Reputation: 1
I am very interested in Node Classification using GCN.
But I don't know how it can be applied when there are many types of labels in GNN.
For example, in the case of drugs, detoxification can be carried out for various diseases, so what I want is that if we proceed with the node classification of drugs, various applicable diseases will come out as a label.
To this end, we are going to proceed with multi-class classification, but since the basic label information has multiple labels, it is difficult to apply the code as a label by applying one hot encoding of [drug x label_class_num].
I would appreciate it if anyone could provide an example or give me some information for reference.
I'm also wondering if it's possible to compare with BCE whether or not that node has that disease when you pick one out of n diseases. In this case, do you need n models to separate each disease into BCE?
My experimental code is based on the code on the site below. https://www.kaggle.com/code/naifislam/node-classification-with-gcn
Upvotes: 0
Views: 362
Reputation: 11
When you say you have many types of labels in GNN
, do you mean that each node in the input graph to the GNN has multiple types of labels? For instance, the same molecule graph can be used to classify the smell as well as the colour of the molecule. However, these would be 2 different classification problems. Or do you mean that there are many labels for a class (eg. multiple colours that a molecule could be)? This is the typical multi-class classification problem.
In the first case, you can break down the problem into multiple multi-class classification problems where each solution performs classification for one type (e.g., one GNN model for classifying colours, another for classifying smells). The second case would then be a subset of the first case.
If you are using DGL for defining your GNNs, a sample multi-class node classification example can be found here (the node classification model on the Cora graph dataset here is a 7-class classification problem).
Upvotes: 1