Course talk:CPSC532:StaRAI2020:Negative Sampling on Knowledge Graphs

Hey Melika, Is there any reference where they have compared these methods quantitatively? If it exists could you please share the information on the page or could you provide a link in bibliography?

Thank you, Maulik. That could be a very nice idea. I added a reference that explored these methods very well. My page is also complete now. Would be happy to have your other feedbacks as well.

Why is the word2vec relevant? In word2vec we know the other words are not in the window. However for knowledge graphs we don't know whether they are true or not; we just know that random variants will be mostly false on average.

I couldn't understand the descriptions of the methods. Try to make them as straightforward as possible. The motivation is good, but someone should be able to reproduce the sampling from your description. (Say, something like.: for each <h,r,t> in knowledge graph, choose k negative examples for <h',r,t> and k negative examples for <h,r,t'> for h' and t'....)

The reason I provided the word2vec example was to make a better understanding of the "Negative Sampling" concept. When I was exploring the negative sampling on knowledge graphs' papers, it seems it was assumed as general knowledge. However, I did not get what does it exactly means and what is its' application. So, I thought it would be helpful for readers to first get to know what could be a negative sample in another context. Not sure whether it helps or became more confusing.

For explaining the sampling methods you are right. I will try to explain them better. Thanks for your feedback.

It would be good to explain how negative sampling in KGs is different from word2vec, then. Making it clear why we need negative sampling would be a good contribution of this paper.

I made some edits and added some details to make it more clear. Hope that is better now!