Collective Non-negative Tensor Factorization (AAAI-19)

This repo contains the PyTorch implementation of the paper Learning Phenotypes and Dynamic Patient Representations via RNN Regularized Collective Non-negative Tensor Factorization in AAAI-19. [paper] [dataset]

Requirements

The codes have been tested with the following packages:

• Python 3.7
• PyTorch 0.4.1

Quick Demo

To run the model with a quick demo data, simply clone the repo and decompress the data archive by executing the following commands:

git clone git@github.com:jakeykj/cntf.git
cd cntf
tar -xzvf demo_data.tar.gz
python train.py demo_data/

A folder ./results/cntf_hitf/ will be automatically created and the results will be saved there.

Use python train.py --help to obtain more information about setting the parameters of the model.

Data Format and Organization

The data are stored in two seperate files contained in a folder (we refer to its path by <DATA_PATH>): <DATA_PATH>/data.pkl and <DATA_PATH>/list.pkl. They should be saved using the pickle package in Python, and the format of the two files are described as follows:

• data.pkl: contains a list of the temporal sparse tensor, one for each patient. It is a python list of dictionary with six keys. subs is a torch.LongTensor object containing the subscripts of the non-zero elements of the tensor (for a 3rd-order tensor, its size should be 3-by-K, where K is the number of non-zero elements). vals is a torch.FloatTensor object containing the values of the tensor that are corresponding to the non-zero subscripts (size: K). size is a torch.Size object describing the size of the temporal sparse tensor. hadm_id is the unique ID of the patient (a python interger). dx_vector is a torch.FloatTensor object containing the one-hot vector of the diagnoses of the patient. rx_vector is a torch.FloatTensor object containing the counting vector of the medications of the patient.
• list.pkl: contains the temporal length and label information of the patient. It is a python list of nested python lists, one for each patient. The format is: [[<hadm_id>, <temporal_length>, <label>], ...].

If you use other datasets, you can organize the input data in the same format described above, and pass the <DATA_PATH> as a parameter to the training script:

python train.py <DATA_PATH>

Citation

If you find the paper or the implementation helpful, please cite the following paper:

@inproceedings{yin2019learning,
  title={Learning phenotypes and dynamic patient representations via {RNN} regularized collective non-negative tensor factorization},
  author={Yin, Kejing and Qian, Dong and Cheung, William K. and Fung, Benjamin C. M. and Poon, Jonathan},
  booktitle={Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence},
  pages={1246--1253},
  year={2019},
  organization={AAAI Press}
}

Contact

If you have any enquires, please contact Mr. Kejing Yin by email: cskjyin [AT] comp [DOT] hkbu.edu.hk, or leave your questions in issues.

---

👉 Check out my home page for more research work by us.