binopt

This package is aiming to categorize labeled data in terms of a global figure of merit. In high energy physics, categorization of collision data is done by maximizing the discovery significance. This package run on unbinned binary datasets.

installation

Install like any other python package:

pip install binopt --user

or:

git clone git@github.com:yhaddad/binopt.git
cd binopt/
pip install .

Getting started

sevent = 1000
bevent = 10000
X = np.concatenate((
             expit(np.random.normal(+2.0, 2.0, sevent)),
             expit(np.random.normal(-0.5, 2.0, bevent))
))
Y = np.concatenate((
             np.ones(sevent),
             np.zeros(bevent)
))
W = np.concatenate((np.ones(sevent), np.ones(bevent)))

binner = binopt.optimize_bin(
             nbins=3, range=[0, 1],
             drop_last_bin=True,
             fix_upper=True,
             fix_lower=False,
             use_kde_density=True
)
opt = binner.fit(
             X, Y, sample_weights=W,
             method="Nelder-Mead",
             breg=None, fom="AMS2"
)

print "bounds : ", opt.x
print "signif : ", binner.binned_score(opt.x)
print "Nsig   : ", binner.binned_stats(opt.x)[0]
print "Nbkg   : ", binner.binned_stats(opt.x)[1]