Gaussian Process

Gaussian process classification, regression, and Cox regression for time-to-event data This package has not yet registered in the METADATA. To install, run the following command.

# Install
Pkg.clone("git://github.com/linxihui/GaussianProcess.jl")

Since it depends on the newest GLMNet, which is in my branch and has not been merge to the main branch, one has to clone and build the package as

Pkg.clone("git://github.com/linxihui/GLMNet.jl")
Pkg.build("GLMNet")

Syntax

The main function of this package is gausspr, which have two methods like R

• gausspr(::Formula, ::DataFrame, ::Distribution)
• gausspr(::Array, ::Array, ::Distribution)

In formula, symbol .. is supported, which means all other variables in DataFrame, equivalent to the . symbol in R formula. For example y ~ .. means y is the response and all other variable except y in the DataFrame are predictors.

For Distribution, currently 5 different types are supported.

• Normal() : regression of continuous response
• Binomial() and Multinomial(): classifications
• Poisson(): count data
• CoxPH(): survival or time-to-event data

Different Kernels are available, including

• kernelLinear() or kernelVanilla()
• kernelPoly(degree=3, offset=1.0)
• kernelRBF(σ=1.0): σ is the standard deviation
• kernelLaplace(σ=1.0): σ is the scale parameter
• kernelMatern(ν=2.0, σ=1.0)
• kernelUser(K): K is a user defined kernel matrix

A Quick Guide

julia> using DataFrames, GaussianProcess

julia> dat = DataFrame(
           A = [0.4, 0.1, 0.7, 0.2],
           B = ["b", "a", "b", "a"],
           C = [0.3, 0.8, 0.1, 0.6],
           D = ["Yes", "No", "Yes", "No"]
           );

julia> gp = gausspr(D ~ A + B + C, dat, Binomial());

# predict class probabilities
julia> pred = predict(gp, dat, outtype = :prob)
4x2 Array{Float64,2}:
 0.432865  0.567135
 0.597388  0.402612
 0.394668  0.605332
 0.575207  0.424793

julia> pred_class = predict(gp, dat, outtype = :class)
4-element Array{String,1}:
 "Yes"
 "No"
 "Yes"
 "No"

# use x-y input (regression)
julia> x, y = modelmatrix(A ~ B + C + D, dat)
(
4x5 Array{Float64,2}:
 0.0  1.0  0.3  0.0  1.0
 1.0  0.0  0.8  1.0  0.0
 0.0  1.0  0.1  0.0  1.0
 1.0  0.0  0.6  1.0  0.0,

[0.4,0.1,0.7,0.2],(("a","b"),(),("No","Yes")),())


julia> gp2 = gausspr(x, y, Normal(0, 1), kernel = kernelPoly(3, 1));

julia> pred2 = predict(gp2, x)
4-element Array{Float64,1}:
 0.409322
 0.0996975
 0.690604
 0.200377 

Binary Classification

julia> using RDatasets

julia> srand(123)

julia> kyphosis = dataset("rpart", "kyphosis");

julia> kyphosis[:Kyphosis] = convert(Array, kyphosis[:Kyphosis]);

julia> iTrain = sample(1:size(kyphosis, 1), 40, replace = false);

julia> iTest = setdiff(1:size(kyphosis, 1), iTrain);

julia> ky_mod = gausspr(Kyphosis ~ .., kyphosis[iTrain, :], Binomial());

julia> ky_pred = predict(ky_mod, kyphosis[iTest, :], outtype = :prob);

julia> out = ROC(kyphosis[iTest, :Kyphosis].data, ky_pred[:, 2]);

julia> out.auc
0.8715277777777777

julia> plot(out)

Regression

julia> srand(456)

julia> boston = dataset("MASS", "Boston");

julia> iTrain = sample(1:size(boston, 1), 300, replace = false);

julia> iTest = setdiff(1:size(boston, 1), iTrain);

julia> mod_boston = gausspr(MedV ~ .., boston[iTrain, :], kernel = kernelMatern(2, 1));

julia> pred_boston = predict(mod_boston, boston[iTest, :]);

julia> err_boston = sqrt(mean((convert(Array, boston[iTest, :MedV]) - pred_boston).^2))
3.3446708896894983