The code for our "CoNet: Collaborative Cross Networks for Cross-Domain Recommendation" paper published at CIKM 2018
"./data". The cross-domain datasets split into Train/Valid. I put one example for each data file since the whole dataset consumes much storage. The Amazon data can be downloaded here and the other Cheetah Mobile cannot be publicly available due to privacy (send email to us).
"./CoNet_mtl_cross_1223hid". This is the CoNet model described in the Section 4.2 in our paper. Here, 'mtl'=multitask learning, 'cross'=cross Connections Unit, '1223hid'=cross units enforced between hidden layer 1 and hidden layer 2, and enforced between hidden layer 2 and hidden layer 3. See the illustration Figure 2 in our paper. Tune these hyperparameters on your own datasets.
"./SCoNet_mtl_lasso_cross_1223". This is the SCoNet model described in the Section 4.3 in our paper. Here, 'lasso'=l1-norm penalty describe in Eq. (9) in our paper.
Our methods are implemented using TensorFlow. For the training time, our models spend about 100 seconds per epoch using one Nvidia TITAN Xp GPU. As a reference, it is 70s for MLP and 90s for CSN models.
Our paper is also implemented in the RecBole-CDR recommendation library.
Please cite the following paper if our code+paper helps your research. arXiv
@inproceedings{hu2018conet,
title={Conet: Collaborative cross networks for cross-domain recommendation},
author={Hu, Guangneng and Zhang, Yu and Yang, Qiang},
booktitle={Proceedings of the 27th ACM international conference on information and knowledge management},
pages={667--676},
year={2018}
}
...Our project page