yaochenzhu / md-cvae

Official codes of "Mutually-Regularized Dual Collaborative Variational Auto-encoder for Recommendation Systems" (WWW'22)

MD-CVAE: Mutually-Regularized Dual Collaborative Variational Auto-encoder for Recommendation Systems

The codes are associated with the following paper:

Mutually-Regularized Dual Collaborative Variational Auto-encoder for Recommendation Systems,
Yaochen Zhu and Zhenzhong Chen,
The Web Conference (WWW) 2022. [paper] [slides]

Environment

The codes are written in Python 3.6.5.

• numpy == 1.16.3
• pandas == 0.21.0
• tensorflow-gpu == 1.15.0
• tensorflow-probability == 0.8.0

Datasets

The raw features of the established movielen-sub dataset can be obtained here: [Google Drive], [Baidu] (PIN:r8f0)

The processed datasets can be found here: [Google Drive], [Baidu] (PIN:l8wr)

For usage, please unzip the processed datasets and copy them into the data folder.

Examples to run the codes

1. To reproduce the in-matrix prediction results:

• Pretrain the dual item content embedding VAE:

  python pretrain_vae.py --dataset Name --split [0-9]

• Iteratively train the collaborative and content VAEs:

  python train_vae.py --dataset Name --split [0-9]

• Evaluate the model and summarize the results into a pivot table

  python predict.py --dataset Name --split [0-9]

  python summarize.py

2. To reproduce the out-of-matrix prediction results:

• First, please change to the cold_start folder.

• Download the processed data [Google Drive], [Baidu] (PIN:f4f7)

• The way to run the codes and summarize the results is similar to the in-matrix case.

For advanced argument usage, run the code with --help argument.

Reference

if you find the codes helpful, please kindly cite our paper. Thanks!

@inproceedings{MDCVAE-WWW2022,
  title={Mutually-Regularized Dual Collaborative Variational Auto-encoder for Recommendation Systems},
  author={Zhu, Yaochen and Chen, Zhenzhong},
  booktitle={Proceedings of the ACM Web Conference 2022},
  year={2022},
}    

md-cvae's Issues

Differences between CVAE and MD-CVAE

Hi, thanks for sharing the code for this paper.

May I have your help to identify the major differences between CVAE and MD-CVAE, as CVAE also ingest item-content for CF.
However, in your experiment, CAVE is compared without item-content?

The general idea and formula seems same between the two papers at my first glance.

https://github.com/eelxpeng/CollaborativeVAE
@eelxpeng

Thanks.

Meaning of datasets

Hi, May I have your kind help to identify the meanings of each file of a dataset used in your code (e.g. cora):
As normally, we only need the adj matrix (user-item graph) and the item feature files for VAE.

• item_features.npz
• meta.csv
• train.csv
• *_obs.csv
• *_unk.csv
• *_X.csv
• *_Y.csv

Thanks a lot.

Word Vocabulary of Chosen Datasets

Hi @yaochenzhu,

Thanks for very interesting work.

Could you kindly provide the word vocabulary of datasets, including citeulike-a, movielen-sub.
I would like to map features stored in item_features.npz to natural language words so as to understand item content.

One more question is what is the meaning of features.npz file? I found that in processed citeulike dataset, shape of features.npz is (5551, 8000). Is this the feature matrix for users? If so, how to derive this matrix.

Thank you again and I'm looking forward to hearing from you soon.