Hi Daisy,

It looks like you need to use labels +-1, rather than 0 and 1. There's also
an easier way of writing logistic regression in Convex.jl. If you pull the
most recent version from master, this should work:

using DataFrames
using RDatasets
using Convex
using SCS

iris = dataset("datasets", "iris")
iris[:Y] = [species == "versicolor" ? 1.0 : -1.0 for species in
iris[:Species]]
X = [ones(size(iris, 1)) iris[:SepalLength] iris[:SepalWidth]
iris[:PetalLength] iris[:PetalWidth]]
Y = array(iris[:Y])

Solve the NLP using Convex.jl and SCS packages

reference:

http://cvxr.com/cvx/examples/cvxbook/Ch07_statistical_estim/html/logistics.html
n, p = size(X)
beta = Variable(p)
problem = minimize(sum([logistic_loss(-Y[i]_(X[i,:]_beta)) for i=1:n])) #
-0.01*norm(beta,1)

scs = SCSSolver()
solve!(problem, scs)

and you can plot the solution like this:

using Gadfly
perm = Base.Sort.sortperm(vec(X_beta.value))
set_default_plot_size(25cm, 12cm)
plot(layer(x=1:n,y=(Y[perm]+1)/2,Geom.point),layer(x=1:n,y=logistic(X_beta.value)[perm],Geom.line))

[image: Inline image 1]

Best,
Madeleine

On Wed, Jan 7, 2015 at 10:09 AM, Daisy Zhuo [email protected]
wrote:

Hello!

I am using convex.jl to test and see if it works on a simple maximum
likelihood estimator for logistic regression. It is giving me unbounded
results though - I am not sure if it's an issue with my formulation or
something else. Would you be able to take a quick look and let me know
where I may have made a mistake? Thank you very much in advance!

using DataFramesusing RDatasetsusing Convexusing SCS

iris = dataset("datasets", "iris")
iris[:Y] = [species == "versicolor" ? 1.0 : 0.0 for species in iris[:Species]]
X = [ones(size(iris, 1)) iris[:SepalLength] iris[:SepalWidth] iris[:PetalLength] iris[:PetalWidth]]
Y = array(iris[:Y])

Solve the NLP using Convex.jl and SCS packages# reference: http://cvxr.com/cvx/examples/cvxbook/Ch07_statistical_estim/html/logistics.html

n, p = size(X)
beta = Variable(p)
problem = maximize(Y'_X_beta-sum(logsumexp([X_beta zeros(n)]))) # -0.01_norm(beta,1)

scs = SCSSolver()solve!(problem, scs)

Output:

Setup time: 8.27e-04s

Iter | pri res | dua res | rel gap | pri obj | dua obj | kap/tau | time (s)

 0|      inf       inf       nan      -inf       inf       inf  3.44e-03
40|      inf       inf       nan      -inf      -inf       inf  5.09e-03

---

Status: Unbounded

—
Reply to this email directly or view it on GitHub
#46.

Madeleine Udell
PhD Candidate in Computational and Mathematical Engineering
Stanford University
www.stanford.edu/~udell

from convex.jl.

Hi Madeleine,

This works like a charm. Logistic loss and using +1/-1 labels does make it look neater and run correctly.

Thank you very much!

Cheers,
Daisy

from convex.jl.