LinkAlike

LinkAlike aims to build recommender systems using graph link prediction. As many have written about, the problem of product recommendation may be described as the prediction of a future edge in a user-item graph. This graph may be bipartite (i.e. only edges from users to items exist) or we may define edges between users or between items, providing a flavour of content-based recommendation.

This is still a work in progress.

Progress:

• General:
  • Create a benchmark dataset.
  • Create a fast implementation of the pairwise dissimilarity in a set of arrays with categorical data.
• Graph class:
  • Create method to create nodes out of df, df_user and df_item.
  • Create method to create interaction (user-item) edges.
  • Discretize numerical metadata if metadata datasets (df_user and df_item) are given.
  • Create similarity edges if metadata datasets (df_user and df_item) are given.
  • Create the fit method for node embeddings (use node2vec, but leave room for other methods).
  • Create the transform method for node embeddings, with strategy for unseen nodes.
• Recommendation:
  • Create Recommender class, which receives node embeddings and predicts the existente of edges.

Instalation

It is recommended to create a separated python environment to run linkalike. If one chooses to install Miniconda (my personal favorite), an appropriate environment is created and open through the lines

conda create --name linkalike_env python=3.7
conda activate linkalike_env

Then, the packages can be pip-installed from Github,

python -m pip install git+https://github.com/gboaviagem/linkalike@main

or one may choose to simply install its dependencies:

git clone https://github.com/gboaviagem/linkalike
cd gspx
bash install.sh

Running unit tests locally

When unit tests are implemented, one may run using pytest:

python -m pytest --cov=linkalike .

To update the coverage badge, run

rm coverage.svg && coverage-badge -o coverage.svg

Update version in production

Update setup.py version and packages and generate package by running:

python setup.py sdist bdist_wheel

Acknowledgements

The pre-commit hook used to verify codestyle was copied from https://github.com/cbrueffer/pep8-git-hook.