Entropy aggregation primitive

[twdobson]

Entropy aggregation primitive

For details on entropy please see:
https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.entropy.html

[twdobson]

Hi @rwedge, had issues with assigning github account to my commits with my previous PR, so I redid the PR.

Previously when running 'make html' the Entropy aggregation primitive was not automatically generating documentation. Is there any extra step that I am missing? In addition, currently I am getting the following error when I run 'make html' now:
"Exception occurred:
File "/usr/local/Caskroom/miniconda/base/envs/dev_featuretools/lib/python3.6/site-packages/IPython/sphinxext/ipython_directive.py", line 586, in process_input
raise RuntimeError('Non Expected warning in {} line {}'.format(filename, lineno))
RuntimeError: Non Expected warning in /usr/local/Caskroom/miniconda/base/envs/dev_featuretools/lib/python3.6/site-packages/featuretools/wrappers/sklearn.py:docstring of featuretools.wrappers.DFSTransformer.__init__ line 101"

[rwedge]

My mistake, you should add entropy to the Aggregation Primitives section of docs/source/api_reference.rst so it gets included in the docs

That docs error is strange, does you environment warn you when running this code:

import featuretools as ft
import pandas as pd
from sklearn.pipeline import Pipeline
from sklearn.ensemble import ExtraTreesClassifier

[twdobson]

There were some issues with the packages in the environment. Have a dedicated featuretools env now, so that should be sorted.

Apologies for the silly questions, still getting the hang of the contributing and getting everything setup. I am getting the following error when trying to make the docs:
"
Configuration error:
There is a programmable error in your configuration file:

Traceback (most recent call last):
File "/usr/local/Caskroom/miniconda/base/envs/featuretools_development/lib/python3.7/site-packages/sphinx/config.py", line 361, in eval_config_file
execfile_(filename, namespace)
File "/usr/local/Caskroom/miniconda/base/envs/featuretools_development/lib/python3.7/site-packages/sphinx/util/pycompat.py", line 86, in execfile_
exec(code, _globals)
File "/Users/timothydobson/work/repos/featuretools_dev/docs/source/conf.py", line 87, in
version = featuretools.version
AttributeError: module 'featuretools' has no attribute 'version'
"

[twdobson]

I also fail the "featuretools/tests/cli_tests/test_cli.py " test locally - however, when I install featuretools into my env this issue (and configuration error) disappear but then the documentation does not build properly. I think it's because everything is imported from the version installed in the env and not from the source code i'm working off.

[rwedge]

If you run

pip install -e /path/to/featuretools/repo

featuretools will get installed in "editable" mode and changes you make to the source code will be reflected when you import featuretools again

I would try re-installing featuretools in the env in editable mode

[twdobson]

@rwedge, thanks for the help - editable mode did the trick.

[twdobson]

Hey @kmax12, @rwedge, I think the PR is ready for review! Thanks for all the help, let me know of any other changes.

[rwedge]

@twdobson code looks good, just thinking about what range we would want to use for the scipy dependency

[twdobson]

@rwedge, thanks.

Would the oldest version of scipy that has entropy, with the required parameters, work?

Alternatively, we could implement the function via numpy, the formula is simple: Entropy = -sum(pk * log(pk), axis=0), where pk = probability of categorical, which is the output from value_counts(normalize =True)

[rwedge]

@twdobson it looks like scipy 0.11.0 is when the base parameter was added to the entropy function but our minimum scikit-learn version requires scipy 0.13.3 or higher so that may be a more practical requirement

[twdobson]

@rwedge, sounds good. I have updated the PR to reflect this.

Quick two questions:

1. how did you go about figuring the which version to use? Did you use pipdeptree?
2. Once new enhancements are made for feature generation, is there a process for testing uplift or importance across different response sets?

[twdobson]

Let me know if there is anything else that needs to be updated.

[rwedge]

PR looks good

As to your questions:

1. I installed featuretools into a clean environment and checked the pip output to see which package required scipy. I did try using pipdeptree after you mentioned it, it also came up with scipy 0.13.3 being necessary for sklearn 0.20.4
2. We don't test how new primitives impact model performance or feature importance