Most YouTube videos trend for just one day. TubeScope predicts which ones will go viral for longer. Using machine learning and the YouTube Data API, we analyze channel metrics, video metadata, and temporal patterns to identify videos with sustained trending potential, catching 2 out of 3 viral hits before they peak.
Goal: Predict which YouTube trending videos will achieve sustained virality (>1 day trending) using machine learning, helping content creators and marketers identify high-potential content before it peaks.
- Viral Video Detection - Predicts which videos will trend >1 day
- Daily Data Collection - Automated YouTube API pipeline
- ML Model - Random Forest with 68.33% recall, 49.48% precision
- Survival Analysis - Kaplan-Meier curves showing dropoff patterns
- Interactive Dashboard - Streamlit web application
- Production Ready - Trained models with proper feature encoding
| Component | Technology |
|---|---|
| API & Data | YouTube Data API v3 (videos.list, videoCategories.list, channels.list) |
| Storage | Daily CSV snapshots |
| Analysis | pandas, numpy, matplotlib, seaborn, lifelines |
| Web App | Streamlit |
| Version Control | Git + GitHub |
| Predictive Modeling | scikit-learn (Random Forest Classifier) |
| Metric | Value | Interpretation |
|---|---|---|
| Accuracy | 82.35% | Overall correct predictions |
| Precision | 48.83% | 2.9x better than 17% baseline |
| Recall | 68.33% | Catches 2 out of 3 viral videos |
| F1-Score | 57.4% | Balanced performance |
Key Insight: We optimized for recall to catch viral videos early, accepting slightly lower precision as a trade-off for early trend detection.
TubeScope takes a hybrid approach: we train on a Kaggle dataset (for time limitations) and test daily pulls on Streamlit.
| Source | Description |
|---|---|
| YouTube Data API v3 | Fetches stats (views, likes, comments, etc.) for trending videos using videos.list |
| Kaggle Dataset | Kaggle Dataset Contains 10,000 historical data from the Youtube trending data (downloaded Nov 20th, static dataset) |
| Trending Videos Feed (U.S.) | Pulled via API access pull_trending.py |
| Category Lookup Table | videoCategories.list endpoint used to tag videos (Music, Gaming, News, etc.) |
Each day’s trending data is stored as a timestamped snapshot, forming a custom time-series dataset of trending videos and their lifecycle.
- Refresh Data - Click button to fetch today's 50 U.S. trending videos
- View Top 5 - See videos with highest viral probability
- Deep Dive - Select any video from dropdown to see detailed analysis
- Get Insights - Understand which features drive viral predictions
- Customize - Customize your own inputs to see how well a video can survive the "24-Hour Cliff"
- Data Collection - Daily YouTube API pulls (50 trending videos + channel metrics)
- Feature Engineering - Extract channel authority, video metadata, temporal patterns.
- ML Prediction - Random Forest Classifier (trained on 10k historical videos)
- Live Dashboard - Streamlit displays ranked predictions and video stats
Model Details:
- Training: Kaggle dataset (10k videos) with stratified 80/20 split
- Features: Channel metrics (subscribers, videos, views), video metadata (duration, tags, description word count), temporal patterns (upload time, day of week)
- Optimization: Class-weighted for imbalanced data (83/17 viral/standard ratio)
Visualize: View results interactively via Streamlit dashboard
Streamlit dashboard: TubeScope Streamlit
- Duration is the #1 predictor
- Categories such as Gaming, Music, and Entertainment tend to perform better
- Channel size/metrics heavily dictates a video's virality
- Upload timing matters, posting at 12PM is stronger than at 2AM
- Started with regression → R² = 0.95 (data leakage!)
- Fixed leakage → R² went negative (too noisy)
- Kaplan-Meier analysis → 83% dropoff at 24 hours
- Pivoted to classification → 68.33% recall, 48.83% precision ✅