I am computing a GNN, where the target feature of some nodes also serves as an input feature to the other nodes. This means that I not only standardize the features but also think it would be best to standardize the target feature.
However, for the time being, this means, that when computing my training error, I am computing the error between the prediction and the standardized ground truth of the target feature. But for my understanding, computing the loss with the non-standardized ground truth would be much better. However, I am struggling with how to apply inverse transform before I backpropagate.