Graph Embedding II: metapath2vec

Topics

Now we know about word2vec and node2vec. We also understand why we can’t directly apply node2vec on HinDroid. This week we are going to study metapath2vec which will address this shortcoming.

Reading

metapath2vec

Pls go through the paper thoroughly. We are going to majorly focus on this paper next Wednesday.

The gist of the paper is that we choose the type of metapath to traverse to generate a neighbourhood. By doing so we don’t let a specific edge type dominate the rest while generating the neighbourhood.

Task

Consider the following graph:

At,

t=0, node = S
t=1, node = a

Landing Nodes:
A. Green (I)
B. Red (B)
C. Yellow (P)

Q1. If we are following node2vec then at t=2, what is the probability of landing on the above nodes for the following three cases:

Q2. If we are following metapath2vec, then at t=2, what is the probability of landing on the above nodes for the following three cases:

Weekly Questions

• From the metapath2vec paper, provide any 3 sentences from the first three sections (Intruduction, Problem Statement and the Framework) that you think cogently explains the problem statement and the approach.
• Last week we learned about node2vec. Why can’t we directly use node2vec to solve our problem (we discussed this on Wednesday)?
• Do you think metapath2vec can address this problem? Explain briefly.