The Movie Recommendation System suggests movies based on a user's favorite selections. It employs Collaborative Filtering and Content-Based Filtering techniques to provide personalized recommendations.
This system integrates data from multiple datasets. To see detailed building of systems, check the provided Jupyter notebooks.
Watch the system in action! 🎬
- Frontend: HTML, CSS (TailwindCSS for styling)
- Backend: Flask
- Database: SQLite
- Vector Database: Qdrant
- Recommendation Engine:
- Collaborative Filtering (with and without Matrix Factorization)
- Content-Based Filtering
- Deployment: Docker, Render (Initially deployed with Docker, later migrated to Render)
- External APIs: OMDB Movie API (for movie metadata)
- Server: Gunicorn
- Caching: Redis (for API and page performance optimization)
- Implemented Redis to cache API responses and improve performance.
- Reduces database queries and speeds up page load times.
- Ensures efficient data retrieval for frequently accessed resources.
- User selects a favorite movie 🎬
- System retrieves similar movies based on cosine similarity in Qdrant 🔄
- Recommended movies are displayed along with their metadata 🖥️
- Integrated a live analytics dashboard displaying:
- Total Visitors
- Unique Visitors
- Active Users
- Uses Socket.IO to provide real-time updates without refreshing the page.
- Optimized UI with Tailwind CSS for a clean and mobile-friendly design.
Collaborative Filtering is based on user interactions with movies. It uses user ratings to find similarities between movies.
- The dataset contains a user-movie interaction matrix, where:
- Rows represent movies
- Columns represent users
- Values represent ratings given by users to movies
- Each movie is treated as a vector representation
- A similarity score (e.g., cosine similarity) is applied to find the closest movies
📌 Example Representation:
Content-Based Filtering focuses on movie attributes rather than user interactions.
-
The dataset contains movie metadata, including:
- Plot
- Cast (lead actor, actress, etc.)
- Crew (director, writer, etc.)
- Genre
- Year of release
- Theme
- Rating
- Vote Count
-
Different vectorization techniques are applied to each attribute:
- TF-IDF, Count Vectorizer for text-based data (e.g., plot)
- Count Vectorizer for categorical data (e.g., genre, cast, crew)
-
Combined vector representations are used for similarity calculations
📌 Example Representation:
The dataset contains vote count and each rating, so I have calculated one true rating based on the Bayesian Average formula to ensure fair ranking across movies:
[ \bar{R} = \frac{\sum (r_i \cdot v_i) + C \cdot m}{\sum v_i + m} ]
where:
- ( \bar{R} ) = Adjusted rating
- ( r_i ) = Average rating of the movie
- ( v_i ) = Number of votes for the movie
- ( C ) = Mean rating across all movies
- ( m ) = Minimum votes required to be considered
This formula prevents movies with very few ratings from getting an unfairly high or low rank by pulling them toward the global average until they receive more votes.
📌 Example Representation:
Matrix Factorization is used to predict missing values in the user-movie rating matrix, improving recommendations.
- User-movie interaction matrices are sparse (most values are missing)
- Traditional Collaborative Filtering struggles with missing ratings
- Matrix Factorization predicts missing values, improving recommendations
- Decomposes the user-movie matrix into two lower-dimensional matrices
- Predicts missing ratings by reconstructing the original matrix
- Techniques used: Neural Network with Mean Squared Error
📌 Example Representation:
This project is open-source and available under the MIT License.
For questions, suggestions, or collaborations, reach out via:
- 📩 Email: [email protected]
- 💼 LinkedIn: Vedant Koppal
- 🐦 Twitter: @vedantkoppal