT embeddings (Silhouette: 0.265 vs 0.058).</li> <li><strong>Killer Words:</strong> The word &ldquo;bye&rdquo; (بای) caused <strong>676 group-wide silences</strong>, acting as a conversation terminator.</li> <li><strong>Bot Fingerprinting:</strong> Bots exhibit <strong>630× distinct emoji signatures</strong> (✔, ➖, ❓), detectable with <strong>99.2% accuracy</strong> without text analysis.</li> <li><strong>Paradox of Toxicity:</strong> Toxic relationships last <strong>3× longer</strong> than healthy ones.</li> <li><strong>Masked Selves:</strong> Average social masking score is <strong>43%</strong> — users hide nearly half their true personality during daytime interactions.</li> </ul> <hr> <h2 id="100-research-analyses">100 Research Analyses</h2> <p>The hypergraph enables <strong>100 distinct analyses</strong> across 9 layers, with potential for publication in venues such as:</p> <ul> <li><em>Nature Human Behaviour</em></li> <li><em>PNAS / Science Advances</em></li> <li><em>Nature Communications</em></li> <li><em>Journal of Personality and Social Psychology</em></li> <li><em>Computational Linguistics</em></li> <li><em>EPJ Data Science</em></li> </ul> <hr> <h2 id="technologies-used">Technologies Used</h2> <p><code>Python</code> <code>PyTorch</code> <code>ParsBERT</code> <code>HuggingFace Transformers</code> <code>Neo4j</code> <code>Cypher</code> <code>PyTorch Geometric</code> <code>XGBoost</code> <code>Scikit-learn</code> <code>Polars</code> <code>NetworkX</code> <code>Prefect</code> <code>Docker</code> <code>CUDA</code> <code>Plotly</code> <code>BERTopic</code> <code>Leiden Algorithm</code></p> <hr> <h2 id="publication-potential">Publication Potential</h2> <blockquote class="border-l-4 border-neutral-300 dark:border-neutral-600 pl-4 italic text-neutral-600 dark:text-neutral-400 my-6"> <p><strong>100 analyses | 4.6M+ data points</strong></p> <p><em>&ldquo;When Emojis Outperform BERT: A 68-Dimensional Framework for Personality, Mental Health, and Relationship Dynamics in Digital Communities&rdquo;</em></p> </blockquote>