opringle / gluonrank

Ranking made easy

GluonRank: Your Choice of Deep Learning for Ranking

GluonRank is a toolkit that enables easy implementation of collaborative filtering models using neural networks, to help your prototyping of state of the art ranking systems.

Installation

Pip

Make sure you are using Python 3.6. You can install MXNet and GluonRank using pip:

pip install --index-url https://test.pypi.org/simple/ gluonrank

Uploading to pypi for testing

Build distribution

python setup.py sdist bdist_wheelbash

Upload to pypi test index

twine upload --repository-url https://test.pypi.org/legacy/ dist/*bash

Docs

Coming soon... (it might be a while actually...)

ToDo

• Categorical features
  • Get running with multiple categorical features, maintain performance when reducing to a single one
  • Gracefully handle missing continuous embedding or categorical variables & user/item biases
  • Do not require user to index their embedding values for a single matrix
• Continuous features
  • Get running with 1 continous feature, maintain performance when excluded
  • Get running with several continuous features
• Increase the efficiency of the evaluation function
• Speed up negative sampling... Negative sampling without collisions results in 5X training time.

(answer)[https://stackoverflow.com/questions/53576915/sample-n-zeros-from-a-sparse-coo-matrix/53577344#53577344]

• Match spotlight performance with implicit interaction model on movielense data
• Build ranking function as network method

• [ ] Create python package

• Create hosted docs

Features

• Allow for sampling more than one negative per interaction
• Allow for feedback that can be in the form of 0, 1 or -1. (eg swiping data)

Ideas