Approaching Graph Deep Learning - mapping vectors on lattice

Models & Apps
#1

Hi,

glad I found DGL, I think it can help much in my researchers and goals to extend knowledge on graph deep learning.

Quickly about me:
I came from areas of data-visualisation and applied network science, don’t have strong knowledge on modelling and asking here for suggestions and dicussion; I self-funded pilots for contextual research, and willing to explore GDL applied on my own model.

As an introduction to GDL to learn more by “getting the hands on” it, I would like to discuss applicability of GDL for a toy-model, how neighbours influence a node:

please see a verbal description of the model here:

Help to hint which model I should look at and tips for implementation will be great, especially if you could explain to a beginner why do you think a model is appropriate, pros and cons I could consider.

Thank you so much.

#2

Sounds like you want to do some spatial-temporal modeling. Maybe you can take a look at Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework
for Traffic Forecasting and see if this can be useful.

1 Like
#3

@mufeili, thank you Mufei, I’ve read the article you suggested.
I went through the math, I grasped the key elements but I lack a background to implement and replicate the described method myself.

I read comments of authors about capability to generalise, both conceptually as well as in tems of large graphs, and I’d liek to experiment on two projects.

Have you tested it?

Do you know if there is any code available for Spatio-temporal GCN ?

__

Edited:

It would be much useful git DGL could offer a mockup code among examples, for replicating results of Yu, Yin and Zhu, so that one could learn from there.
Is there maybe in the examples?

As example, looking at:

and the paper you suggested, which describes a model completely based on convolutions, GCN may be the only model required.
As a beginner and with my current background, a code implementing the basic of the model will be definitely helpful to understand the logic described in the math passages.

#4

The authors released their code here: https://github.com/VeritasYin/STGCN_IJCAI-18.

It will be nice to have a DGL implementation, but I cannot guarantee you if and when it will be ready.

1 Like
#5

well meanwhile thank you for the hints!