hdvvip/CS224W_Winter2021

On hw1 question 6

Closed this issue · 1 comments

aspatic commented

hello there i am just wondering why do we need a nega-sampling here, isn't nega-sampling only for random walk?

hdvvip commented

We are training something like node2vec. Given a node, we want a vector to represent that node.
So, if nodes are connected to each other then their corresponding vectors must have high cosine similarity.
Nodes that are not connected to each other must have low cosine similarity.

We know which nodes are connected to each other by looking at the graph. (positive examples)
Nodes that are not connected to each other are retrieved by nega-sampling. (negative examples)

This idea of Node2Vec is very similar to Word2Vec.
I recommend read the Word2Vec paper for understanding why we need neg-sampling (negative examples)