Use MurmurHash3 for interaction features

[stegben]

#23
MurmurHash3 is used by sklearn's FeatureHash, which is fast and robust enough.

Before

         3 function calls in 9.690 seconds

   Ordered by: standard name

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    0.000    0.000 base.py:99(get_shape)
        1    0.000    0.000    0.000    0.000 {method 'disable' of '_lsprof.Profiler' objects}
        1    9.690    9.690    9.690    9.690 {method 'fit' of 'kaggler.online_model.ftrl.FTRL' objects}

After

         3 function calls in 2.265 seconds

   Ordered by: standard name

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    0.000    0.000 base.py:99(get_shape)
        1    0.000    0.000    0.000    0.000 {method 'disable' of '_lsprof.Profiler' objects}
        1    2.265    2.265    2.265    2.265 {method 'fit' of 'kaggler.online_model.ftrl.FTRL' objects}

[stegben]

As for the original index, a function that directly returns the input integer is a good enough hash function, so I remove the hash function for it.

[jeongyoonlee]

Have you tried 'mmh3' python module (https://pypi.python.org/pypi/mmh3/2.3.1) instead of the C++ one? Unless there's a significant performance difference, the python module is preferred for simplicity.

[jeongyoonlee]

Regarding not using hash for original features, I agree with you. There might be more collision for features with low index if the number of features is greater than n, but as long as n is set big enough, which is usually the case, it should be fine.

Thanks for the improvement!

[stegben]

mmh3 looks great, I'll try it in 2 days

[stegben]

Sorry I was on my vacation these days.

I found that if I use mmh3, which cannot be imported by cimport, there will be extra function call overheads. On the other hand, the implementation of it use lots of PyObjects, which is much slower.

[jeongyoonlee]

No worries. Then let's merge this. Thanks for looking into it.