The code repository for SOFTS# (SOFTSsharp) in PyTorch. The manuscript is currently in preparation.
SOFTS# extends SOFTS with a stochastic variable-position encoding procedure and multiple dropout layers inside the aggregation-redistribution component, aiming to improve long-term forecasting accuracy while preserving linear computational complexity.
This work builds on the original SOFTS repository (https://github.com/Secilia-Cxy/SOFTS) and its SOFTS++ predecessor with minimal changes (https://github.com/hljubic/SOFTSpp).
The preceding SOFTS++ work has been published in Intelligent Data Analysis: https://journals.sagepub.com/doi/10.1177/1088467X251380055
Full results are in the experiment_results/ directory and image below.
You can directly download the datasets used in the paper from [Google Drive] (https://drive.google.com/drive/folders/1QPM7MMKlzVffdzbGGkzARDuIqiYRed_f?usp=drive_link) or NJU Box.
Once downloaded, place the datasets under folder dataset/, like dataset/ETT-small, dataset/traffic, etc.