GHCF
This is our implementation of the paper:
Chong Chen, Weizhi Ma, Min Zhang, Zhaowei Wang, Xiuqiang He, Chenyang Wang, Yiqun Liu and Shaoping Ma. 2021. Graph Heterogeneous Multi-Relational Recommendation. In AAAI'21.
Please cite our AAAI'21 paper if you use our codes. Thanks!
@inproceedings{chen2021graph,
title={Graph Heterogeneous Multi-Relational Recommendation},
author={Chen, Chong and Ma, Weizhi and Zhang, Min and Wang, Zhaowei and He, Xiuqiang and Wang, Chenyang and Liu, Yiqun and Ma, Shaoping},
booktitle={Proceedings of AAAI},
year={2021},
}
Example to run the codes
Train and evaluate our model:
python GHCF.py
Reproducibility
parser.add_argument('--wid', nargs='?', default='[0.1,0.1,0.1]',
help='negative weight, [0.1,0.1,0.1] for beibei, [0.01,0.01,0.01] for taobao')
parser.add_argument('--decay', type=float, default=10,
help='Regularization, 10 for beibei, 0.01 for taobao')
parser.add_argument('--coefficient', nargs='?', default='[0.0/6, 5.0/6, 1.0/6]',
help='Regularization, [0.0/6, 5.0/6, 1.0/6] for beibei, [1.0/6, 4.0/6, 1.0/6] for taobao')
parser.add_argument('--mess_dropout', nargs='?', default='[0.2]',
help='Keep probability w.r.t. message dropout, 0.2 for beibei and taobao')
Suggestions for parameters
Several important parameters need to be tuned for different datasets, which are:
parser.add_argument('--wid', nargs='?', default='[0.1,0.1,0.1]',
help='negative weight, [0.1,0.1,0.1] for beibei, [0.01,0.01,0.01] for taobao')
parser.add_argument('--decay', type=float, default=10,
help='Regularization, 10 for beibei, 0.01 for taobao')
parser.add_argument('--coefficient', nargs='?', default='[0.0/6, 5.0/6, 1.0/6]',
help='Regularization, [0.0/6, 5.0/6, 1.0/6] for beibei, [1.0/6, 4.0/6, 1.0/6] for taobao')
parser.add_argument('--mess_dropout', nargs='?', default='[0.2]',
help='Keep probability w.r.t. message dropout, 0.2 for beibei and taobao')
Specifically, we suggest to tune "wid" among [0.001,0.005,0.01,0.02,0.05,0.1,0.2,0.5]. It's also acceptable to simply make the three weights the same, e.g., self.weight = [0.1, 0.1, 0.1] or self.weight = [0.01, 0.01, 0.01]. Generally, this parameter is related to the sparsity of dataset. If the dataset is more sparse, then a small value of negative_weight may lead to a better performance.
The coefficient parameter determines the importance of different tasks in multi-task learning. In our datasets, there are three loss coefficients λ 1 , λ 2 , and λ 3 . As λ 1 + λ 2 + λ 3 = 1, when λ 1 and λ 2 are given, the value of λ 3 is determined. We suggest to tune the three coefficients in [0, 1/6, 2/6, 3/6, 4/6, 5/6, 1].
32 Dec 04, 2022
37 Dec 18, 2022
70 Dec 05, 2022
258 Dec 22, 2022
16 Nov 07, 2022
2 Feb 04, 2022
14 Dec 11, 2022
3.6k Dec 30, 2022
794 Dec 23, 2022
5 Oct 02, 2022
85 Oct 04, 2021
1 Nov 15, 2021
7 Mar 12, 2022
3 Aug 31, 2021
0 Mar 31, 2022
3k Dec 31, 2022
18 Jan 02, 2023
1.4k Dec 27, 2022
3 Oct 30, 2022
5 Jul 19, 2022