Hello, Recently I am trying to draw Figure 2 and Figure 3，how did you draw the diagrams about the relationship in Figure 2 and Figure 3 in the paper? Do you have a reference code for this? Thank you very much!

Hello, I am an undergraduate who has just come into contact with the knowledge map, and I have just read your paper. I would like to ask whether the code is running in Linux environment? Can I run and operate under windows? Second, what are the running steps of the code?

We look forward to your answers~

Hi,

First of all, thanks for sharing the code behind your paper! I have a few questions; mainly about how the details in the paper maps to the code.

1. My main concern is how a semantic hierarchies is translated into the embedding. How do we go from text values from the knowledge graphs to an embedding with modulus values that obey the hierarchy; for instance "dog" and "mammal" gets embedded with modulus values 2 and 1? Is it a consequence of the score function?

2. In the code, we have the parameters modulus_weight and phase_weight that are used to multiply with the modulus and phase score. In the paper, however, we only multiply with a weight in the phase part (the lambda parameter) as seen below.
   
   So the question is, are we missing a weight in the modulus part? If not, where does the above lambda parameter map to in the code?

3. I can not find a description of the embedding_range parameter in the paper. Is there an explanation behind its definition? Furthermore, is the purpose of the epsilon value just to get a minimum value for the embedding range? And why 2.0?

4. The initialisation of the HAKE and ModE classes seems to differ in how the relation embedding is initialised. The HAKE initialises the modulus and bias to ones and zeroes respectively. What is the purpose of this?

I hope the questions make sense. Thanks in advance!

As we all know, Poincare embeddings main contribution is representation hierarchy struct data in hyperbolic space.

目前用于实体对齐比较多的是DBP15K，但现在DBP15K里貌似没有类似的层级数据

In have a question about the code written in "codes/models.py".
In "codes/models.py", there is an "if statement" as below.
I want to know if there is a reason for dividing the condition with the 'if statement'.

        if batch_type == BatchType.HEAD_BATCH:
            phase_score = phase_head + (phase_relation - phase_tail)
        else:
            phase_score = (phase_head + phase_relation) - phase_tail

Of course, since it is a float operation, the value is different.
But this is due to the limitation of the operation device.
So I wonder if there is a logical difference.

Thank you for reading.

Hello, I am currently a final year student and I have a graduation project on this knowledge graph topic. I am very impressed with your approach to the problem. I have a question about "Figure 3" in the article, I don't quite understand how to draw this graph. Could you please be more specific about this?
Thank you very much and I look forward to hearing from you

Thank you very much for the code! I read your code carefully, I have one question,I would be grateful if you could help me with the answer

in data.py:131

    def collate_fn(data):
        positive_sample = torch.stack([_[0] for _ in data], dim=0)
        negative_sample = torch.stack([_[1] for _ in data], dim=0)
        subsample_weight = torch.cat([_[2] for _ in data], dim=0)
        batch_type = data[0][3]# this line
        return positive_sample, negative_sample, subsample_weight, 

this function is processing one batch, all data in this batch have only a kind of batch_type ?

应该在代码中如何计算modulus，从而来判断两个实体是否是在一个层次呢？
以及对于不在一个层次的实体，怎么判断他们所属的层次级别呢，比如说动物 鸟 麻雀，动物在麻雀两层之上，鸟在麻雀一层之上。
可以给出具体的代码示例来解决我这边提到的例子问题么。
目前我已经完成了HAKE模型的训练。

你好，我用HAKE得到了实体的embedding。请问我如何根据实体的embedding，判断两个实体是否属于同一层次呢？就是我想问如何利用模型训练好的实体embedding，来区分实体的层级呢？

I find that the hyperparameters from the HAKE is almost the same as RotatE expect the new parameters

• MODULUS_WEIGHT

• PHASE_WEIGHT

Could you explain how you grid seach them?(such as the range of grid search)

For the phase part, why the model score function is sin((h+p-t)/2). If possible, could you please tell us why you divide the (h+p-t) by 2?

Hello,

I was wondering where one could find the number of parameters of the ModE and HAKE on WN18RR, FB15k-237 and YAGO3-10.

Cheers

I'm very curious about the 2D pic from the paper.
Could you tell me how you plot them?

Hello, I have a question about how HAKE models hierarchy in the input graph. Specifically, does the embedding of an entity using HAKE only consider information at a 1-hop distance or is information considered at a multi-hop distance?

For example, if an entity is at the highest level of hierarchy in the graph, does HAKE's embedding of this entity capture information all the way from it's hierarchy down to the lowest level of hierarchy, or is it just considering information at a 1-hop distance?

Thank you!

您好，负采样算法您是否实现了

Thank you for your work
In HAKE, the model has to consider type relation 1-n, n-n?

Looking forword to your reply, thank you.

Hello, thank you so much for your excellent work! When I use my own data set, it has the following error, can you ask me why? The same environment configuration I used with the FB15k-237 dataset you provided did not have this error. But my data runs on TransE, and I converted it to the same format as the FB15k-237 data set you provided. This problem has been bothering me for a long time. I hope you can help me。The following is my data set information and error information.