Question regarding observed mask and predicting a "gap" in the data
Hi,
I was wondering if it's possible to use this model to make predictions for a "gap" in the data, i.e., feeding the values before and after and predicting the missing interval? I suppose that maybe the observed mask can play a role to achieve this, but I am not entirely certain if it's possible.
Thanks,
Gil
Thanks, @giltinde for the question!
I suppose you are asking about the imputation task, rather than the prediction one. In theory, these models should be able to do the imputation task but we would need to cook up a separate model for that and get rid of the causal mask in the decoder. Something like a TimeSeriesTransformerForImputationusing Bert style model should be doable I believe.
But to answer your question, right now that is not possible.
Ok, thanks for the reply!
@giltinde I will check how much effort it is to make an imputation model. Also, and you probably know this, but imputation is not the same as prediction, in that with imputation we need information from the past and future values for consistency, whereas with the prediction task, we have information only from the past. Thus it is technically not correct to use a prediction model as an imputer.
Ok, so it wouldn't be possible to devise a model which could handle both imputation and prediction? I.e., have one model which performs imputation, and view prediction as a special case where the window to predict simply isn't followed by any data?