Details in simple graph training and embedding about graphsage HOT 2 OPEN

Hi, in this case we can use one-hot indicator vectors as the features for each node. I.e., we can assign node $i$ as a vector of size |V| where all values in this vector or zero except for the entry corresponding to node $i$.

However, this leads to very sparse features and so for implementation we have a special flag called "identity_dim" for this case. If you have no node features than you can set "identity_dim" to a positive value. This flag will associate each node with a dense embedding vector, and we backpropagate gradients through these embeddings. (Mathematically, this is equivalent to representing each node by a one-hot indicator vector and then multiplying these vectors by a learned feature matrix of size |V|xidentity_dim).

Hope that makes some sense...

from graphsage.

@williamleif Wow, this comment perfectly solves my question! When I first got this TF code, I wondered why you used trainable variables to initialize the embeddings of nodes that have no features. So you are using this method to model the multiplication of one-hot encoding with a trainable weight matrix?

from graphsage.