MVP for regression task

[Roxyi]

Hi, the default mask of MVP is subsequence_mask, which may result in data leakage in a regression task(?). If the downstream task is a regression task, I should set future_mask to True and subsequence_mask to False to pretrain an MVP. Am I right?

[oguiza]

Great question, @Roxyi!
The only way to know it is to try it and see the performance in the test set.
However, I'd say that if your task is truly a regression task I don't think there's a risk of leakage and you could use the default mask. If your task is a forecasting task, then there's a risk, and you should use a future mask.
By a regression task, I mean one where the label is a continuous value for each univariate or multivariate time series. For example, if you need to estimate the user's age based on his/her resting heart rate. In this case, the label is not contained in the time series.

[Roxyi]

Thank you! I will try it.