Multi-Stage Two-Tower Recommender

A multi-stage movie recommendation system using YouTube's Tow-Tower architecture.

Getting Started

Installation

To set up the virtual environment and install the necessary dependencies, run the following command:

make install

This will ensure your environment is prepared with all required packages, ready to run the project.

Usage

Run the test.py for a quick start.

Dataset

Different versions of the MovieLens dataset are used for both training and evaluation purposes. Due to significant data redundancy in the original datasets, I created the script dataset.py to generate separate files for movies, users, and ratings, effectively eliminating duplication and reducing the overal size of each datasets along side feature engineering and feature selection for the task. This is most similar to the data at hand in a production environment.

Feature Selection

• Retained all features from the original dataset, except for raw_user_age, which is only available in the 100k version of the dataset.
• Removed user_occupation_text since user_occupation_label was derived from it, making it redundant.

Feature Engineering

• Most of the movie_title values in the dataset include their release years in parentheses, which were extracted and treated as a separate feature. For movies without a listed release year, the corresponding values are left as NaN.
• Converted the user_gender feature from boolean values to integer representations.

Project Structure

• data/ - Directory for storing 100k and 1m versions of the Movielense dataset.
• src/ - Contains the core implementation, including model definitions, architecture, and api functions.
• scripts/ - Useful scripts.

🖇References

Videos

• Building recommendation systems with TensorFlow

Papers

• Deep Neural Networks for YouTube Recommendations

Libraries

• Tensorflow Ranking (TFR)
• TensorFlow Recommenders (TFRS)

Codes

• xei/recommender-system-tutorial

Contributions

Contributions are welcome and greatly appreciated! If you have an idea for improvement, or if you find a bug, feel free to contribute by opening an issue or a pull request.

License

This project is licensed under the MIT License. See the LICENSE file for details.