Unfortunately the travis failure highlights that itertools.combinations_with_replacement needs to be backported in sklearn.utils.fixes to support Python 2.6.

The great thing about this approach is that we can turn the actual polynomials (x², y²) off and get only the interaction terms (xy) by substituting combinations for combinations_with_replacement.

Wouldn't it be interesting to add an option in the public API to let the user choose between the 2 (only interaction terms or all degree 2 terms)?

Coverage decreased (-0.04%) when pulling c77e3f6 on larsmans:faster-poly-features into c3ed6c2 on scikit-learn:master.

Tests pass, except the known OMP failure on 2.6. Added a second commit for interaction features only.

Thanks! +1 for merge.

I forgot to ask, have you quickly benchmarked the empirical runtime complexity of this method vs master?

Current master:

>>> %timeit PolynomialFeatures._power_matrix(3, 2, False)
1000 loops, best of 3: 244 us per loop
>>> %timeit PolynomialFeatures._power_matrix(10, 3, False)
1 loops, best of 3: 847 ms per loop
>>> PolynomialFeatures._power_matrix(30, 2, False)
# Ctrl+C after waiting 1m30

This PR:

>>> %timeit PolynomialFeatures._power_matrix(3, 2, False, False)
1000 loops, best of 3: 337 us per loop
>>> %timeit PolynomialFeatures._power_matrix(10, 3, False, False)
100 loops, best of 3: 9.73 ms per loop
>>> %timeit PolynomialFeatures._power_matrix(30, 2, False, False)
100 loops, best of 3: 17.1 ms per loop

Thanks!

Neat! I'm not sure about the clarity of range(not include_bias, ...) but otherwise this LGTM.

Coverage decreased (-0.04%) when pulling de42d69 on larsmans:faster-poly-features into 9f468f3 on scikit-learn:master.

Merged after taking care of @jnothman and @arjoly's concerns.

@hamsal reports an issue that appears to result from this merge at
#3203 (comment)

On 6 June 2014 02:03, Lars Buitinck [email protected] wrote:

Merged after taking care of @jnothman https://github.com/jnothman and
@arjoly https://github.com/arjoly's concerns.

—
Reply to this email directly or view it on GitHub
#3239 (comment)
.