Team ThunderCoding (벼락코딩)
Project Period: Sep 25, 2023 – Feb 1, 2024
Project Type: Youth-led Real-World Experience Project
Developed By: Hanbin Im, Jeonghoon Jeong
PredicTube is a web service that leverages artificial intelligence to predict YouTube video performance based on title, thumbnail, and metadata before the video is uploaded.
It also provides data-driven optimization suggestions, such as recommended titles and keyword insights, for creators and marketers.
- YouTube’s Ubiquity: 81% of Korea’s population watches YouTube monthly (KOSIS, 2022), making it a dominant platform for content and marketing.
- Creator Saturation: With increasing creators and limited viewer attention, titles and thumbnails have become crucial for competitiveness.
- Lack of Objective Guidelines: There are few tools that provide data-based support for optimizing these elements.
PredicTube helps creators and marketers make better decisions with AI-based performance prediction and analytics.
- YouTube Creators seeking growth, engagement, and monetization
- Video Marketing Specialists and Brands using YouTube as a promotional tool
- Predict video performance (view count range) based on title, thumbnail, subscriber count, and category.
- Input: title, thumbnail (image), subscriber count → Output: predicted view range.
- Extracts core keywords from user input and generates 3 optimized titles using ChatGPT API.
- Saves each user’s past prediction attempts and results for comparison and refinement.
- Shows variable-wise correlation with view count.
- Includes "simple" and "detailed" views with toggle support.
- Presents top-ranking keywords by category to support keyword planning and content ideation.
- Frontend: JSP, HTML5, CSS3, JavaScript
- Backend: Flask (Python), Tomcat
- Database: MySQL
- AI Model: Python (Keras, TensorFlow)
- APIs:
- YouTube Data API v3
- Google Cloud Vision API (OCR, face detection, NSFW scoring)
- ChatGPT API
-
Data Collection:
- ~100K videos crawled via YouTube Data API
- OCR + face detection + safety detection via Google Cloud Vision API
-
Preprocessing:
- Tokenization, padding, stopword removal, text scaling, embedding
-
Model Design:
- LSTM for title & thumbnail text
- MLP for numerical metadata (subscriber count, length, faces, etc.)
- Combined MIMO (Multi-Input Multi-Output) architecture
-
Evaluation:
- Category-specific models
- Hyperparameter tuning and visual performance comparison
- Main Page
- Google Social Login
- Channel ID Registration
- Thumbnail Upload via Drag & Drop
- Prediction Result Page
- Interactive Graphs & Keyword Rankings
- User History Modal with Thumbnail Previews
- Service Info & FAQ
- Bug Reporting Page
- Developer Intro Page
Compared to tools like Noxinfluencer:
| Feature | PredicTube | Noxinfluencer |
|---|---|---|
| AI-based prediction | ✅ | ❌ |
| Title optimization (AI-generated) | ✅ | ❌ |
| Thumbnail analysis (OCR/face) | ✅ | ❌ |
| Keyword frequency analysis | ✅ | ❌ |
| User history management | ✅ | ❌ |
- Beta Deployment: Host the service externally and collect feedback.
- Business Expansion:
- Paid version with improved AI model and advanced analytics
- Collaboration with influencers and agencies
- Use Cases:
- Creators for performance optimization
- Brands for ad campaign effectiveness
- Researchers for behavioral and psychological analysis
| Name | Role | GitHub |
|---|---|---|
| Hanbin Im | Planning, Full-stack Dev, AI Modeling, UI/UX, API Integration | @Hanbeeen |
| Jeong Jeonghoon | Data Management, Visualization Design | – |
This repository is for educational and prototyping purposes.
Commercial licensing inquiries welcome upon request.