Hyperbolic (ordinary and variational) autoencoders for recommender systems

Accompanying code for the paper Performance of Hyperbolic Geometry Models on Top-N Recommendation Tasks, accepted at ACM RecSys 2020.

Results

Data

To reproduce our code, please put the corresponding data files in the following folder structure:

data

• troublinganalysis
  • mvae
    • netflix
    • ml20m
  • neumf
    • ml1m
    • pinterest
• recvae
  • ml20m

Also, please install geoopt package geoopt for Riemannian optimization and hyptorch for computations in hyperbolic spaces.

Wandb

In our experiments, we have used wandb framework for result tracking. Our test scripts are based on wandb configs.

Acknowledgments

In our code we have used the following repositories:

• mvae
• geoopt
• hyptorch