Node2Vec Question: how does the implementation generate input-target context pairs from the sampled random walks for training with the Skip-gram architecture?

[ghost]

Hi,

Thank you first of all for your great work!

In your Pytorch Geometric Node2Vec tutorial you explain how the "context size" is used to generate smaller random walks from a larger random walk (see below for the example you give in the tutorial).

Now, how does your implementation generate input-target context pairs as training data for the Skip-gram architecture from this random walk? Which pairs in the first mini random walk {u: s4, s5, s6} will serve as training pairs from the algorithm? Will those be {u,s4}, {u,s5} and {u,s6} or {u,s4}, {s4,s5},{s5,s6}?

Or, is the training data generated in the same way as for sentences that serve as input to Skip-gram in word2vec, where a sliding window is used over the random walk? (See image below):

Thank you in advance!
Ruth

[rusty1s]

Context nodes are generated via sliding window based on the original random walk (like in your first image). Training data for skipgram will then be {u, s4}, {u, s5}, ..., see here.

[ghost]

Thank you!