Data science for lazy people, Automated Machine Learning

Diego Hueltes

Big Data Spain 2017

Original image from https://github.com/rhiever/tpot

Data cleaning

Data cleaning

Data cleaning

Data cleaning

Feature selection

Feature selection

Feature preprocessing

Feature construction

Model selection

Model selection

Parameter optimization

RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
            max_depth=2, max_features='auto', max_leaf_nodes=None,
            min_impurity_decrease=0.0, min_impurity_split=None,
            min_samples_leaf=1, min_samples_split=2,
            min_weight_fraction_leaf=0.0, n_estimators=10, n_jobs=1,
            oob_score=False, random_state=0, verbose=0, warm_start=False)

Model validation

lazy Oxford dictionary

Unwilling to work or use energy

lazy Oxford dictionary

Unwilling to work or use energy

in repetitive tasks

Diego dictionary

naive_bayes.GaussianNB

naive_bayes.BernoulliNB

naive_bayes.MultinomialNB​

tree.DecisionTreeClassifier

ensemble.ExtraTreesClassifier

ensemble.RandomForestClassifier

ensemble.GradientBoostingClassifier

neighbors.KNeighborsClassifier

svm.LinearSVC

linear_model.LogisticRegression

xgboost.XGBClassifier

preprocessing.Binarizer

decomposition.FastICA

cluster.FeatureAgglomeration

cluster.FeatureAgglomeration

preprocessing.MaxAbsScaler

preprocessing.MinMaxScaler

preprocessing.Normalizer

kernel_approximation.Nystroem

decomposition.PCA

preprocessing.PolynomialFeatures

kernel_approximation.RBFSampler

preprocessing.RobustScaler

preprocessing.StandardScaler

tpot.builtins.ZeroCount

feature_selection.SelectFwe

feature_selection.SelectPercentile

feature_selection.VarianceThreshold

feature_selection.RFE

feature_selection.SelectFromModel

linear_model.ElasticNetCV

linear_model.ElasticNetCV

ensemble.ExtraTreesRegressor

ensemble.GradientBoostingRegressor

ensemble.AdaBoostRegressor

tree.DecisionTreeRegressor

neighbors.KNeighborsRegressor

linear_model.LassoLarsCV

svm.LinearSVR

ensemble.RandomForestRegressor

linear_model.RidgeCV

xgboost.XGBRegressor

Automated Machine Learning

TPOT is a Python tool that automatically creates and optimizes machine learning pipelines using genetic programming.

auto-sklearn frees a machine learning user from algorithm selection and hyperparameter tuning. It leverages recent advantages in Bayesian optimization, meta-learning and ensemble construction

auto-sklearn

Genetic programming

Source: http://www.genetic-programming.org/gpbook4toc.html​

Photo: Diego Hueltes

Crossover

Mutation

Source http://w3.onera.fr/smac/?q=tracker

Bayesian optimization

Bayesian optimization

source: https://advancedoptimizationatharvard.wordpress.com/2014/04/28/bayesian-optimization-part-ii/

Bayesian optimization

source: https://advancedoptimizationatharvard.wordpress.com/2014/04/28/bayesian-optimization-part-ii/

TPOT is a Python tool that automatically creates and optimizes machine learning pipelines using genetic programming.

Scikit-learn pipelines

Warning

Code is coming

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.svm import LinearSVC

count_vectorizer = CountVectorizer(ngram_range=(1, 4), analyzer='char')
X_train = count_vectorizer.fit_transform(train)
X_test  = count_vectorizer.transform(test)

linear_svc = LinearSVC()
model = linear_svc.fit(X_train, y_train)

y_test = model.predict(X_test)
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.svm import LinearSVC

pipeline = Pipeline([
    ('count_vectorizer', CountVectorizer(ngram_range=(1, 4), analyzer='char')),
    ('linear_svc', LinearSVC())
])
model = pipeline.fit(train)
y_test = model.predict(test)

TPOT

from tpot import TPOTClassifier, TPOTRegressor

tpot = TPOTClassifier()
tpot.fit(X_train, y_train)
predictions = tpot.predict(X_test)

tpot = TPOTRegressor()
tpot.fit(X_train, y_train)
predictions = tpot.predict(X_test)

Basic usage

Config dict

TPOTClassifier(config_dict = {
    'sklearn.ensemble.RandomForestClassifier': {
        'n_estimators': [100],
        'criterion': ["gini", "entropy"],
        'max_features': np.arange(0.05, 1.01, 0.05),
        'min_samples_split': range(2, 21),
        'min_samples_leaf':  range(1, 21),
        'bootstrap': [True, False]
    },
    'sklearn.feature_selection.RFE': {
        'step': np.arange(0.05, 1.01, 0.05),
        'estimator': {
            'sklearn.ensemble.ExtraTreesClassifier': {
                'n_estimators': [100],
                'criterion': ['gini', 'entropy'],
                'max_features': np.arange(0.05, 1.01, 0.05)
                }
        }
    }
})

auto-sklearn frees a machine learning user from algorithm selection and hyperparameter tuning. It leverages recent advantages in Bayesian optimization, meta-learning and ensemble construction

auto-sklearn

auto-sklearn

Source: http://papers.nips.cc/paper/5872-efficient-and-robust-automated-machine-learning.pdf

import autosklearn.classification
import autosklearn.regression

automl = autosklearn.classification.AutoSklearnClassifier()
automl.fit(X_train, y_train)
predictions = automl.predict(X_test)


automl = autosklearn.regression.AutoSklearnRegressor()
automl.fit(X_train, y_train)
predictions = automl.predict(X_test)

Basic usage

auto-sklearn

Custom configuration

auto-sklearn

  • include_estimators
  • exclude_estimators
  • include_preprocessors
  • ex​clude_preprocessors

Automated Machine Learning — A Paradigm Shift That Accelerates Data Scientist Productivity @ Airbnb

Automated Machine Learning

  • Exploratory analysis
  • Selective discovering
  • New ideas for your model
  • Model optimization

Thank you!

Data science for lazy people, Automated Machine Learning

By J. Diego Hueltes Vega

Data science for lazy people, Automated Machine Learning

Data science is fun… right? Data cleaning, feature selection, feature preprocessing, feature construction, model selection, parameter optimization, model validation… oh wait… are you sure? What about automating 80% of the work even doing better choices than you? Automated Machine Learning has arrived to be your personal assistant in Data Science

  • 880
Loading comments...

More from J. Diego Hueltes Vega