JointNets

JointNets is an end-to-end R package for An end-to-end package for learning multiple sparse Gaussian graphical models and nonparanormal models from Heterogeneous Data with Additional Knowledge. This package supports joint estimation, visualization, simulation and evaluation.

It contains the following estimators: SIMULE, WSIMULE, JEEK, DIFFEE and kDiffNet.

Table of Contents

• Installation
• Usage
  • Estimation
    • Methods
    • Dataset
    • Joint Estimation
    • Difference Estimation
  • Visualization
    • plot
    • plotbrain
    • Demo
    • Sample plots
  • Simulation and Evaluation
    • Simulation
    • Evaluation
      • F1
      • BIC
      • AUC
      • Demo
  • GUI
• License
• Getting Help

Installation

1. From CRAN

   install.packages('JointNets', dependencies=TRUE)
   library(JointNets)
   

2. From the GitHub repo. You need to first install the devtools package

   install.packages("devtools")
   

   Load the devtools package

   library(devtools)
   

   Install from Github

   install_github('QData/JointNets', dependencies = TRUE)
   library(JointNets)
   

Usage

Estimation

Methods

JointNets currently support six estimators: simule, wsimule, fasjem, jeek, diffee and diffeek and they serve different estimation purposes:

• joint estimation from multiple contexts:
  • simule and fasjem
• joint estimation from multiple contexts, incorporating prior knowledge:
  • wsimule and jeek
• difference estimation from two contexts:
  • diffee
• difference estimation from two contexts, incorporating prior knowledge:
  • diffeek

Dataset

Each estimator function takes input as a list of data matrices (for multiple contexts) or two separate data matrices (for difference estimation).

The data matrix should be arranged so that each row represents a cell of data (eg, matrix of size 500 x 20 represents 500 instances of data of dimensionality 20).

JointNets comes with a couple of synthetic and real-world datasets for demonstration purposes.

To load them, please run:

data(exampleData)                ## 2 contexts, dimensionality 100
data(cancer)                     ## 2 contexts, dimensionality 26
data(nip_37_data)                ## 2 contexts, dimensionality 37
data(ABIDE_aal116_timeseries)    ## 2 contexts, dimensionality 116

Joint Estimation

simule, fasjem, wsimule and jeek are intended for joint estimation from multiple contexts.

Take simule as an example.

First load a dataset:

data(nip_37_data)

Then simply run the function simule:

result = simule(
  X = nip_37_data,
  lambda = 0.13,
  epsilon = 0.5,
  covType = "kendall",
  parallel = TRUE
)

The estimator simule will return to result a list of named elements:

• result$graphs
  • a list of graphs for each context (in the format of covariance/correlation matrices)
  • In this example, result$graphs is a list of two matrices of size 37x37
• result$share
  • shared graph (in the format of covariance/correlation matrices) among multiple contexts
  • In this example, result$share is a matrix of size 37x37
  • $share is NULL for jeek and fasjem (They do not explicitly estimate for shared graph among contexts. For more details please read FASJEM andJEEK)

To learn about the parameters for each estimator, please try run ? followed by the estimator name in the R console:

?simule
?wsimule
?jeek
?fasjem

Difference Estimation

diffee and diffeek are intended for difference estimation from two contexts.

Take diffee as an example.

First load a dataset:

data(nip_37_data)

Then simply run the function diffee:

result = diffee(C = nip_37_data[[1]], D = nip_37_data[[2]] , lambda = 0.0035))

The estimator diffee will return to result a list of named elements:

• result$graphs
  • a list of the difference graph between two contexts (in the format of covariance/correlation matrix)
  • In this example, result$graphs is a list of a single matrix of size 37x37
• result$share
  • NULL
  • Difference estimation doesn't produce shared graph among contexts. For more details please read diffee

To learn about the parameters for each estimator, please run ? followed by the estimator name in the R console:

?diffee
?diffeek

Visualization

JointNets provides a general purpose plot function as well as a specialized plotbrain for 3d brain graph visualization.

plot and plotbrain are intended for results produced by any of the JointNets estimators.

plot

Take the result from simule estimator as an example.

First, produce a graph estimation:

data(nip_37_data)
result = simule(
  X = nip_37_data,
  lambda = 0.13,
  epsilon = 0.5,
  covType = "kendall",
  parallel = TRUE
)

Then simply run plot on result. It will plot all graphs from the joint estimation with a descriptive title as well as color legend to distinguish between different contents/tasks.

plot(result)

plot extends plot.igraph from the igraph package and there are a lot of options to view and manipulate the graph visualization.

For example:

graph = returngraph(result)                                       ## get the graph from result
layout = layout_nicely(graph, dim = 2)                            ## produce a layout from graph

## plot task specific graph from context 1 using layout
plot(result, type = "taskspecific", subID = 1, layout = layout)

nodeid = which(label %in% c("data", "probability"))               ## obtain node id for data and probability

## Zoom into the plot at nodes data and probability using the same layout
plot(result, type = "neighbour", index = nodeid, layout = layout)

Since estimators in JointNets produce two types of estimation (difference and multiple), plot takes in slightly different parameters for each estimation.

For more details on the parameters in plot, please run ?plot. followed by the estimator name in the R console.

For example:

?plot.simule

plotbrain

plorbrain is similar to plot but instead plots graph in 3d with a brain background as reference.

Please try demo(brain) or ?plotbrain. followed by the estimator name to learn the function.

Demo

JointNets provides demos for various estimators and visualization options.

Please try demo() with the estimator name. (eg, demo(simule))

Sample plots

<img src="JN_demo.png" width="400"> <img src="JN_brain.png" width="400">

Simulation and Evaluation

JointNets offers both a simulation and evaluation module for users to produce synthetic graphs and datasets and evaluate the effectiveness of the estimators based on the simulated ground truth.

Simulation

Simulating sparse Gaussian graphical models (in the format of covariance/correlation matrix) as well as generating samples from the same graphical models are encapsulated in a single function simulation.

To simulate 3 related sparse Gaussian graphical models of dimensionality 20 and at the same time generate 100 slices of samples for each context:

## s and ss control sparsity of all graphs and shared graph, respectively
simulationresult = simulation(p = 20, n=c(100,100,100), seedNum = 37, s = 0.1, ss = 0.1)

The function simulation will return to simulationresult a list of named elements:

• $simulatedgraphs
  • a list of named elements $graphs and $share
    • $graphs is a list of simulated graphs (in the format of covariance and correlation matrices)
    • $share is the shared graph among all contexts
• $simulatedsamples
  • a list of data matrices

To access graphs and samples from simulationresult:

simulationresult$simulatedgraphs$graphs   ## access graphs
simulationresult$simulatedgraphs$share    ## access shared graph
simulationresult$simulatedsamples         ## access generated samples

Evaluation

JointNets offers three score-based metrics for graphical model estimation evaluation: F1, BIC and AUC.

F1

To calculate F1 score from the estimation result obtained by simule and the ground truth:

truth = simulationresult$simulatedgraphs
result = simule(simulationresult$simulatedsamples, 0.2, 0.5, covType = "cov", TRUE)
F1(result,truth)

F1 returns a list of scores for each graph and shared graph (if applicable).

BIC

To obtain BIC score from the estimation result obtained by simule and the simulated samples:

BIC(simulationresult$simulatedsamples,result)

BIC returns a single score for the entire estimation (currently not support for difference estimation)

AUC

To compute AUC score for a sequence of estimations with varying lambdas:

AUC_result = A