#Graph data structures. Looking for the right ML technique

8 messages · Page 1 of 1 (latest)

delicate edge
#

Hello friends, I have a list of graph data structures. Each graph has a number of nodes and relationships. I also have a description for each graph. Given a new graph, I need to generate the description. How can I find the closest graph from my list? Is there an ML technique I can use? I have tried just vectorizing the graph data and using classification techniques to predict. It's not very accurate.

viscid lagoon
#

@untold shale @opaque dove thoughts?

limpid moat
#

@delicate edge How many graph instances do you have in your list? If they are sufficient in quantity, I might recommend building a graph autoencoder. The autoencoder should learn an intermediate (and vector-like) representation. This representation can then be compared to your test data point using the nearest neighbor algorithm. Just my 2 cents.

delicate edge
#

I have 3000 or so graph instances. I will read up on autoencoders. Thanks for your help! I wonder if it makes a difference to say that the same node types and relationship types can occur across multiple graphs. Like 'X' type node can be connected to 'Y' type node across graphs. There will also be cases where 'X' is not connected to 'Y' for certain graphs. Thank you!

untold shale
delicate edge
#

Thanks Tyler!

signal sparrow
#

graph convolutions?

#

if its smaller graphs u can also use graph attention transformers