SigmaRidgeRegression.jl

Optimally tuned ridge regression when features can be partitioned into groups.
Author nignatiadis
Popularity
5 Stars
Updated Last
3 Months Ago
Started In
January 2020

SigmaRidgeRegression.jl

Build Status Coverage

Automatically and optimally-tuned Ridge regression when the features may be partitioned into groups.

See the manuscript below for a theoretical description of the method.

Ignatiadis, Nikolaos, and Panagiotis Lolas. "σ-Ridge: group regularized ridge regression via empirical Bayes noise level cross-validation." arXiv:2010.15817 (2020+)

The folder reproduction_code in this repository contains code to reproduce the results of the paper.

Installation

The package is available on the Julia registry (for Julia version 1.5) and may be installed as follows:

using Pkg
Pkg.add("SigmaRidgeRegression")

Example usage

SigmaRidgeRegression.jl can be used alongside the MLJ framework for machine learning in Julia.

using MLJ
using SigmaRidgeRegression
using Random

# Suppose we have three groups of features, each with n observations
# and 25, 50 and 100 features respectively
n = 400
Random.seed!(1)
p1 = 25 ; X1 = randn(n, p1)
p2 = 50 ; X2 = randn(n, p2)
p3 = 100; X3 = randn(n, p3)

# The signal in the regression of the coefficients across these groups varies
α1_sq = 4.0 ;  βs1 = randn(p1) .* sqrt(α1_sq / p1)
α2_sq = 8.0 ;  βs2 = randn(p2) .* sqrt(α2_sq / p2)
α3_sq = 12.0;  βs3 = randn(p3) .* sqrt(α3_sq / p3)

# Let us concatenate the results and create a response
X = [X1 X2 X3]
βs = [βs1; βs2; βs3]
σ = 4.0
Y = X*βs .+ σ .* randn(n)

# Let us make a `GroupedFeatures` object that describes the feature grouping
# !!NOTE!! Right now the features are expected to be ordered consecutively in groups
# i.e., the first p1 features belong to group 1 etc.
groups = GroupedFeatures([p1;p2;p3])

# Create MLJ machine and fit SigmaRidgeRegression:
sigma_model = LooSigmaRidgeRegressor(;groups=groups)
mach_sigma_model = machine(sigma_model,  MLJ.table(X), Y)
fit!(mach_sigma_model)

# How well are we estimating the true X*βs in mean squared error?
mean(abs2, X*βs .- predict(mach_sigma_model))  # 4.612726430034071

# In this case we may compare also to the Bayes risk
λs_opt = σ^2 ./ [α1_sq; α2_sq; α3_sq] .* groups.ps ./n
bayes = MultiGroupRidgeRegressor(;groups=groups, λs=λs_opt, center=false, scale=false)


mach_bayes = machine(bayes, MLJ.table(X), Y)
fit!(mach_bayes)
mean(abs2, X*βs .- predict(mach_bayes)) #4.356913540118585

TODOs

  • Fully implement the MLJ interface.
  • Wait for the following MLJ issue to be fixed: JuliaAI/MLJBase.jl#428 (comment), in the meantime this package uses type piracy as in the linked comment to accommodate a large number of features.