jaeyow/matrix_factorization_recommender

End to end ML recommender using Matrix Factorization (for educational use only)

Matrix Factorization Recommender

This repo contains code to generate data and train an ML recommender using a Matrix Factorization library called implicit

• we have identified that we can leverage the Alternating Least Squares (ALS) algorithm to build this recommender system as described in the paper Collaborative Filtering for Implicit Feedback Datasets.

• We will generate our own dataset similar to the method used in Cevo Shopping Demo

• Our dataset does not have any rating information (dataset does not have explicit feedback), so we will use the implicit feedback method to generate the recommendations.

• There is a python library called implicit which implements this algorithm as defined in the paper. We will use this library instead of implementing the algorithm from scratch.

• This model is online serving (meaning that the predictions are done in real time), and because we are hosting this model in a serverless in AWS Lambda, it may not be grunty enough for a larger dataset.

• The MLOps pipeline, we're using Metaflow in this project, will take care of model building, hyperparameter tuning, model evaluation, then model serving. We will deploy the model in a Lambda function, and we will use the API Gateway to expose the model as an API endpoint.

Tech Stack

• GitHub Actions
• Metaflow
• AWS S3
• Comet ML for experiment tracking
• SAM for Lambda deployment
• GridSearch CV (or manual grid search)
• Model Registry (?)
• Feature Store (?)

To run:

1. python recommender.py show
2. python recommender.py run