qcactus/cpr

TensorFlow implementation of our paper: Cross Pairwise Ranking for Unbiased Item Recommendation (WWW'22)

cpr's Issues

Request to Add a License to the Code Repository

First, thank you for sharing your work with the community!

I kindly ask if you could consider adding an open-source license to your repository for enabling others to reference your work.

Thank you for considering our request!

请问如何得到推荐给每个用户前K个商品的id？

如题，谢谢~

没有找到recq.tools.plot文件

data_preprocess中使了recq.tools.plot.degree_plot方法，但仓库中似乎没有找到plot.py文件

Incorrect row/column switching in cpr.py

In the file cpr.py, the code on line 38 is supposed to switch columns, but is, to my understanding, switching rows.

CPR/recq/recommenders/cpr.py

Line 38 in 782edb8

for u, i in zip(u_list, i_list[1:] + [i_list[0]])

This is causing incorrect model behavior. According to equation (13) of the paper and line 7 of the pseudo-algorithm, the procedure should switch the first item of the tuple to the last positon of each tuple. Each row of i_list seems to contain the concatenated embedding vectors of all idx + 2 items of the tuple. So its shape is (idx + 2)*embed_size. Hence, we should be shifting the first embed_size columns to the end.

To fix this issue, the code should be changed to switch the first embed_size columns to the end instead of the first row.

This issue was discovered during testing of the cpr.py file, and appears to be affecting model training. For validation, I tried to input a shifted item sequence directly and computed pos_scores. They should have been identical to neg_scores of the non-shifted sequence but they were not.

Apologies if I have misunderstood your code and there is no error. Please let me know if you require any additional information.

Is there pytorch version?

Hi, thanks for your excellent work, Is there pytorch version?

没有找到recq.tools.plot文件

data_preprocess中使了recq.tools.plot.degree_plot方法，但仓库中似乎没有找到plot.py文件