🎬 Movie Recommendation System (Content-Based)

A lightweight content-based movie recommendation system built using TMDB metadata. It extracts key movie features, processes them into vectors, computes similarity scores, and serves the recommendations through an interactive Streamlit UI.

---

🚀 Features

• Content-based filtering using cosine similarity

• Metadata extraction:

  • Genres
  • Keywords
  • Top 3 cast members
  • Director
  • Overview

• Stemming of normalized text (NLTK)

• Vectorization via CountVectorizer

• Fast, precomputed similarity matrix

• Supports downloading large models from Hugging Face

• Clean, interactive Streamlit interface

---

📁 Project Structure

.
├── app.py                   # Streamlit UI application
├── main.ipynb               # Data preprocessing + similarity computation
├── movies.pkl               # Cleaned movie metadata
├── requirements.txt         # Dependencies for app & notebook
├── runtime.txt              # Python version (for Streamlit Cloud)
├── README.md                # Documentation
└── data/
    ├── tmdb_5000_movies.csv
    └── tmdb_5000_credits.csv

---

🧠 System Architecture

 TMDB CSVs
    │
    ▼
 Data Cleaning & Feature Extraction (main.ipynb)
    │
    ├── create movies.pkl
    └── compute similarity matrix → similarity.pkl
    │
    ▼
 Streamlit App (app.py)
    │
    ├── load movies.pkl
    ├── load similarity from local OR Hugging Face
    └── recommend top similar movies

---

🛠️ Local Setup

1️⃣ Create a virtual Environment

python -m venv venv

2️⃣ Activate the Environment

source venv/Scripts/activate  # Windows
source venv/bin/activate      # Macos 

3️⃣ Install Dependencies

pip install -r requirements.txt

4️⃣ Run the Streamlit App

streamlit run app.py

Your app will open at:

👉 http://localhost:8501

---

🧪 Regenerating Artifacts (optional)

Open and run:

main.ipynb

This notebook:

• Cleans the TMDB dataset

• Creates a tags column

• Computes cosine similarity

• Saves artifacts:

  • movies.pkl
  • similarity.pkl or similarity.npz (depending on your notebook code)

☁️ Deploying on Streamlit Cloud

1. Push the repo to GitHub

(NO large model files like similarity.pkl — use HF for that)

2. Add these two files:

runtime.txt

python-3.10.12

requirements.txt (example)

streamlit
numpy==1.25.3
pandas
scikit-learn
joblib
requests

3. Ensure app.py contains your HF URL:

HF_RAW_URL = "https://huggingface.co/<username>/<repo>/resolve/main/similarity.pkl"

4. Deploy

Go to: https://share.streamlit.io

• New App
• Choose your GitHub repo
• Branch: main
• Entry point: app.py
• Deploy 🎉

Streamlit will download the model from Hugging Face on first run.

---

🌐 Hosting the similarity model on Hugging Face

Upload similarity.pkl to your HF repo.

Use the raw URL:

https://huggingface.co/<username>/<repo>/resolve/main/similarity.pkl

⚠️ Do NOT use the blob link — it won’t work. Use /resolve/main/ or /raw/main/.

---

🧠 Recommendation Logic

movie_index = movies[movies['title'] == movie].index[0]
distances = similarity[movie_index]
movie_list = sorted(
    list(enumerate(distances)),
    reverse=True,
    key=lambda x: x[1]
)[1:6]
return [movies.iloc[i[0]].title for i in movie_list]

---

❗ Common Issues & Fixes

🔥 GitHub rejecting large files?

GitHub doesn’t allow >100MB. Solution:

• Upload large similarity file to Hugging Face
• Let app.py download it at runtime

🔥 Streamlit build failing on numpy?

Use a wheel-friendly version:

numpy==1.25.3

Add runtime.txt:

python-3.10.12

🔥 Using private Hugging Face files?

Add a token to Streamlit Secrets.

---

🌟 Future Enhancements

• TMDB poster integration
• Movie detail pages
• Hybrid recommender (Content + Collaborative Filtering)
• Semantic similarity with Sentence Transformers
• Compressed sparse similarity matrix

---

❤️ Acknowledgements

• TMDB for the dataset
• Streamlit for the UI
• Hugging Face for large file hosting
• Scikit-learn & Pandas for preprocessing