"""Weight Boosting This module contains weight boosting estimators for both classification and regression. The module structure is the following: - The ``BaseWeightBoosting`` base class implements a common ``fit`` method for all the estimators in the module. Regression and classification only differ from each other in the loss function that is optimized. - ``AdaBoostClassifier`` implements adaptive boosting (AdaBoost-SAMME) for classification problems. - ``AdaBoostRegressor`` implements adaptive boosting (AdaBoost.R2) for regression problems. """ # Authors: Noel Dawe # Gilles Louppe # Hamzeh Alsalhi # Arnaud Joly # # License: BSD 3 clause from abc import ABCMeta, abstractmethod import numpy as np from .base import BaseEnsemble from ..base import ClassifierMixin, RegressorMixin, is_regressor, is_classifier from ..externals import six from ..externals.six.moves import zip from ..externals.six.moves import xrange as range from .forest import BaseForest from ..tree import DecisionTreeClassifier, DecisionTreeRegressor from ..tree.tree import BaseDecisionTree from ..tree._tree import DTYPE from ..utils import check_array, check_X_y, check_random_state from ..utils.extmath import stable_cumsum from ..metrics import accuracy_score, r2_score from sklearn.utils.validation import has_fit_parameter, check_is_fitted __all__ = [ 'AdaBoostClassifier', 'AdaBoostRegressor', ] class BaseWeightBoosting(six.with_metaclass(ABCMeta, BaseEnsemble)): """Base class for AdaBoost estimators. Warning: This class should not be used directly. Use derived classes instead. """ @abstractmethod def __init__(self, base_estimator=None, n_estimators=50, estimator_params=tuple(), learning_rate=1., random_state=None): super(BaseWeightBoosting, self).__init__( base_estimator=base_estimator, n_estimators=n_estimators, estimator_params=estimator_params) self.learning_rate = learning_rate self.random_state = random_state def fit(self, X, y, sample_weight=None): """Build a boosted classifier/regressor from the training set (X, y). Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. COO, DOK, and LIL are converted to CSR. The dtype is forced to DTYPE from tree._tree if the base classifier of this ensemble weighted boosting classifier is a tree or forest. y : array-like of shape = [n_samples] The target values (class labels in classification, real numbers in regression). sample_weight : array-like of shape = [n_samples], optional Sample weights. If None, the sample weights are initialized to 1 / n_samples. Returns ------- self : object """ # Check parameters if self.learning_rate <= 0: raise ValueError("learning_rate must be greater than zero") if (self.base_estimator is None or isinstance(self.base_estimator, (BaseDecisionTree, BaseForest))): dtype = DTYPE accept_sparse = 'csc' else: dtype = None accept_sparse = ['csr', 'csc'] X, y = check_X_y(X, y, accept_sparse=accept_sparse, dtype=dtype, y_numeric=is_regressor(self)) if sample_weight is None: # Initialize weights to 1 / n_samples sample_weight = np.empty(X.shape[0], dtype=np.float64) sample_weight[:] = 1. / X.shape[0] else: sample_weight = check_array(sample_weight, ensure_2d=False) # Normalize existing weights sample_weight = sample_weight / sample_weight.sum(dtype=np.float64) # Check that the sample weights sum is positive if sample_weight.sum() <= 0: raise ValueError( "Attempting to fit with a non-positive " "weighted number of samples.") # Check parameters self._validate_estimator() # Clear any previous fit results self.estimators_ = [] self.estimator_weights_ = np.zeros(self.n_estimators, dtype=np.float64) self.estimator_errors_ = np.ones(self.n_estimators, dtype=np.float64) random_state = check_random_state(self.random_state) for iboost in range(self.n_estimators): # Boosting step sample_weight, estimator_weight, estimator_error = self._boost( iboost, X, y, sample_weight, random_state) # Early termination if sample_weight is None: break self.estimator_weights_[iboost] = estimator_weight self.estimator_errors_[iboost] = estimator_error # Stop if error is zero if estimator_error == 0: break sample_weight_sum = np.sum(sample_weight) # Stop if the sum of sample weights has become non-positive if sample_weight_sum <= 0: break if iboost < self.n_estimators - 1: # Normalize sample_weight /= sample_weight_sum return self @abstractmethod def _boost(self, iboost, X, y, sample_weight, random_state): """Implement a single boost. Warning: This method needs to be overridden by subclasses. Parameters ---------- iboost : int The index of the current boost iteration. X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. COO, DOK, and LIL are converted to CSR. y : array-like of shape = [n_samples] The target values (class labels). sample_weight : array-like of shape = [n_samples] The current sample weights. random_state : RandomState The current random number generator Returns ------- sample_weight : array-like of shape = [n_samples] or None The reweighted sample weights. If None then boosting has terminated early. estimator_weight : float The weight for the current boost. If None then boosting has terminated early. error : float The classification error for the current boost. If None then boosting has terminated early. """ pass def staged_score(self, X, y, sample_weight=None): """Return staged scores for X, y. This generator method yields the ensemble score after each iteration of boosting and therefore allows monitoring, such as to determine the score on a test set after each boost. Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. y : array-like, shape = [n_samples] Labels for X. sample_weight : array-like, shape = [n_samples], optional Sample weights. Returns ------- z : float """ for y_pred in self.staged_predict(X): if is_classifier(self): yield accuracy_score(y, y_pred, sample_weight=sample_weight) else: yield r2_score(y, y_pred, sample_weight=sample_weight) @property def feature_importances_(self): """Return the feature importances (the higher, the more important the feature). Returns ------- feature_importances_ : array, shape = [n_features] """ if self.estimators_ is None or len(self.estimators_) == 0: raise ValueError("Estimator not fitted, " "call `fit` before `feature_importances_`.") try: norm = self.estimator_weights_.sum() return (sum(weight * clf.feature_importances_ for weight, clf in zip(self.estimator_weights_, self.estimators_)) / norm) except AttributeError: raise AttributeError( "Unable to compute feature importances " "since base_estimator does not have a " "feature_importances_ attribute") def _validate_X_predict(self, X): """Ensure that X is in the proper format""" if (self.base_estimator is None or isinstance(self.base_estimator, (BaseDecisionTree, BaseForest))): X = check_array(X, accept_sparse='csr', dtype=DTYPE) else: X = check_array(X, accept_sparse=['csr', 'csc', 'coo']) return X def _samme_proba(estimator, n_classes, X): """Calculate algorithm 4, step 2, equation c) of Zhu et al [1]. References ---------- .. [1] J. Zhu, H. Zou, S. Rosset, T. Hastie, "Multi-class AdaBoost", 2009. """ proba = estimator.predict_proba(X) # Displace zero probabilities so the log is defined. # Also fix negative elements which may occur with # negative sample weights. np.clip(proba, np.finfo(proba.dtype).eps, None, out=proba) log_proba = np.log(proba) return (n_classes - 1) * (log_proba - (1. / n_classes) * log_proba.sum(axis=1)[:, np.newaxis]) class AdaBoostClassifier(BaseWeightBoosting, ClassifierMixin): """An AdaBoost classifier. An AdaBoost [1] classifier is a meta-estimator that begins by fitting a classifier on the original dataset and then fits additional copies of the classifier on the same dataset but where the weights of incorrectly classified instances are adjusted such that subsequent classifiers focus more on difficult cases. This class implements the algorithm known as AdaBoost-SAMME [2]. Read more in the :ref:`User Guide `. Parameters ---------- base_estimator : object, optional (default=None) The base estimator from which the boosted ensemble is built. Support for sample weighting is required, as well as proper ``classes_`` and ``n_classes_`` attributes. If ``None``, then the base estimator is ``DecisionTreeClassifier(max_depth=1)`` n_estimators : integer, optional (default=50) The maximum number of estimators at which boosting is terminated. In case of perfect fit, the learning procedure is stopped early. learning_rate : float, optional (default=1.) Learning rate shrinks the contribution of each classifier by ``learning_rate``. There is a trade-off between ``learning_rate`` and ``n_estimators``. algorithm : {'SAMME', 'SAMME.R'}, optional (default='SAMME.R') If 'SAMME.R' then use the