Allow for Flexible Preprocessing

[SourdoughCat]

[please review the Contribution Guidelines prior to submitting your pull request. go ahead and delete this line if you've already reviewed said guidelines.]

What does this PR do?

Too many things - so many things that I'm pretty sure there will be discussions on whether this is the way to proceed. Putting it up here before I invest more time on it. This PR addresses several items (possibly too many again) including

#507
#836
Handling of categorical data #771 - I know its not directly related; but it is the closest issue, and in my mind it would allow for extending it to using different encodings like using this: https://github.com/scikit-learn-contrib/categorical-encoding

Which are all related to how TPOT does preprocessing. This PR does a number of things including re-introducing "RandomTree" option in templates that allows specifying templates in the form Transformer-RandomTree; which will then allow things like My_Preprocessing-RandomTree.

The high level approach is to inject additional preprocessing steps when _fit_init is called which then alters the behaviour of TPOT.

Where should the reviewer start?

tpot/base.py - will add comments in files as part of this PR with my thoughts...

Also can't get relative imports working - so that would be appreciated for the tpot.drivers.load_scoring_function part

How should this PR be tested?

Happy to add new tests later; when the design is approved...

Any background context you want to provide?

NIL see above

What are the relevant issues?

#507
#836
Handling of categorical data #771

Screenshots (if appropriate)

Example of API using modified Iris dataset

from tpot import TPOTClassifier
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
import numpy as np
import pandas as pd

iris = load_iris()
X_train, X_test, y_train, y_test = train_test_split(iris.data.astype(np.float64),
    iris.target.astype(np.float64), train_size=0.75, test_size=0.25)

tpot = TPOTClassifier(generations=1, population_size=5, verbosity=2, template="PCA-RandomTree")

X_train_df = pd.DataFrame(X_train, columns=["num1", "num2", "num3", "num4"])
X_train_df['text'] = np.random.choice(["hello", "world", "foo", "bar world", "bar hello"], X_train.shape[0])

tpot2 = TPOTClassifier(generations=1, population_size=5, verbosity=2, template="PCA-LogisticRegression", 
                      preprocess_config_dict = {
                          'numeric_columns': ["num2"]
                      })
tpot2.fit(X_train_df, y_train)

tpot2 = TPOTClassifier(generations=1, population_size=5, verbosity=2, template="PCA-LogisticRegression", 
                      preprocess_config_dict = {
                          'numeric_columns': ["num2", "num3", "num4"]
                      })
tpot2.fit(X_train_df, y_train)

tpot2 = TPOTClassifier(generations=1, population_size=5, verbosity=2, template="PCA-LogisticRegression", 
                      preprocess_config_dict = {
                          'numeric_columns': ["num2"],
                          'text_columns': ['text']
                      })
tpot2.fit(X_train_df, y_train)

Generation 1 - Current best internal CV score: 0.5908385093167702

Best pipeline: LogisticRegression(PCA(PreprocessTransformer(input_matrix, numeric_columns=['num2']), iterated_power=2, svd_solver=randomized), C=20.0, dual=True, penalty=l2)

Generation 1 - Current best internal CV score: 0.9556935817805383

Best pipeline: LogisticRegression(PCA(PreprocessTransformer(input_matrix, numeric_columns=['num2', 'num3', 'num4']), iterated_power=9, svd_solver=randomized), C=20.0, dual=False, penalty=l1)

Generation 1 - Current best internal CV score: 0.5096273291925466

Questions:

• Do the docs need to be updated? Yes - will do later
• Does this PR add new (Python) dependencies? No

[SourdoughCat]

help for getting relative imports working would be appreciated here... tpot.drivers.load_scoring_function

[weixuanfu]

Hmm, directly importing via from ..driver import load_scoring_function will cause some conflicts with

from .tpot import TPOTClassifier, TPOTRegressor
from ._version import __version__

A workaround is to move this function to tpot/metrics.py and then add from ..metrics import load_scoring_function to tpot/builtins/preprocessing.py

[SourdoughCat]

This injection could be dangerous - do we have opinions on how it is supposed to be handled?

[weixuanfu]

I think preprocess_config_dict should be a argument within PreprocessTransformer instead of TPOT. And users should be able to customize it via config_dict.

[SourdoughCat]

Yes this is certainly possible (and possible right now with no changes to TPOT master technically) via the use of templates - I think the question arises related to #507; if its possible to have a "built-in" configuration with text or not.

Maybe the answer is we can't

[SourdoughCat]

This line deals with #836 - might be overloading this PR and might be an item for later?

[weixuanfu]

I think it is more related to #889. I think we need add imputation into config_dict too. We may allow TPOT skip imputation if the pipeline only has XGBClassifier or XGBRegressor.

[SourdoughCat]

Ok,

Given the higher level comments @weixuanfu how should we proceed. In my opinion, I think this PR is just too big, what we probably want is:

• a PR that just adds back "RandomTree" as an option so that we can do things like "PCA-RandomTree" or similar (have to ensure than RandomTree is the very last option"; in doing so this enables
• Another PR that proposes how the preprocessing transform should work - there are downsides to this, as it means someone using the preprocessing has to directly alter config options for every model. e.g.

(as per comments above)

X_train, y_train = ...
my_meta_data_info = {<insert meta data information>}
config_dict = {**my_meta_data_info, **TPOT_DEFAULT_CONFIG}
tpot = TPOTClassifier(config_dict=config_dict, template='Preprocess-RandomTree')
tpot.fit(X_train, y_train)

versus

(as per approach taken in this PR)

X_train, y_train = ...
tpot = TPOTClassifier(config_dict=None, preprocess_config_dict ={<insert meta data>})
tpot.fit(X_train, y_train)