ethanhezhao/nbvae

Negative-Binomial VAE for discrete data

NBVAE

This is the demo code for the paper "Variational Autoencoders for Sparse and Overdispersed Discrete Data" in AISTATS 2020 Link.

Datasets

The datasets of 20NG and ML-10M are provided.

The other text data can be downloaded from the code repo of DPFA.

The other collaborative-filtering data can be downloaded from the links in the paper and preprocessed with the code of MultiVAE.

Installation & Set-Up

The code is implemented with Python 3.5.2 and Tensorflow 1.10.0, and also requires Numpy, Scipy, Scikit-learn, Pandas, and Bottleneck installations.

Run the demos

The demos of NBVAE and NBVAE_dm on text data are in demo_NBVAE.sh.

The demo of NBVAE_b on binary collaborative-filtering data is in demo_NBVAE_b.sh.

nbvae's Issues

A question about Loss

First I have to say it is a nice work!
I have noticed that you focused on the generative process in Section 2.1 where describing the NBVAE. And in the code, the KL loss means that the encoder was still sampling from Gaussian distribution.
And this is my question, why not sample from X~NB(r,p)? And if features were learned from Gaussian in the encoder, how can the decoder reconstruct the information following NB distribution?
Thank you!