Communication-Efficient Federated Linear and Deep Generalized Canonical Correlation Analysis (GCCA)

Implementation of the paper "Communication-Efficient Federated Linear and Deep Generalized Canonical Correlation Analysis" published in IEEE Transactions on Signal Processing. A shorter conference version of the paper was published in ICASSP 2022, found here.

Federated GCCA setting

Results

Synthetic data experiment

Real data experiment in EHR data

Setup

Currently, Linear federated GCCA (CuteMaxVar) is implemented in MATLAB whereas Deep federated GCCA (Deep-CuteMaxVar) is implemented in python. The folder gcca contains the source code for CuteMaxVar, whereas dgcca contains the source code for Deep-CuteMaxVar.

For Linear GCCA

Download tensor_toolbox and make sure it is in your matlab path.

For Deep GCCA, run the following command to install the required packages

conda install pytorch torchvision -c pytorch pandas scikit-learn scipy tqdm jupyter matplotlib

Executing the Code

For CuteMaxVar:

$ cd gcca $ matlab -nodisplay -nosplash -nodesktop -r "run('demo.m');exit;"

For Deep-CuteMaxVar:

$ cd dgcca

The folder dgcca contains two subfolders toy_2d and ehr. The former contains the code for the synthetic data experiment, whereas the latter contains the code for the EHR data experiment. Training and visualization for synthetic data experiment is in

dgcca/toy_2d/toy_dgcca.ipynb

Whereas training and visualization for EHR data experiment is in

dgcca/ehr/ehr_dgcca.ipynb