SentiTube — YouTube Sentiment Analysis

Analisis sentimen komentar YouTube menggunakan AI multilingual (XLM-RoBERTa).

🚀 Quick Start (Development)

1. Setup Environment

# Copy env file dan isi API Key
cp .env.example .env
# Edit .env → isi YOUTUBE_API_KEY

2. Jalankan Backend

cd backend
pip install -r requirements.txt
uvicorn main:app --reload --port 8000

3. Jalankan Frontend

cd frontend
npm install
npm run dev
# Buka http://localhost:5173

---

🧠 Fine-tuning Model (Penting!)

Lihat model/README_training.md untuk panduan lengkap.

Cara cepat (Google Colab — Gratis):

1. Upload model/train.py ke Google Colab
2. Aktifkan GPU: Runtime → Change runtime type → T4 GPU
3. Install deps: !pip install transformers datasets accelerate scikit-learn -q
4. Jalankan: !python train.py --output_dir ./fine_tuned_model --epochs 3
5. Download hasil dan extract ke backend/fine_tuned_model/

Tanpa fine-tuned model, backend otomatis pakai model pre-trained cardiffnlp/twitter-xlm-roberta-base-sentiment

---

🐳 Deploy dengan Docker

# Pastikan fine_tuned_model/ sudah ada di backend/
# Pastikan .env sudah diisi

docker compose up --build -d

# Frontend: http://localhost:80
# Backend API: http://localhost:8000

---

📁 Struktur Project

YoutubeSentimentAnalysis/
├── backend/           # FastAPI API server
│   ├── main.py        # Entry point + routes
│   ├── services/      # Scraper, filter, model, cache
│   ├── fine_tuned_model/  # ← Taruh model di sini setelah training
│   └── Dockerfile
├── frontend/          # React + Vite UI
│   ├── src/
│   │   ├── pages/     # Home, Results
│   │   └── components/# Chart, WordCloud, Stats, dll.
│   └── Dockerfile
├── model/             # Fine-tuning pipeline
│   ├── train.py       # Training script
│   ├── evaluate.py    # Evaluasi model
│   └── README_training.md
├── docker-compose.yml
└── .env.example

---

🔑 YouTube API Key

1. Buka Google Cloud Console
2. Buat project baru
3. Enable YouTube Data API v3
4. Buat API Key di Credentials
5. Masukkan ke .env: YOUTUBE_API_KEY=your_key

---

Features

• ✅ Scraping komentar YouTube (hingga 10.000)
• ✅ Filter otomatis: spam, promo link, emoji-only
• ✅ Sentimen multilingual: Indonesia + Inggris
• ✅ Donut chart interaktif
• ✅ Statistik detail (total, waktu, kecepatan)
• ✅ Top 5 komentar representatif per sentimen
• ✅ Word cloud terpisah positif/negatif
• ✅ Cache hasil (1 jam per video)
• ✅ Non-blocking UI (background task + polling)
• ✅ Docker ready