I am working on training a Seq2Seq Variational Autoencoder (VAE) model using healthcare data. In my dataset, I have features that exhibit varying levels of variance across patients. For instance, blood glucose values are highly variant for each patient, while HbA1c levels (which reflect an average of blood glucose over around three months) are less variant.
My issue is that the VAE model is not accurately reconstructing the highly variant blood glucose feature, whereas it performs well on reconstructing the less variant HbA1c feature. To address this, I want to assign a higher weight to the blood glucose reconstruction loss, which is calculated using gamma Negative Log-Likelihood (NLL), compared to the HbA1c reconstruction loss.
Is it appropriate to weight the VAE's NLL reconstruction loss for each feature independently?
