project for fast retrieval of movie recomendations via approximate nearest neighbor searches in embedding vector space using user-to-movie, user-to-user, movie-to-user, and or movie-to-movie similarities. There is also a cold-start list made from calculating the bayesian average of all movie ratings.

The embeddings were trained by a listwise contrastive bi-encoder in the project: https://github.com/nking/recommender_systems.git

instructions: set up a virtual environment using conda or virtualenv with a python version that is >= 3.10.0

activate the virtual environment

to install the dependencies, the easiest way is to install this project: pip install --editable . else you can find the required libraries in pyproject.toml or setup.py

the unit tests show how to run the code.

Local testing:

pycharm:

using right click menu, mark the source tree directory:
 src/main/python
using right click menu, mark the test tree directory:
 src/test/python/movie_lens_retrieval
then pycharm tests will correctly resolve paths.

bash or other shell environment:

python and pytest can be used from the project's base
directory

Misc: the cold start, bayesian shrinkage could redone regularly. the code is in the recommendations project. could use for the "m" estimate of 0.75 quantile, data sketches like q-digest or t-digest on live data.