Open tangxiaoyue163 opened 1 month ago
Canonical pre-alignment means using the representation method described in the paper's section 3.2. You can find the code for such canonical pre-alignment representation in the function 'variable_time_collate_fn' in lib/parse_datasets.py.
After the canonical pre-alignment, time and mask will have the same sequence length as the IMTS, then you can regard them as additional variables. As forecasting queries have different sequence length to the IMTS, we postpend them to the IMTS when we conduct forecasting.
Could you please introduce in detail the experimental methods of MODELS FOR MTS FORECASTING in the baseline? How to understand the thesis "For MTS forecasting models, we input sequences after canonical pre-alignment and incorporate the observed time, mask,information, and forecasting queries as additional features into these models. " in the paper? ①Does "canonical pre-alignment" mean first using timesnet to imputation the complete sequence and then input the complete sequence for subsequent prediction? ②Does "as other features" mean directly viewing observed time, mask, information, and forecasting queries as multivariate variables? Thank you for your work and great support for my experiment!Looking forward to your reply!