The GPCCA package provides functions to implement Generalized Probabilistic
Canonical Correlation Analysis (GPCCA) on multi-modal dataset. GPCCA is
an unsupervised method designed to perform data integration and joint
dimensionality reduction on a dataset consisting of multiple modalities.
The package can be installed from GitHub with:
devtools::install_github("Kaversoniano/GPCCA")A tutorial of GPCCA comes as a built-in vignette with the installed package:
devtools::install_github("Kaversoniano/GPCCA", build_vignettes = TRUE)GPCCA is a probabilistic model built based on the Gaussian latent variable framework, serving as an extension of the probabilistic CCA (PCCA) with several key merits:
- It generalizes PCCA from two to more than two modalities.
- It learns joint low-dimensional embeddings probabilistically.
- It imputes missing values inherently during its parameter estimation.
GPCCA is capable of handling incomplete data with various patterns of missing values. See below for an illustration of the GPCCA model applied to a three-modality dataset. White boxes are used to indicate missing data and the missing pattern of each data modality is described as follows:
- Modality 1 contains random missing values
- Modality 2 is fully observed and complete
- Modality 3 involves modality-wise missingness
vignette("tutorial", package = "GPCCA")Please see the package vignette for:
- Generation of example multi-modal datasets
- How to fit the GPCCA model on a multi-modal dataset
- How to extract the *learned latent factors*
- Example downstream analysis on the *joint latent embeddings*
- Visualizations of the *low-dimensional embeddings*
- Model selection on the target dimension, i.e., the number of latent factors
GPCCAmodel is the primary function for users. It implements GPCCA and fit it
to a multi-modal dataset, which generates the joint low-dimensional embeddings.
For high-dimensional data, feature selection is strongly recommended so
that only the highly informative features are included in the model fit.
This helps to reduce the computational resources required and boost the
computational speed.
GPCCAselect provides a built-in technique of model selection on
target dimension (hyper-parameter). Its output gives the optimal choice(s)
of the target dimension based on a given set of candidate values.
GPCCAmodel2 is another primary function for users. It implements GPCCA
extension models: GPCCA-Pois and GPCCA-Bern, where
GPCCA-Pois is designed for multi-modal count data and
GPCCA-Bern is designed for multi-modal binary data. By contrast,
the original GPCCA model is developed for general continuous data,
which is implemented in a separate framework by GPCCAmodel.