This is our implementation of the paper:

Chong Chen, Min Zhang, Weizhi Ma, Yongfeng Zhang, Yiqun Liu and Shaoping Ma. 2020. Efficient Heterogeneous Collaborative Filtering without Negative Sampling for Recommendation. In AAAI'20.

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

@inproceedings{chen2020efficient,
  title={Efficient Heterogeneous Collaborative Filtering without Negative Sampling for Recommendation},
  author={Chen, Chong and Zhang, Min and Ma, Weizhi and Zhang, Yongfeng and Liu, Yiqun and Ma, Shaoping},
  booktitle={Thirty-Fourth AAAI Conference on Artificial Intelligence},
  year={2020},
}

Author: Chong Chen ([email protected])

• python
• Tensorflow
• numpy
• pandas

The parameters for Beibei datasets is:

self.weight = [0.1, 0.1, 0.1]

self.coefficient = [0.05, 0.8, 0.15]

The parameters for Taobao datasets is:

self.weight = [0.01, 0.01, 0.01]

self.coefficient = [1.0/6, 4.0/6, 1.0/6]

We would like to correct one typo error of the paper. In table 2, the HR@10 and NDCG@10 results of our EHCF on Beibei dataset should be 0.1551 and 0.0831, which are much better than baseline methods.

Train and evaluate the model:

python EHCF.py

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

self.weight = [0.1, 0.1, 0.1]
self.coefficient = [0.05, 0.8, 0.15]
deep.dropout_keep_prob: 0.5

Specifically, we suggest to tune "self.weight" 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].

The performance of our EHCF is much better than existing multi-behavior models like CMF, NMTR, and MBGCN (SIGIR2020). You can also contact us if you can not tune the parameters properly.

For some reasons, we do not recognize MBGCN (Multi-behavior Recommendation with Graph Convolution Networks, SIGIR 2020) as a state-of-the-art method for multi-behavior recommendation. We also call on researchers to not only compared with MBGCN in future research, so as to avoid getting an inaccurate conclusion.

First Update Date: May 27, 2020