LogisticOptTools.jl

Build Status

A Julia package to benchmark optimization solvers on logistic regression problems.

MIT license
Install using julia> ] add LogisticOptTools

Basic usage

Suppose you import LogisticOptTools in your REPL

julia> using LogisticOptTools
julia> const LOT = LogisticOptTools

Suppose you have available a matrix of features X and a vector of observations y, and you want to fit a logistic model onto this data. You could instantiate a new logistic model simply by typing

julia> logreg = LOT.LogisticRegressor(X, y,
                                      fit_intercept=true,
                                      penalty=LOT.L2Penalty(1.0))

and then fit the logistic regression with Optim.jl:

julia> p = LOT.nfeatures(logreg)
julia> x0 = zeros(p) ; algo = LBFGS()
julia> res = Optim.optimize(logreg, x0, algo)
# Fetch optimal parameters
julia> p_opt = res.minimizer

Benchmarks

LogisticOptTools could use the different algorithms implemented in Optim.jl. We depict in the following figure a comparison of three algorithms, when fitting a logistic model on the covtype dataset (581,012 data, 54 features).

For an example on how to use other solvers, we have implemented in examples/tron.jl a resolution of a logistic regression problem with tron, a solver implemented JSOSolvers.jl.

Use-cases

Import LIBSVM datasets

LogisticOptTools supports the libsvm format. Once a dataset downloaded from the website, you could load it in the Julia REPL with

shell> ls .
covtype.binary.bz2
# Parse as Float64
julia> dataset = LOT.parse_libsvm("covtype.binary.bz2", Float64)
# Load as dense matrix
julia> X = LOT.to_dense(dataset)
julia> y = dataset.labels

You could load the dataset as a sparse matrix just by replacing LOT.to_dense with LOT.to_sparse.

Advanced usages

You could find in examples/ a few examples on:

optimizing the L2 penalty parameter with Optim.jl
fitting a sparse regression (l0-l2 logistic regression) with JuMP and a MILP solver