Recommender

A coursework for TIETS43 Recommender Systems at University Of Tampere.

General

The purpose of this coursework is to implement recommender system which recommends open source software repositories for user based on his Github metadata.

Github metadata currently include:

Repository languages and topics from

users' own repositories
starred repositories
followed repositories

Data also includes both the README and LICENSE -files in base64-encoded format.

Requirements

Anaconda or Minicoda (or whatever Python distribution you like, but these are easiest, though)
Git (if you want to clone the repository, otherwise you can just download the repository as ZIP-archive)
Create new environment by running conda env create --name recommender --file requirements.txt in the repository root directory.
Run source activate recommenderto activate the environment

And now you're ready to use or contribute to the Recommender. For using the dataset retrieval and prerpocessing tools, see "Tools"-section of this README.

Repository search

Using Github search-API, one can search repositories by license. Following query searches for repositories that uses one of these licenses:

BSD 2-clause
BSD 3-clause
MIT

The query:

?q=license:bsd-3-clause+license:bsd-2-clause+license:mit

This should be way more than enough, since the query shows there are more than 2 700 00 repositories meeting this criteria.

Full API query:

https://api.github.com/search/repositories?q=license:bsd-3-clause+license:bsd-2-clause+license:mit

Tools

There are four Python scripts that can be used to retrieve and vectorize the dataset. They are listed here and described below:

user.py
data.py
vectorize.py
preprocess.py
helper.py (contains helper methods used in user and data retrieval scripts)

Before using these, do the following:

Make an environment variable called GITHUB, where you put your Github access-token.

Login to your Github-account
Go to https://github.com/settings/tokens
Create your personal access token
Make an environment variable called GITHUB and store your token there

Users

File user.py contains code to retrieve repository metadata for user to whom we're going to recommend other repositories. Result is saved in JSON-format.

Run python code/user.py github-username on your terminal and it will report what it is doing finally saving the results to github-username.json. For example:

$ cd recommender
$ python code/user.py mkoske
$ # ...
$ ls -lh
$ # ...
$ -rw-r--r-- 1 miika miika 392K marra 24 00:44 mkoske.json

Data

Similar to user.py-script. An example below:

$ cd recommender
$ python code/data.py 50
$ # ...
$ ls -lh
$ # ...
$ -rw-r--r-- 1 miika miika 392K marra 24 00:44 data.json

Above example retrieves metadata on 50 repositories and stores it in JSON-format.

Vectorize

This script is used to convert each repository to a vector and then store the result in CSV-format.

$ cd recommender
$ python code/vectorize.py output/<user>.json
$ python code/vectorize.py output/data.json

Of course, replace with your target Github username.

Preprocess

This script is used to apply a series of preprocessing steps to the README document texts of all the repositories. The output is two csv-files, <user>_tok.csv and data_tok.csv .

$ cd recommender
$ python code/preprocess.py output/<user>.csv output/data.csv

NOTE: The above step must be performed for <user>.csv and data.csv together. The data_tok.csv and <user>_tok.csv files are both specific to that user.

Workflow

Run user.py to get user profile
Run data.py to get dataset
Run vectorize.py to make CSV-files out of both JSON-files
Run preprocess.pyto tokenize all README files (must be done again when changing username)
Run ./run.sh in your local repository root directory to start the web app
Navigate to http://localhost:5000 and enter <user> to get the recommendations.

Dataset statistics

For the statistics of the dataset, please use the notebook found in notebooks-subdirectory.

HouPeiyu/recommender