SAMME.R real boosting algorithm. ``base_estimator`` must support calculation of class probabilities. If 'SAMME' then use the SAMME discrete boosting algorithm. The SAMME.R algorithm typically converges faster than SAMME, achieving a lower test error with fewer boosting iterations. random_state : int, RandomState instance or None, optional (default=None) If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; If None, the random number generator is the RandomState instance used by `np.random`. Attributes ---------- estimators_ : list of classifiers The collection of fitted sub-estimators. classes_ : array of shape = [n_classes] The classes labels. n_classes_ : int The number of classes. estimator_weights_ : array of floats Weights for each estimator in the boosted ensemble. estimator_errors_ : array of floats Classification error for each estimator in the boosted ensemble. feature_importances_ : array of shape = [n_features] The feature importances if supported by the ``base_estimator``. See also -------- AdaBoostRegressor, GradientBoostingClassifier, sklearn.tree.DecisionTreeClassifier References ---------- .. [1] Y. Freund, R. Schapire, "A Decision-Theoretic Generalization of on-Line Learning and an Application to Boosting", 1995. .. [2] J. Zhu, H. Zou, S. Rosset, T. Hastie, "Multi-class AdaBoost", 2009. """ def __init__(self, base_estimator=None, n_estimators=50, learning_rate=1., algorithm='SAMME.R', random_state=None): super(AdaBoostClassifier, self).__init__( base_estimator=base_estimator, n_estimators=n_estimators, learning_rate=learning_rate, random_state=random_state) self.algorithm = algorithm def fit(self, X, y, sample_weight=None): """Build a boosted classifier from the training set (X, y). Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. y : array-like of shape = [n_samples] The target values (class labels). sample_weight : array-like of shape = [n_samples], optional Sample weights. If None, the sample weights are initialized to ``1 / n_samples``. Returns ------- self : object """ # Check that algorithm is supported if self.algorithm not in ('SAMME', 'SAMME.R'): raise ValueError("algorithm %s is not supported" % self.algorithm) # Fit return super(AdaBoostClassifier, self).fit(X, y, sample_weight) def _validate_estimator(self): """Check the estimator and set the base_estimator_ attribute.""" super(AdaBoostClassifier, self)._validate_estimator( default=DecisionTreeClassifier(max_depth=1)) # SAMME-R requires predict_proba-enabled base estimators if self.algorithm == 'SAMME.R': if not hasattr(self.base_estimator_, 'predict_proba'): raise TypeError( "AdaBoostClassifier with algorithm='SAMME.R' requires " "that the weak learner supports the calculation of class " "probabilities with a predict_proba method.\n" "Please change the base estimator or set " "algorithm='SAMME' instead.") if not has_fit_parameter(self.base_estimator_, "sample_weight"): raise ValueError("%s doesn't support sample_weight." % self.base_estimator_.__class__.__name__) def _boost(self, iboost, X, y, sample_weight, random_state): """Implement a single boost. Perform a single boost according to the real multi-class SAMME.R algorithm or to the discrete SAMME algorithm and return the updated sample weights. Parameters ---------- iboost : int The index of the current boost iteration. X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. y : array-like of shape = [n_samples] The target values (class labels). sample_weight : array-like of shape = [n_samples] The current sample weights. random_state : RandomState The current random number generator Returns ------- sample_weight : array-like of shape = [n_samples] or None The reweighted sample weights. If None then boosting has terminated early. estimator_weight : float The weight for the current boost. If None then boosting has terminated early. estimator_error : float The classification error for the current boost. If None then boosting has terminated early. """ if self.algorithm == 'SAMME.R': return self._boost_real(iboost, X, y, sample_weight, random_state) else: # elif self.algorithm == "SAMME": return self._boost_discrete(iboost, X, y, sample_weight, random_state) def _boost_real(self, iboost, X, y, sample_weight, random_state): """Implement a single boost using the SAMME.R real algorithm.""" estimator = self._make_estimator(random_state=random_state) estimator.fit(X, y, sample_weight=sample_weight) y_predict_proba = estimator.predict_proba(X) if iboost == 0: self.classes_ = getattr(estimator, 'classes_', None) self.n_classes_ = len(self.classes_) y_predict = self.classes_.take(np.argmax(y_predict_proba, axis=1), axis=0) # Instances incorrectly classified incorrect = y_predict != y # Error fraction estimator_error = np.mean( np.average(incorrect, weights=sample_weight, axis=0)) # Stop if classification is perfect if estimator_error <= 0: return sample_weight, 1., 0. # Construct y coding as described in Zhu et al [2]: # # y_k = 1 if c == k else -1 / (K - 1) # # where K == n_classes_ and c, k in [0, K) are indices along the second # axis of the y coding with c being the index corresponding to the true # class label. n_classes = self.n_classes_ classes = self.classes_ y_codes = np.array([-1. / (n_classes - 1), 1.]) y_coding = y_codes.take(classes == y[:, np.newaxis]) # Displace zero probabilities so the log is defined. # Also fix negative elements which may occur with # negative sample weights. proba = y_predict_proba # alias for readability np.clip(proba, np.finfo(proba.dtype).eps, None, out=proba) # Boost weight using multi-class AdaBoost SAMME.R alg estimator_weight = (-1. * self.learning_rate * ((n_classes - 1.) / n_classes) * (y_coding * np.log(y_predict_proba)).sum(axis=1)) # Only boost the weights if it will fit again if not iboost == self.n_estimators - 1: # Only boost positive weights sample_weight *= np.exp(estimator_weight * ((sample_weight > 0) | (estimator_weight < 0))) return sample_weight, 1., estimator_error def _boost_discrete(self, iboost, X, y, sample_weight, random_state): """Implement a single boost using the SAMME discrete algorithm.""" estimator = self._make_estimator(random_state=random_state) estimator.fit(X, y, sample_weight=sample_weight) y_predict = estimator.predict(X) if iboost == 0: self.classes_ = getattr(estimator, 'classes_', None) self.n_classes_ = len(self.classes_) # Instances incorrectly classified incorrect = y_predict != y # Error fraction estimator_error = np.mean( np.average(incorrect, weights=sample_weight, axis=0)) # Stop if classification is perfect if estimator_error <= 0: return sample_weight, 1., 0. n_classes = self.n_classes_ # Stop if the error is at least as bad as random guessing if estimator_error >= 1. - (1. / n_classes): self.estimators_.pop(-1) if len(self.estimators_) == 0: raise ValueError('BaseClassifier in AdaBoostClassifier ' 'ensemble is worse than random, ensemble ' 'can not be fit.') return None, None, None # Boost weight using multi-class AdaBoost SAMME alg estimator_weight = self.learning_rate * ( np.log((1. - estimator_error) / estimator_error) + np.log(n_classes - 1.)) # Only boost the weights if I will fit again if not iboost == self.n_estimators - 1: # Only boost positive weights sample_weight *= np.exp(estimator_weight * incorrect * ((sample_weight > 0) | (estimator_weight < 0))) return sample_weight, estimator_weight, estimator_error def predict(self, X): """Predict classes for X. The predicted class of an input sample is computed as the weighted mean prediction of the classifiers in the ensemble. Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. Returns ------- y : array of shape = [n_samples] The predicted classes. """ pred = self.decision_function(X) if self.n_classes_ == 2: return self.classes_.take(pred > 0, axis=0) return self.classes_.take(np.argmax(pred, axis=1), axis=0) def staged_predict(self, X): """Return staged predictions for X. The predicted class of an input sample is computed as the weighted mean prediction of the classifiers in the ensemble. This generator method yields the ensemble prediction after each iteration of boosting and therefore allows monitoring, such as to determine the prediction on a test set after each boost. Parameters ---------- X : array-like of shape = [n_samples, n_features] The input samples. Returns ------- y : generator of array, shape = [n_samples] The predicted classes. """ n_classes = self.n_classes_ classes = self.classes_ if n_classes == 2: for pred in self.staged_decision_function(X): yield np.array(classes.take(pred > 0, axis=0)) else: for pred in self.staged_decision_function(X): yield np.array(classes.take( np.argmax(pred, axis=1), axis=0)) def decision_function(self, X): """Compute the decision function of ``X``. Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. Returns ------- score : array, shape = [n_samples, k] The decision function of the input samples. The order of outputs is the same of that of the `classes_` attribute. Binary classification is a special cases with ``k == 1``, otherwise ``k==n_classes``. For binary classification, values closer to -1 or 1 mean more like the first or second class in ``classes_``, respectively. """ check_is_fitted(self, "n_classes_") X = self._validate_X_predict(X) n_classes = self.n_classes_ classes = self.classes_[:, np.newaxis] pred = None if self.algorithm == 'SAMME.R': # The weights are all 1. for SAMME.R pred = sum(_samme_proba(estimator, n_classes, X) for estimator in self.estimators_) else: # self.algorithm == "SAMME" pred = sum((estimator.predict(X) == classes).T * w for estimator, w in zip(self.estimators_, self.estimator_weights_)) pred /= self.estimator_weights_.sum() if n_classes == 2: pred[:, 0] *= -1 return pred.sum(axis=1) return pred def staged_decision_function(self, X): """Compute decision function of ``X`` for each boosting iteration. This method allows monitoring (i.e. determine error on testing set) after each boosting iteration. Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. Returns ------- score : generator of array, shape = [n_samples, k] The decision function of the input samples. The order of outputs is the same of that of the `classes_` attribute. Binary classification is a special cases with ``k == 1``, otherwise ``k==n_classes``. For binary classification, values closer to -1 or 1 mean more like the first or second class in ``classes_``, respectively. """ check_is_fitted(self, "n_classes_") X = self._validate_X_predict(X) n_classes = self.n_classes_ classes = self.classes_[:, np.newaxis] pred = None norm = 0. for weight, estimator in zip(self.estimator_weights_, self.estimators_): norm += weight if self.algorithm == 'SAMME.R': # The weights are all 1. for SAMME.R current_pred = _samme_proba(estimator, n_classes, X) else: # elif self.algorithm == "SAMME": current_pred = estimator.predict(X) current_pred = (current_pred == classes).T * weight if pred is None: pred = current_pred else: pred += current_pred if n_classes == 2: tmp_pred = np.copy(pred) tmp_pred[:, 0] *= -1 yield (tmp_pred / norm).sum(axis=1) else: yield pred / norm def predict_proba(self, X): """Predict class probabilities for X. The predicted class probabilities of an input sample is computed as the weighted mean predicted class probabilities of the classifiers in the ensemble. Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. Returns ------- p : array of shape = [n_samples, n_classes] The class probabilities of the input samples. The order of outputs is the same of that of the `classes_` attribute. """ check_is_fitted(self, "n_classes_") n_classes = self.n_classes_ X = self._validate_X_predict(X) if n_classes == 1: return np.ones((X.shape[0], 1)) if self.algorithm == 'SAMME.R': # The weights are all 1. for SAMME.R proba = sum(_samme_proba(estimator, n_classes, X) for estimator in self.estimators_) else: # self.algorithm == "SAMME" proba = sum(estimator.predict_proba(X) * w for estimator, w in zip(self.estimators_, self.estimator_weights_)) proba /= self.estimator_weights_.sum() proba = np.exp((1. / (n_classes - 1)) * proba) normalizer = proba.sum(axis=1)[:, np.newaxis] normalizer[normalizer == 0.0] = 1.0 proba /= normalizer return proba def staged_predict_proba(self, X): """Predict class probabilities for X. The predicted class probabilities of an input sample is computed as the weighted mean predicted class probabilities of the classifiers in the ensemble. This generator method yields the ensemble predicted class probabilities after each iteration of boosting and therefore allows monitoring, such as to determine the predicted class probabilities on a test set after each boost. Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. Returns ------- p : generator of array, shape = [n_samples] The class probabilities of the input samples. The order of outputs is the same of that of the `classes_` attribute. """ X = self._validate_X_predict(X) n_classes = self.n_classes_ proba = None norm = 0. for weight, estimator in zip(self.estimator_weights_, self.estimators_): norm += weight if self.algorithm == 'SAMME.R': # The weights are all 1. for SAMME.R current_proba = _samme_proba(estimator, n_classes, X) else: # elif self.algorithm == "SAMME": current_proba = estimator.predict_proba(X) * weight if proba is None: proba = current_proba else: proba += current_proba real_proba = np.exp((1. / (n_classes - 1)) * (proba / norm)) normalizer = real_proba.sum(axis=1)[:, np.newaxis] normalizer[normalizer == 0.0] = 1.0 real_proba /= normalizer yield real_proba def predict_log_proba(self, X): """Predict class log-probabilities for X. The predicted class log-probabilities of an input sample is computed as the weighted mean predicted class log-probabilities of the classifiers in the ensemble. Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. Returns ------- p : array of shape = [n_samples, n_classes] The class probabilities of the input samples. The order of outputs is the same of that of the `classes_` attribute. """ return np.log(self.predict_proba(X)) class AdaBoostRegressor(BaseWeightBoosting, RegressorMixin): """An AdaBoost regressor. An AdaBoost [1] regressor is a meta-estimator that begins by fitting a regressor on the original dataset and then fits additional copies of the regressor on the same dataset but where the weights of instances are adjusted according to the error of the current prediction. As such, subsequent regressors focus more on difficult cases. This class implements the algorithm known as AdaBoost.R2 [2]. Read more in the :ref:`User Guide `. Parameters ---------- base_estimator : object, optional (default=None) The base estimator from which the boosted ensemble is built. Support for sample weighting is required. If ``None``, then the base estimator is ``DecisionTreeRegressor(max_depth=3)`` n_estimators : integer, optional (default=50) The maximum number of estimators at which boosting is terminated. In case of perfect fit, the learning procedure is stopped early. learning_rate : float, optional (default=1.) Learning rate shrinks the contribution of each regressor by ``learning_rate``. There is a trade-off between ``learning_rate`` and ``n_estimators``. loss : {'linear', 'square', 'exponential'}, optional (default='linear') The loss function to use when updating the weights after each boosting iteration. random_state : int, RandomState instance or None, optional (default=None) If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; If None, the random number generator is the RandomState instance used by `np.random`. Attributes ---------- estimators_ : list of classifiers The collection of fitted sub-estimators. estimator_weights_ : array of floats Weights for each estimator in the boosted ensemble. estimator_errors_ : array of floats Regression error for each estimator in the boosted ensemble. feature_importances_ : array of shape = [n_features] The feature importances if supported by the ``base_estimator``. See also -------- AdaBoostClassifier, GradientBoostingRegressor, sklearn.tree.DecisionTreeRegressor References ---------- .. [1] Y. Freund, R. Schapire, "A Decision-Theoretic Generalization of on-Line Learning and an Application to Boosting", 1995. .. [2] H. Drucker, "Improving Regressors using Boosting Techniques", 1997. """ def __init__(self, base_estimator=None, n_estimators=50, learning_rate=1., loss='linear', random_state=None): super(AdaBoostRegressor, self).__init__( base_estimator=base_estimator, n_estimators=n_estimators, learning_rate=learning_rate, random_state=random_state) self.loss = loss self.random_state = random_state def fit(self, X, y, sample_weight=None): """Build a boosted regressor from the training set (X, y). Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. y : array-like of shape = [n_samples] The target values (real numbers). sample_weight : array-like of shape = [n_samples], optional Sample weights. If None, the sample weights are initialized to 1 / n_samples. Returns ------- self : object """ # Check loss if self.loss not in ('linear', 'square', 'exponential'): raise ValueError( "loss must be 'linear', 'square', or 'exponential'") # Fit return super(AdaBoostRegressor, self).fit(X, y, sample_weight) def _validate_estimator(self): """Check the estimator and set the base_estimator_ attribute.""" super(AdaBoostRegressor, self)._validate_estimator( default=DecisionTreeRegressor(max_depth=3)) def _boost(self, iboost, X, y, sample_weight, random_state): """Implement a single boost for regression Perform a single boost according to the AdaBoost.R2 algorithm and return the updated sample weights. Parameters ---------- iboost : int The index of the current boost iteration. X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. y : array-like of shape = [n_samples] The target values (class labels in classification, real numbers in regression). sample_weight : array-like of shape = [n_samples] The current sample weights. random_state : RandomState The current random number generator Returns ------- sample_weight : array-like of shape = [n_samples] or None The reweighted sample weights. If None then boosting has terminated early. estimator_weight : float The weight for the current boost. If None then boosting has terminated early. estimator_error : float The regression error for the current boost. If None then boosting has terminated early. """ estimator = self._make_estimator(random_state=random_state) # Weighted sampling of the training set with replacement # For NumPy >= 1.7.0 use np.random.choice cdf = stable_cumsum(sample_weight) cdf /= cdf[-1] uniform_samples = random_state.random_sample(X.shape[0]) bootstrap_idx = cdf.searchsorted(uniform_samples, side='right') # searchsorted returns a scalar bootstrap_idx = np.array(bootstrap_idx, copy=False) # Fit on the bootstrapped sample and obtain a prediction # for all samples in the training set estimator.fit(X[bootstrap_idx], y[bootstrap_idx]) y_predict = estimator.predict(X) error_vect = np.abs(y_predict - y) error_max = error_vect.max() if error_max != 0.: error_vect /= error_max if self.loss == 'square': error_vect **= 2 elif self.loss == 'exponential': error_vect = 1. - np.exp(- error_vect) # Calculate the average loss estimator_error = (sample_weight * error_vect).sum() if estimator_error <= 0: # Stop if fit is perfect return sample_weight, 1., 0. elif estimator_error >= 0.5: # Discard current estimator only if it isn't the only one if len(self.estimators_) > 1: self.estimators_.pop(-1) return None, None, None beta = estimator_error / (1. - estimator_error) # Boost weight using AdaBoost.R2 alg estimator_weight = self.learning_rate * np.log(1. / beta) if not iboost == self.n_estimators - 1: sample_weight *= np.power( beta, (1. - error_vect) * self.learning_rate) return sample_weight, estimator_weight, estimator_error def _get_median_predict(self, X, limit): # Evaluate predictions of all estimators predictions = np.array([ est.predict(X) for est in self.estimators_[:limit]]).T # Sort the predictions sorted_idx = np.argsort(predictions, axis=1) # Find index of median prediction for each sample weight_cdf = stable_cumsum(self.estimator_weights_[sorted_idx], axis=1) median_or_above = weight_cdf >= 0.5 * weight_cdf[:, -1][:, np.newaxis] median_idx = median_or_above.argmax(axis=1) median_estimators = sorted_idx[np.arange(X.shape[0]), median_idx] # Return median predictions return predictions[np.arange(X.shape[0]), median_estimators] def predict(self, X): """Predict regression value for X. The predicted regression value of an input sample is computed as the weighted median prediction of the classifiers in the ensemble. Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. Returns ------- y : array of shape = [n_samples] The predicted regression values. """ check_is_fitted(self, "estimator_weights_") X = self._validate_X_predict(X) return self._get_median_predict(X, len(self.estimators_)) def staged_predict(self, X): """Return staged predictions for X. The predicted regression value of an input sample is computed as the weighted median prediction of the classifiers in the ensemble. This generator method yields the ensemble prediction after each iteration of boosting and therefore allows monitoring, such as to determine the prediction on a test set after each boost. Parameters ---------- X : {array-like, sparse matrix} of shape = [n_samples, n_features] The training input samples. Sparse matrix can be CSC, CSR, COO, DOK, or LIL. DOK and LIL are converted to CSR. Returns ------- y : generator of array, shape = [n_samples] The predicted regression values. """ check_is_fitted(self, "estimator_weights_") X = self._validate_X_predict(X) for i, _ in enumerate(self.estimators_, 1): yield self._get_median_predict(X, limit=i)