Hybrid Movie Recommender System

---

Project Overview

This project implements a Movie Recommender System using Content-Based Filtering, Collaborative Filtering, and a Hybrid Recommendation approach. The system focuses on high-rated movies released between 2015 and 2025 and leverages both movie metadata and user–movie interaction data to generate accurate and personalized recommendations.

---

Objectives

• Build a content-based recommender using movie metadata
• Build a collaborative filtering model using user ratings
• Combine both approaches into a hybrid recommender system
• Evaluate recommendation quality using ranking-based metrics

---

Dataset Description

Data Sources

• TMDB API – Movie metadata (titles, overview, genres, cast, keywords)
• MovieLens Dataset – User–movie rating interactions
• IMDb (Scraping) – Scraping pipeline implemented (reviews optional)

Filtering Criteria

• Release years: 2015–2025
• Minimum rating: 6.0
• Minimum vote count: 50

Generated Files

File Name	Description
movies_2015_2025.csv	Raw movie metadata from TMDB
movies_tmdb_enriched.csv	Enriched metadata (genres, cast, keywords)
ratings_for_cf.csv	User–movie ratings for collaborative filtering
movies_master.csv	Final dataset with combined textual features

---

Methodology

1. Content-Based Filtering

• Feature engineering using movie overview, genres, keywords, cast
• Unified text feature: combined_text
• TF-IDF vectorization
• Cosine similarity for movie-to-movie recommendations

2. Collaborative Filtering

• User–item interaction matrix from MovieLens
• Matrix factorization (SVD) using Surprise library
• Personalized rating prediction

3. Hybrid Recommendation

• Weighted fusion of content-based and collaborative scores
• Fusion parameter α controls contribution of each model

Final Score = α × Content Score + (1 − α) × Collaborative Score

---

Evaluation Metrics

The system was evaluated using ranking-based and prediction metrics:

• HitRate@5
• HitRate@10
• MRR@5
• RMSE (for collaborative filtering)

Performance Highlights

• HitRate@10 ≈ 0.95
• HitRate@5 ≈ 0.92
• MRR@5 ≈ 0.90
• Best performance achieved at α ≈ 0.7–0.9
• Hybrid model significantly outperformed collaborative-only (Surprise) baseline

---

Tools and Technologies

• Programming Language: Python
• Libraries: Pandas, NumPy, Scikit-learn, Surprise
• Data Sources: TMDB API, MovieLens
• Web Scraping: BeautifulSoup
• Visualization: Matplotlib
• Environment: Kaggle, Jupyter Notebook
• Version Control: Git, GitHub

---

Ethical Considerations

• Only public and academic datasets were used
• Ethical and rate-limited web scraping
• No personal or sensitive user data collected
• Project developed strictly for educational purposes

---

Future Work

• Deep learning–based recommenders (Neural CF, embeddings)
• Real-time user feedback integration
• Web or mobile application deployment
• Larger and multilingual datasets

---

Authors

• Awais Asghar
• Ahmad Hussain
• M. Hammad Sarwar

---

⭐ Acknowledgments

• TMDB for providing the movie metadata API
• GroupLens Research for the MovieLens dataset

---

⭐ If you find this project useful, feel free to the repository ⭐

---