tsinghua-fib-lab/graph-denoising-socialrec

The official implementation of "Robust Preference-Guided Denoising for Graph based Social Recommendation" (WWW'23)

GDMSR

This is the official implementation of the WWW 2023 paper Robust Preference-Guided Denoising for Graph based Social Recommendation.

In this paper, we instead propose to improve graph based social recommendation by only retaining the informative social relations to ensure an efcient and efective infuence difusion, i.e., graph denoising. Our designed denoising method is preference-guided to model social relation confdence and benefts user preference learning in return by providing a denoised but more informative social graph for recommendation models. Moreover, to avoid interference of noisy social relations, it designs a self-correcting curriculum learning module and an adaptive denoising strategy, both favoring highly-confdent samples. Experimental results on three public datasets demonstrate its consistent capability of improving three state-of-the-art social recommendation models by robustly removing 10-40% of original relations.

Architecture

Environment

• Tested OS: Linux
• Python >= 3.7
• PyTorch == 1.7.1

Data

We provide all three datasets we used in our experiments, all in the data folder

For each dataset, in addition to the u-i and u-u relation data, we also provide the negative samples used to calculate the metrics, as well as some other pre-processed data to reduce computation, such as user_item_dict.pkl, user_visited_dict.pkl and visited_and_mask_matrix_30.pkl.

For the new dataset, we have provided the data processing code that you can use by input the u-i and u-u relation data in data_process.py

Model Training

To run the experiments, the scripts we provide can be used directly:

./run.sh

graph-denoising-socialrec's Issues

您好，我对贵团队的工作非常感兴趣，能否更新readme中的信息，感谢！

How to get social recommendation outputs?

I am using the code from your paper for a social recommendation task and I cannot figure out how to obtain the recommendation outputs. It seems like your NDCG score is calculated based on user-item interaction but not user-user interaction. If I have to obtain which new users can be recommended to a given user, how would I get that output from the model?

实验

老师你好，请教一下您提出的这个方法如何运用到已有模型中？如lightgcn模型，文章中将方法加入基线模型中得到结果，您提供的代码是不是不包括基线模型，

跪求超参！

您好，我对贵团队的这篇文章很感兴趣，可以提供一下生成pkl数据处理代码吗，谢谢！

baseline

作者你好！我看文中采用lightgcn等模型作为基线，附加去噪方法，但是我通读代码并没有发现基线这块代码，请问这块代码是在哪里？

可以提供一下源代码吗，谢谢

您好，我对贵团队的这篇文章很感兴趣，可以提供一下源代码吗，谢谢！