Question about MAE

[How-Will]

Hi, Wenjie

def masked_mae_cal(inputs, target, mask):
    """ calculate Mean Absolute Error"""
    return torch.sum(torch.abs(inputs - target) * mask) / (torch.sum(mask) + 1e-9)

I have a little doubt about the calculation of MAE. I found you normalizes the dataset with standard scaling, it means the target and input are standard normalized. So why not calculate MAE after inverse the scaling to them?

[WenjieDu]

Hi there,

Thank you so much for your attention to SAITS! If you find SAITS is helpful to your work, please star⭐️ this repository. Your star is your recognition, which can let others notice SAITS. It matters and is definitely a kind of contribution.

I have received your message and will respond ASAP. Thank you again for your patience! 😃

Best,
Wenjie

[How-Will]

I applied SAITS to my data and found some fluctuations in the imputated data, is this normal?

[WenjieDu]

Hi William, thank you for raising this issue.

For your 1st question about the error metric calculation, sure, you can invert the transformation before calculating, but this asks for one more unnecessary inverse step. Please remember that the purpose of error metrics is to make fair comparisons between all methods. No matter we make the inverse or not, the comparisons in our experiments are impartial.

For the 2nd question about your imputation results, I personally think this is normal because there're fluctuations in your original data, e.g. data from 2013/6/4 to 2013/7/24 in the bottom fig in your screenshot. I know you may want to make the imputation look more smooth, namely, not so many flips. I'd suggest you try to 1). tune the hyperparameters of SAITS to make the model not overfit the original data; 2). try to add additional methods to smooth the imputation, e.g. add additional loss constraints, or add post-processing steps to manually eliminate these too many flips.

[How-Will]

Thanks very much, your reply help me a lot.

[WenjieDu]

My pleasure.