Feature Request: KitchenSink classification (Fits every classification algorithm in a single line)

Description

Example: Machine Learning practioners want to try every single classification algorithm for a dataset, how about having a modularized module which is an additional feature of GridSearchCV that now takes MULTIPLE classification algorithms with multiple parameters for each classification algorithm and shoots out the model that best classifies the data based on a chosen scorer

Steps/Code to Reproduce

Example:

hyperparametertuning.py

(C) 2017 by Abhishek Babuji <[e-mail]>

Contains methods to return a pipeline object and a dictionary containing
classifier parameters
"""

from sklearn.ensemble import GradientBoostingClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.naive_bayes import MultinomialNB
from sklearn.neighbors import KNeighborsClassifier
from sklearn.pipeline import Pipeline
from sklearn.svm import SVC


class HyperParameterTuning:
    """
    Contains methods to return a pipeline object and a dictionary containing
    classifier parameters
    """

    def __init__(self, classifier, vectorizer):
        """
        Args:
            classifier (One of 6 sklearn classifier objects): 'logreg', 'svm', 'nb',
                                                              'knn', 'xgboost', 'randomforests'
            vectorizer (CountVectorizer or TfidfVectorizer): Type of vector space model


        Returns:
            pipeline (sklearn pipeline object): Returns a pipeline object which is used
                                                by GridSearchCV
            model_params[self.classifier] (dict): Returns a dictionary of parameters
                                                  for the specified type of classifier

        """

        self.classifier = classifier
        self.vectorizer = vectorizer

    def get_pipeline(self):
        """
        Args:

            classifier (One of 6 sklearn classifier objects): 'logreg', 'svm', 'nb',
                                                              'knn', 'xgboost', 'randomforests'
            vectorizer (CountVectorizer or TfidfVectorizer): Type of vector space model


        Returns:
            pipeline (sklearn pipeline object): Returns a pipeline object which is
                                                used by GridSearchCV
            model_params[self.classifier] (dict): Returns a dictionary of parameters
                                                  for the specified type of classifier
        """

        classifier_objects = {'logreg': LogisticRegression(),
                              'svm': SVC(),
                              'knn': KNeighborsClassifier(),
                              'xgboost': GradientBoostingClassifier(),
                              'randomforests': RandomForestClassifier(),
                              'nb': MultinomialNB()}
        pipeline = Pipeline([('vect', self.vectorizer),
                             ('clf', classifier_objects[self.classifier])])

        return pipeline

    def get_params(self):
        """
        Args:
            self


        Returns:
            model_params[self.classifier] (dict): Returns a dictionary of parameters for the
                                                  specified type of classifier

        """
        model_params = {'logreg': {'clf__C': (1, 10, 100), 'clf__penalty': ('l1', 'l2')},
                        'svm': {'clf__C': (1, 10, 100),
                                'clf__kernel': ('linear', 'poly', 'rbf', 'sigmoid')},
                        'knn': {'clf__n_neighbors': (5, 10, 50, 100)},
                        'xgboost': {'clf__n_estimators': (100, 500, 1000)},
                        'randomforests': {'clf__n_estimators': (100, 500, 1000)},
                        'nb': {'clf__alpha': (0, 1), 'clf__fit_prior': (True, False)}}
        return model_params[self.classifier]

An example like this, that tries EVERY classifier. Of course this is very computationally and time expensive… But if you’re trying to fit it for small datasets where Machine Learning is applicable, we could brainstorm further to maybe understand which default HyperParameters to tune for each algorithm using a scorer of your choice

[amueller]

Hey. There’s auto-sklearn that does that (somewhat smarter).

Some of their recent work is described here: https://ml.informatik.uni-freiburg.de/papers/18-AUTOML-AutoChalleng

I’ll probably soon post a simple implementation of that, but outside of sklearn. I think the space is still too volatile to do this in sklearn.

[AbhishekBabuji]

@amueller Absolutely. Feel free to close it. I am going to make it my serious hobby to contribute to sklearn. I write potato code though. So I have some ways to go!