Chaining Node2Vec and GraphSAGE #1873
Replies: 3 comments 3 replies
-
I want to ask a similar question - perhaps even more generalized: #1877 |
Beta Was this translation helpful? Give feedback.
-
Please let me know if my understanding is correct here. You'd like to:
Some notes for this approach:
An important point here is that both Node2Vec or MetaPath2Vec are NOT inductive. Any embeddings trained on one graph will not transfer to another graph, and furthermore for a graph that is disconnected (ie. is not a single connected component) the embeddings in one component will not have any relation to the embeddings for a disconnected component. I hope this helps, please ask any follow up questions on this and some more information on how you are training GraphSAGE (do you use both types of links, are you training for link prediction or node classification?) and how you envisage the inductive training working would be useful. |
Beta Was this translation helpful? Give feedback.
-
Can Node2vec or other Embedding algorithms like Struc2Vec operate on completely disconnected nodes from the graph? like the node below> |
Beta Was this translation helpful? Give feedback.
-
Eventually I will run this experiment. But while I find the time, I wanted to find out if it had been attempted before.
My end goal is inductive node embeddings via GraphSAGE. I want the inductive property so I can generate new embeddings without retraining the entire model.
My graph is an ontology. Generally speaking, there are two types of links: Taxonomic and semantic. Taxonomic links define the skeleton or structure of the ontology. Every node has an "IS-A" link such that there is a path between all nodes in a tree-like structure. The semantic links are the interesting pieces of information that connect nodes (social relationships, part-whole, etc.)
Currently I train it just in GraphSAGE and each node has a binary feature vector of length 12. My intuition tells me this isn't enough information for quality embeddings. So the experiment is: First, train embeddings using just the taxonomic links fed into Node2Vec. Second, use the embeddings from Node2Vec as the feature vectors for training embeddings using just the semantic links fed into GraphSAGE. Logically this makes sense to me as Node2Vec learns structures like what the taxonomic links describe. And GraphSAGE clusters related nodes using semantic information. Mathematically I don't know if this gains me anything with the added complexity...
Beta Was this translation helpful? Give feedback.
All reactions