pyrvm: Ranking Vector Machines in Python
Learn and predict orderings of vectors using large-margin criteria.
This is an implementation of the ranking-vector-machine algorithm from Yu, Hwanjo and Kim, Sungchul. "SVM Tutorial: Classification, Regression and Ranking." Handbook of Natural Computing. Springer Berlin Heidelberg, 2012.
RVMs are similar to 1-norm ranking SVMs, but can be trained more efficiently.
Installation
pip install pyrvm
Examples
>>> from math import cos, sin, pi
>>> import numpy as np
>>> from pyrvm import RVM
>>> # Create points that spiral along the z axis, with higher rankings at lower
>>> # values of z.
>>> n_points = 50
>>> points = [[cos(2*pi*5*t), sin(2*pi*5*t), t] for t in np.linspace(0, 1, n_points)]
>>> X = np.array(points)
>>> # This algorithm is sensitive to constant shifts. To see this, uncomment
>>> # this mean subtraction; we'll see the algorithm fail.
>>> X = X - X.mean(0)
>>> y = range(n_points)
>>> # Train a linear RVM using half of the data. We keep the slack penalty C high
>>> # here because we know that the points can be ranked linearly with no errors.
>>> ranker = RVM(C=100.0)
>>> ranker.fit(X[0::2, :], y[0::2])
Out<rvm.RVM at 0x116f2d0d0>
>>> # Since we used a linear kernel, we can determine the weight vector in the
>>> # original space that determines ranking. It should be in the direction
>>> # of -z.
>>> print sum(ranker._alpha[ranker._alpha != 0, np.newaxis] * ranker._rank_vectors, 0)
[ 1.88497489e-05 1.58543033e-05 -2.45000255e+01]
>>> # Now let's see how we do on the other half of the data.
>>> print ranker.predict(X[1::2, :])
[ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24]Documentation
The API reference can be found here.
Dependencies
pyrvm depends on numpy, sklearn, pulp, and (assuming default use)
GLPK.
