Application for regression tasks

chao1224 / SGNN-EBM

Structured Multi-task Learning for Molecular Property Prediction, AISTATS'22 (https://proceedings.mlr.press/v151/liu22e.html)

https://chao1224.github.io/SGNN-EBM

MIT License

13 stars 0 forks source link

Application for regression tasks #3

Open sangyeon-hits opened 1 year ago

sangyeon-hits commented 1 year ago

Hello! Is there a way to train a model for multi-task regression? And will it be desirable to do so with SGNN-EBM?

I'd like to train a property predictor that predicts two or so properties for molecules, where those properties are presumably highly correlated, considering their similar physics.

I'd be grateful to have some advice about the two questions.

chao1224 commented 1 year ago

Hi @sangyeon-hits,

Yes, SGNN-EBM can be naturally adopted for regression tasks. The high-level idea is as follows:

For SGNN in Eq 8, the objective stays the same with neg-log-likelihood. The only change is MSE instead of BCE.
For EBM in Eq 9 10, the objective also stays the same.

sangyeon-hits commented 1 year ago

Thank you very much for the quick reply.

Great to know it'll also work for regressions.

Could you advise which parts of the code I should amend? I've quickly looked through the code, but it seems a bit involved, so I'd like to get some help.