MrRCE

Implementation of Capturing Between-Tasks Covariance and Similarities Using Multivariate Linear Mixed Models, Volume 14, Number 2 (2020), Electronic Journal of Statistics.

Specifically, this is an implementation of the Multivariate random Regression with Covariance Estimation (MrRCE) algorithm, designed to take advantage of correlations and similarities among responses and coefficients, in a multi-task regression framework (see the paper for details).

Useful Links

• Paper
• GitHub repo
• Example

Installation and Requirements

Clone this repo and run

pip install -r requirements.txt

Simulations

There are five simulations that can be easily executed:

• Autoregressive (AR) error covariance with dense coefficient matrix (ar_dense)
• AR error covariance with sparse coefficient matrix (ar_sparse)
• Fractional Gaussian Noise (FGN) error covariance (fgn)
• Equicorrelation error covariance (equi)
• Identity error covariance (identity)

Running the simulations will create a file with the name simulation_results_<simulation name>.csv with Model Error (ME) for each method and replication. In addition, it will create a plot of ME against the correlation parameter, and save it as simulation_plot_<simulation name>.png. The files will be saved into a results and plots folders.

Running Simulations

For help run:

$ python run_simulation.py --help

usage: run_simulation.py [-h] --simulation-name SIMULATION_NAME [--n N]
                         [--output-path OUTPUT_PATH] [--save-data]

MrRCE simulations.

optional arguments:
  -h, --help            show this help message and exit
  --simulation-name SIMULATION_NAME
                        simulation mane, one of ['ar_dense', 'ar_sparse',
                        'fgn', 'equi', 'identity']
  --n N                 number of repetitions
  --output-path OUTPUT_PATH
                        output folder
  --save-data           whether to save the simulation data

python run_simulation.py --simulation-name <simulation name>

This will run the simulation with the default 200 replication. You can also run:

python run_simulation.py --simulation-name <simulation name> --n <N>

where <N> is an integer for the number of replications. For example, the following line,

python run_simulation.py --simulation-name equi --n 200

will run the equicorrelation (covariance matrix) simulation with 200 replications (for each value of the correlation coefficient, rho), and the outcome should look like the following:

Example

Example of running MrRCE:

from mrrce import MrRCE
mrrce = MrRCE()
mrrce.fit(X, Y) # X and Y are matrices of shapes (n,p) and (n,q) correspondingly
mrrce.Gamma     # estimated coefficient matrix
mrrce.rho       # estimated correlation coefficient
mrrce.sigma     # estimated sd for coefficients
mrrce.Sigma     # estimated covariance matrix for the error terms
mrrce.Omega     # estimated precision matrix for the error terms

See full example at this notebook.

Citation

If you find MrRCE to be useful in your own research, please consider citing the following paper:

@ARTICLE{NavRos2020,
    AUTHOR = {Aviv Navon and Saharon Rosset},
     TITLE = {Capturing between-tasks covariance and similarities using multivariate linear mixed models},
   JOURNAL = {Electron. J. Statist.},
  FJOURNAL = {Electronic Journal of Statistics},
      YEAR = {2020},
    VOLUME = {14},
    NUMBER = {2},
     PAGES = {3821-3844},
      ISSN = {1935-7524},
       DOI = {10.1214/20-EJS1764},
      SICI = {1935-7524(2020)14:2<3821:CBTCAS>2.0.CO;2-2},
}