Functional-Bayesian-Tucker-Tensor

This authors' official PyTorch implementation for paper:"Functional Bayesian Tucker Decomposition for Continuous-indexed Tensor Data"[OpenReview][Arxiv] (ICLR 2024).

Key Idea: Decompose Continuous-indexed Tensor Data as mode-wise latent functions in Tucker format

Each function is modeled as a group of State-Space Gaussian Processes(SSGPs):

Example of latent functions of spatial and temporal modes learned from real-world data.

Requirements:

The project is mainly built with pytorch 1.13.1 under python 3.10. Besides that, make sure to install tqdm and tensorly==0.70 before running the project. The detailed package info can be found in requirement.txt.

Instructions:

Clone this repository.
To play with the model quickly, we offer several notebooks at notebook(on synthetic & real data)
To run the real-world datasets with scripts, see script_Funbat.sh and script_Funbat_CP.sh for example.
To tune the (hyper)parametrs of model, modify the .yaml files in config folder
To apply the model on your own dataset, please follow the process_script. or generating-synthetic-data to process the raw data into appropriate format.
GPU choice: the models are run on CPU by default, but you can change the device to GPU by modifying the device as gpu of .yaml files in the config folder.

Data

We offer the raw data, processed scripts, and processed data(Beijing,US-TEMP) for all four datasets used in paper. The code for generating the synthetic data is also provided in the data folder.

If you wanna customize your own data to play the model, please follow the process_script.

Citation

Please cite our work if you would like it

@article{fang2023functional,
  title={Functional Bayesian Tucker Decomposition for Continuous-indexed Tensor Data},
  author={Fang, Shikai and Yu, Xin and Wang, Zheng and Li, Shibo and Kirby, Mike and Zhe, Shandian},
  journal={arXiv preprint arXiv:2311.04829},
  year={2023}
}