hwwang55 / pathcon Goto Github PK

When I want to view the complete PathCon model code, I found that there are some missing codes? The following are something puzzle me.

• Point one: Regarding the experiment in Figure 6 in the paper, it cannot be found in the provided code about the Embedding+Mean and Embedding+Attention.
• Point two: The formula 15 in the paper uses softmax, but the code provided uses sigmoid?
• Point three: The formulas 13 and 14 in the paper have no traces of use in the embedding mode in the provided code, but the paper mentions that PathCon is more effective after introducing formulas 13 and 14. Can you explain it here?
  The above are some of my questions, I look forward to your reply!!!

您好！
请问您们是按照什么样的策略将知识图谱的数据集划分成训练集、测试集的？能否告知呢？

请问我要如何保存训练后的头尾实体的向量表示，谢谢！

您好，十分欣赏您及其团队所提出来的PathCon模型。由于我自己实验的需要，关于TransE, DistMult等基于Embedding的baseline方法，有些实现细节需要向您请教：

1. APPENDIX D中给出了baseline的代码的链接，请问超参的话是直接采用它们之前所采用的最佳超参吗？
2. 之前的方法在训练的时候是对头实体和尾实体进行破坏进行负采样，而就像4.1 Evaluation Protocol中提到的一样，本文是对关系进行预测，所以是对关系进行破坏进行负采样。那这样的话，负采样率这个超参还是沿用之前代码链接的负采样率吗？如果不是的话，想了解一下本文baseline实现时的负采样率。
3. 由公式16可知，本文是通过交叉熵，将预测关系转换为一个分类的问题，所以按照我的理解PathCon模型是没有进行负采样的（当然也可能理解错了）。我好奇baseline中的loss是直接沿用之前代码各自的实现，还是也采用PathCon这种方式的实现，即过完打分函数之后直接softmax再交叉熵（我的判断是直接采用的之前的实现，不然也不会提到对关系进行负采样）。因为基于Ruffinelli et al.，我觉得不同loss的选择对模型的结果影响很大。不知您是否有做相关的实验。

（当然了，如果您愿意开源baseline的实验细节那是最好不过的了

4. 很好奇PathCon为什么没有像之前知识图谱链接预测的工作一样，给出头（尾）实体和关系，来预测尾（头）实体，而是给出两个实体预测它们之间的关系。我目前的推测是因为模型对关系上下文进行传递并结合两节点之间路径表示这种范式导致，所以直接给出头实体和尾实体预测关系更为直接和方便。但毕竟之前的论文都是采用链接预测这种方式，不知您是否有将PathCon在链接预测这种实验条件下进行实验？或者对于将PathCon用于链接预测的loss的构建有什么好的建议吗？

您好，PathCon这个工作十分的有意义！我很感兴趣！
在复现代码是，我发现代码实现中在model.py的line 106，待预测的关系被加入了edge_list，然后在line 170，当hop=0，i=0时，待预测的关系将会被当作self_vector被聚合，也就是在训练的时候就使用到的待预测的关系信息，但是理论上不应该有这样的信息进入模型训练？请问是这样吗？

Can you explain the purpose or effect of mask in the function "_get_neighbors_and_masks()" of model.py?

specifically for the following two lines of codes?

mask = neighbor_edges - train_edges # [batch_size, -1]
mask = (mask != 0).float()
it is an excellent job of your work ,hope your reply,thanks.

May i know whether the relation embedding will be updated during training?

Hi,

I'm trying to reproduce the results presented in the paper using the WN18RR dataset. However, I'm running into some kind of trouble since I'm not able to replicate the results. Using the default parameters, i.e., running only python main.py the obtained results are, for example:

final results
acc: 1.0000   mrr: 0.4864   mr: 3.9359   h1: 0.3086   h3: 0.5443   h5: 0.7096

What I'm doing wrong?

Also, the other models' differ greatly from everything reported so far, as an example, the RotatE results for the WN18RR dataset reported in the original paper are mmr: 0.440 h1: 36.1 h3: 48.3 h10: 58.1, while in the PathCon paper are mmr: 0.799 h1: 0.735 h3: 0.823. How do they were evaluated?

Thanks in advance.

您好，请问可以告知一下 bert.npy文件的生成方式吗？十分感谢。
Hello,sir!
Would you tell me how I can generate the bert.npy files by myself? Thank you very much!!

嗨，您好，想问一下您的FB15k路径文件生成大概需要多久？谢谢！

您好！我在学习贵团队pathCon模型时，发现所有数据集的path_type均为embedding，请问是因为rnn的处理效果略差的原因吗？
感谢～

Section 4.2 describes an experiment for inductive KG completion which includes sampling a subset of nodes in the test set and removing them from the training set.
How was this implemented in the code? I am looking at data_loader.py, but I cannot seem to find it.
Thank you.

Hi there! Can you help me with the question about using the model in an inductive setting, please?

Let's assume, we want to get the model scores for a pair of vertices not presented in training set. According to the code, we need to use calculated relational paths which rely only on the training set. It means that for new vertices there will be no paths connecting them with vertices in the training set. What should we do in such a case?