emiyalzn / graphde Goto Github PK

Thanks for your paper and code. I have some puzzles about OOD detection and down-weight outliers, and I hope you can help me.

1. The OOD detection is the classifier of graph neural networks, why is it? During the training process, the GNN classifier is trained as the classifies for each graph, so why can it ultimately be used as a detector for OOD?
2. I didn't find the code for down-weight outliers in the dataset drugood. I want to know how to deal with this part.

Thank you very much. I would like to receive your reply as soon as possible.

感谢作者如此精彩的工作！在参考代码过程有以下疑问：

在Instantiations of the Recognition Model中，提到利用graphde-a来对环境的$e$来进行优化，我找到以下部分的具体代码：

疑问如下：

e_inferred_logprob 为什么要concat起来呢？最后计算的时候，用了一个手动设置prior的阈值来进行先验处理，为什么不是利用e_inferred进行处理？

这些在论文没有得到解释，烦请作者指点！期待您的回复！