yipenguva / pesca_matlab Goto Github PK

The Matlab implementaion of the pESCA (penalized exponential familay simultaneous component analysis) model with various concave penalties. The pESCA model with the group concave (concave L2 norm) penalty is used in my previous paper "Separating common (global and local) and distinct variation in multiple mixed types data sets", \url{https://arxiv.org/abs/1902.06241}.

knitr::opts_chunk$set(echo = TRUE)

Make sure the current folder is in your Matlab path and run the help function from the Matlab console (e.g. help ESCA_group_concave) for more information on the input/output of the algorithms.

The demos are used to show how to simulate multiple data sets with underlying global, local common and distinct structures according to the ESCA model, how to construct a P-ESCA model, and how to do the model selection. The explanation and the results of these demos can be found in the corresponding html files in the examples fold.

-P-ESCA model on three quantitative data sets:

• demo_PESCA_GGG.m: this demo shows how to do data simulation, model selection and fit the final P-ESCA model for multiple quantitative data sets.

• demo_PESCA_GGG_fullInfo.m: this demo is same as above, except that the simulated parameters are used to evaluate the model selection process.

-P-ESCA model on three binary data sets:

• demo_PESCA_BBB.m: this demo shows how to do data simulation, model selection and fit the final P-ESCA model on multiple binary data sets.

• demo_PESCA_BBB_fullInfo.m: this demo is same as above, except that the simulated parameters are used to evaluate the model selection process.

-P-ESCA model on mixed binary and quantitative data sets:

• demo_PESCA_GBB.m: this demo shows how to do data simulation, model selection and fit the final P-ESCA model for mixed quantitative-binary-binary data sets.

• demo_PESCA_GGB.m: this demo shows how to do data simulation, model selection and fit the final P-ESCA model for mixed quantitative-quantitative-binary data sets.

• The html documents of the above demos published using Matlab are in the examples fold.

The algorithms are used to do data simulation, to fit a P-ESCA model, and to do model selection are in the algorithms fold.

-dataSimulation:

• This folder contains the functions used to do the data simulations.

-functions:

• This folder contains some functions used in the algorithm.

-Dispersion parameter estimation:

• alpha_estimation.m: the developed \alpha estimation procedure.

• SVD_CV_modelSelection.m: missing value based cross validation procedure for the rank selection of a SVD (PCA) model. Mainly used in alpha_estimation.m.

• SVD_missingValues.m: a SVD (PCA) algorithm, which is capable to tackle missing value problem.

-Algorithms to fit a P-ESCA model:

• ESCA_group_concave.m: the main algorithm developed to fit the P-ESCA model.

• ESCA_group_concave_composite.m: the algorithm used to fit a P-ESCA model when composite concave penalty is used in the P-ESCA model. This function is a part of ESCA_group_concave.m.

• ESCA_group_concave_L1.m: the algorithm used to fit a P-ESCA model when concave L1-norm is used in the P-ESCA model. This function is a part of ESCA_group_concave.m.

• ESCA_group_concave_L2.m: the algorithm used to fit a P-ESCA model when concave L2-norm penalty is used in the P-ESCA model. This function is a part of ESCA_group_concave.m. This algorithm is used in the P-ESCA paper.

-Algorithms to select the P-ESCA model:

• ESCA_modelSelection_KCV.m: missing value based cross validation to select the P-ESCA model.

• ESCA_modelSelection_KCV_fullInfo.m: same as the above procedure. And the simulated parameters are used to evaluate the model selection procedure.

• ESCA_modelSelection_KCV_twoSteps_bg.m: the model selection procedure for mixed quantitative and binary data sets.

• ESCA_modelSelection_KCV_twoSteps_Gaussian.m and ESCA_modelSelection_KCV_twoSteps_nonGaussian.m are used in ESCA_modelSelection_KCV_twoSteps_bg.m.

-Others:

• scaling_centering.m: a column centering and scaling procedure for the processing of the real data sets.