Hi there,
I am wondering if anyone has any knowledge on the difference in performance of 2 approaches. I am creating a graph dataset to use with DGL, and I have the option of either creating weighted edges using a feature, or to create the edges using KNN and Euclidean distance.
Would it be more beneficial to performance to have extra features? or to have weighted edges?
Many Thanks
Could you elaborate more about your case? What’s the task and scenario? What kind of graphs do you have? What are the features?
I am working with multiple different brain imaging modalities, namely:
I am currently planning on defining the node positions using the anatomical MRI, and using 20 nearest neighbours with Euclidean distance to define which nodes have an edge. I am currently planning on using the dMRI and fMRI as features.
However, I could instead use the dMRI in order to define which nodes are connected to each other.
I was wondering which would benefit performance the most, extra node features or weighted edges?
Thanks for the response,
Chris
Hi Chris,
Thank you for your elaboration. I don’t know much about brain imaging so I probably cannot help much here. I guess it’s not a bad idea to start with KNN graphs when there is not a prior graph structure.
Cheers,
Mufei
Hi Mufei,
Thank you for your responses I will start off with KNN.
Many thanks,
Chris
Someone recommended this paper to me for working with EGG data, in case you find it helpful: EEG Emotion Recognition Using Dynamical Graph Convolutional Neural Networks