A graph-based Github recommendation system for machine learning repositories

Task

Implement a graph-based recommendation system for GitHub. Particularly we recommend repositories for a target user that he may find useful or even want to contribute to, by utilizing his previous starred repositories.

Project setup

Create virtual environment

$ python3 -m venv venv
$ source venv/bin/activate

Upgrade pip

$ python -m pip install --upgrade pip

Install dependencies

$ pip install -r requirements.txt

Architecture

The overall architecture of the recommendation system is shown below:

Dataset

For the constructed dataset we utilize the GitHub API to gather users that have given a star (stargazers) to specific repositories.

To end up with a relatively dense graph the criteria for the seed repositories (i.e. the repos from which we will gather the corresponding users) are the following:

1. Topic is “Machine Learning”
2. Main language of repositories is “Python”
3. Each repository has at least 1000 stars.

The results are 250 seed repositories.

Those seed repos are available at data/repos.json

For these repos we gather using the API the corresponding stargazers.

From all the collected users, we keep only the top users according to their given stars to the seed repos, in three flavors: top-3000, top-6000, top-9000.

For the complete dataset with the repositories and their corresponding users please contact the contributors of this repository.

Recommendation algorithm

After constructing the bipartite graph from the above dataset, we implemented a random-walk algorithm. Particularly given a target user for which we want to recommend new repositories, we find the neighbors of the user that correspond to the repositories that have already received a star by the user. These repositories are set as the seed repositories for the recommendation and will be used as the starting point of the random walks that will be conducted over the bipartite graph.

The parameters of our algorithm are random_walks_per_repo and double_steps_per_random_walk.

The evaluation is based on the Precision and Recall metrics.

Demonstration of the system

The k value is the parameter for requesting top-k repositories to be recommended for a target user. Particularly by modifying the target users and the topk value in recommendation/random_walk.py you can request recommended machine learning repositories for those target users that they haven't already starred.

To run it:

$ cd recommendation/
$ python random_walk.py

Example output:

For GitHub user: fly51fly, the top-10 recommended repositories are ['datamade/usaddress']
    
For GitHub user: gaomingweig, the top-10 recommended repositories are ['deepfakes/faceswap', 'flairNLP/flair', 'RasaHQ/rasa', 'stanfordnlp/stanza', 'microsoft/nni', 'microsoft/nlp-recipes', 'iterative/dvc', 'alteryx/featuretools', 'streamlit/streamlit', 'lanpa/tensorboardX']
    
For GitHub user: izdi, the top-10 recommended repositories are ['PyTorchLightning/pytorch-lightning', 'microsoft/nni', 'scikit-learn/scikit-learn', 'ddbourgin/numpy-ml', 'ageitgey/face\_recognition', 'tensorflow/tensor2tensor', 'pyro-ppl/pyro', 'RasaHQ/rasa', 'JaidedAI/EasyOCR', 'PaddlePaddle/Paddle']