variational-autoencoders-for-collaborative-filtering-pytorch

variational-autoencoders-for-collaborative-filtering-pytorch copied to clipboard

Variational autoencoders for collaborative filtering in PyTorch.

This repo implements Variational autoencoders for collaborative filtering in PyTorch presented in [1], and also does conditional VAE [2] to use user profiles in collaborative filtering. Numerical expriments are done for MovieLens dataset for VAE and Last.fm for conditional VAE.

Dataset

Two datasets are used in this repo.

• MovieLens 20M Dataset: This contains rating for movies by users. This collaborative filtering recommends movies user would like to watch.
• Last.fm Dataset - 360K users: This contains play counts of artists by users and user profile, gender, age, country and signup data. This collaborative filtering recommends artists user would like to listen to, using play counts and user profiles. You can use VAE conditioned on the the user profiles.

After downloading datasets, run

create_data.sh

Training

Usage:

python demo.py train <options>

Results

NDCG@N and Recall@N of VAE(PyTorch) for MovieLens 20M

References

1. Dawen Liang, Rahul G. Krishnan, Matthew D. Hoffman, Tony Jebara. Variational Autoencoders for Collaborative Filtering, The Web Conference (WWW), 2018.
   arXiv, github

2. Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning Structured Output Representation Using Deep Conditional Generative Models. Advances in Neural Information Processing Systems, 2015. NIPS

3. Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, Alexander Lerchner. beta-VAE: Learning basic visual concepts with a constrained variational framework. ICLR, 2017. Paper link