chenchongthu / JNSKR

This is our implementation of JNSKR: Jointly Non-Sampling Learning for Knowledge Graph Enhanced Recommendation (SIGIR 2020)

JNSKR

This is our implementation of the paper:

Chong Chen, Min Zhang, Weizhi Ma, Yiqun Liu and Shaoping Ma. 2020. Jointly Non-Sampling Learning for Knowledge Graph Enhanced Recommendation. In SIGIR'20.

Please cite our SIGIR'20 paper if you use our codes. Thanks!

@inproceedings{chen2020jointly,
  title={Jointly Non-Sampling Learning for Knowledge Graph Enhanced Recommendation},
  author={Chen, Chong and Zhang, Min and Ma, Weizhi and Liu, Yiqun and Ma, Shaoping},
  booktitle={Proceedings of SIGIR},
  year={2020},
}

You also need to cite the KDD'19 paper if you use the datasets. Thanks!

@inproceedings{KGAT19,
  author    = {Xiang Wang and
               Xiangnan He and
               Yixin Cao and
               Meng Liu and
               Tat{-}Seng Chua},
  title     = {{KGAT:} Knowledge Graph Attention Network for Recommendation},
  booktitle = {{KDD}},
  pages     = {950--958},
  year      = {2019}
}

Author: Chong Chen ([e-mail])

Baselines and Datasets

We follow the previous work, KGAT, and you can get the detailed information about the baselines and datasets in https://github.com/xiangwang1223/knowledge_graph_attention_network.

Example to run the codes

Train and evaluate our model:

python main_JNSKR.py

Train and evaluate baselines:

python main_Baselines.py

Reproducibility

For Amazon dataset:

parser.add_argument('--dropout', type=float, default=[0.8,0.7],
                        help='dropout keep_prob')
parser.add_argument('--coefficient', type=float, default=[1.0, 0.01],
                        help='weight of multi-task')
parser.add_argument('--c0', type=float, default=300,
                        help='initial weight of non-observed data')
parser.add_argument('--c1', type=float, default=600,
                        help='initial weight of non-observed knowledge data')
                        

For Yelp dataset:

parser.add_argument('--dropout', type=float, default=[0.9,0.7],
                        help='dropout keep_prob')
parser.add_argument('--coefficient', type=float, default=[1.0, 0.01],
                        help='weight of multi-task')
parser.add_argument('--c0', type=float, default=1000,
                        help='initial weight of non-observed data')
parser.add_argument('--c1', type=float, default=7000,
                        help='initial weight of non-observed knowledge data')

Suggestions for parameters

Several important parameters need to be tuned for different datasets, which are:

parser.add_argument('--dropout', type=float, default=[0.8,0.7],
                        help='dropout keep_prob')
parser.add_argument('--coefficient', type=float, default=[1.0, 0.01],
                        help='weight of multi-task')
parser.add_argument('--c0', type=float, default=300,
                        help='initial weight of non-observed data')
parser.add_argument('--c1', type=float, default=600,
                        help='initial weight of non-observed knowledge data')

Specifically, c0 and c1 determine the overall weight of non-observed data. The coefficient parameter determines the importance of different tasks in joint learning.

You can also contact us if you can not tune the parameters properly.