huangtinglin

图推荐是讲原本的推荐问题，即判断一个用户与某个商品的相关性，看成是一个边预测问题。实际运用即现在某个user-item-KG上训练，然后预测user-item边存在的概率，与我们文中的实验与实验数据类似。

代表了user和正样本或者负样本的相似性 Sent from my iPhone On May 21, 2022, at 14:10, cabbage ***@***.***> wrote:  还有一点，在代码中，公式14体现为 [image] 能具体说说此时pos_score和neg_score分别代表什么吗，有什么具体含义。非常感谢您能在百忙之中抽出时间来回答！ — Reply to this email directly, view it on GitHub, or unsubscribe....

Hi 感谢您的关注！请问下您的显卡显存大小有多少呢，以及跑的是哪个数据集。此外算法是transductive setting，需要每次生成全部结点的embedding，您可以拓展成inductive setting的，每次只生成batch内结点的embedding。

抱歉迟复，对于transductive learning的算法，结点的个数是制约算法在大规模数据集上训练的主要因素，拓展的思路是改写成inductive learning，您可以参考pyg的训练逻辑： https://github.com/pyg-team/pytorch_geometric/blob/master/examples/graph_sage_unsup.py

您好。这个bug我也不太确定呢，我并没遇到，您能分享下您的操作系统环境、Python版本之类的信息吗？

Could you please provide the information about your running environment? Actually, I have just run the code and it can reproduce the performance.

Sorry for the late reply. The time cost of training on a graph comes from the aggregation process of GNN. A way to accelerate the sampling process of MixGCF is...

It's a trick widely used in RL. You can search for the keyword: off-policy, importance sampling. Hope it can help you!

Thanks, @brucejwittmann. Actually, I found that the error is due to the scaling factor which is related to the number of rows. I have created a new issue regarding this...