The Latent Causality Alignment (LCA) is a model designed for time series domain adaptation. Based on the characteristic that high-dimensional data is generated by low-dimensional latent variables, it restores stable latent causal mechanisms and solves the difficulties faced by traditional methods in constructing causal structures when dealing with high-dimensional time series data. It has demonstrated good performance in time series classification and prediction tasks on multiple benchmark tests.
Overall structure of LCA
-
Set up the running environment. Install Python and PyTorch, see Install for details.
-
Download data. You can obtain all the datasets from [Google Drive(TSForcast)] or [AdaTime(TSClassif)] or [TranSVAE(VideoClassif)].
-
Train the model. Run the script in the scripts directory, you can run the shell command to execute the experiment.Examples are as follows:
# cd TSForcast/
bash ./scripts/human_action.sh
# cd TSClassif/
bash ./scripts/HAR.sh
# cd /VideoClassif
bash ./scripts/humdb_ucf.sh