Node Classification with NLP

[glacial bolt]

Hello all,
I have a text classification where labels have inherent tree relationships. I can do simple text classification where given "text", predict "label", but since labels are actually end nodes of a tree with various level of nodes, such as:

Parent 1 > Parent 2 > Terminal Node (label 1) 
or 
Parent 3 > Parent 4 > Parent 5 >Terminal Node (label 2) 

I'd like to see if I can turn this into given "text", predict node to utilize the inherent tree structure behind the labels. Given that there are hundreds of labels, the search space in general is not that small.

Is this a plausible approach worth exploring in general?

Update:
I also just realized Node Classification is the wrong terminology. I'm classifying text into nodes

@viscid crypt @serene grotto Pinging as you guys are more experienced with this

[viscid crypt]

perhaps a hybrid approach would work best. get feature vectors for the text, and train a decision tree model based on the inherent tree structure?

[glacial bolt]

Hmm training a decision tree is doable, but what should be the label be? If the label is just encoded label 0, 1, 2..., then it's not really using the inherent tree structure

[viscid crypt]

the label should, presumably, be the vector representation of the text ascribed to the label

[serene grotto]

This is a bit outside my area of experience. But I think I've seen a similar problem on the Keras code examples page. Link here:
https://keras.io/examples/graph/gat_node_classification/
https://keras.io/examples/graph/gnn_citations/

The GAT example performs node prediction based on node subject & linked citations

The GNN example does a similar thing

I hope this brings inspiration to solving your problem @glacial bolt. Sorry I couldn't provide more input/help

[glacial bolt]

So if I'm understanding correctly:

1. Turn the labels into node vector features
2. Encode text to with something like spacy or bert
3. Train a decision tree to decide on encoded text to predict node vector?

[glacial bolt]

It does remind me that I need to understand this deeper! Thank you for the resources. I have 3 day hackathon internally next week, so I wanna try this approach rather than always go for text classification only

[viscid crypt]

Yes, that's my back of the envelop thoughts on the approach

[glacial bolt]

Will look into this then! What are some libraries that you use for dealing with graph? I'm watching one that's using DGL for example

[viscid crypt]

Dionysus 2 is the main one, but I think it might be more specific than what you need.

[glacial bolt]

I love the name haha thanks a ton guys

[serene grotto]

Question: what relation does each node have to the top level parent?

Also, just to be clear, the classification label is for the top level parent right?

[glacial bolt]

It's actually the opposite. So it's predicting which category should a product be in. This image is taken from Amazon, since they have similar structure. As you can see, Leaf Browse Node is basically terminal node of the tree, and it's not certain which level they could be. The idea is the same, given a product description, predict which Leaf Browse Node does the product belong to. In this case, if a product is Boots, Sneakers, or something other Leaf Browse Node under Heels

[serene grotto]

So each node serves as context for the leaf, & the label is for the leaf node?

Does each leaf have 1 label or do the have multiple labels?

[glacial bolt]

Each leaf would have one label only. I’m trying to see if I can utilize the fact that the leaf itself is from Shoes, that relationship could provide more context to the leaf itself