Source code for yellowbrick.classifier.rocauc

# yellowbrick.classifier.rocauc
# Implements visual ROC/AUC curves for classification evaluation.
#
# Author:   Rebecca Bilbro
# Author:   Benjamin Bengfort
# Author:   Neal Humphrey
# Created:  Tue May 03 18:15:42 2017 -0400
#
# Copyright (C) 2016 The scikit-yb developers
# For license information, see LICENSE.txt
#
# ID: rocauc.py [5388065] neal@nhumphrey.com $

"""
Implements visual ROC/AUC curves for classification evaluation.
"""

##########################################################################
## Imports
##########################################################################

import numpy as np

from sklearn.metrics import auc, roc_curve
from sklearn.preprocessing import label_binarize
from sklearn.utils.multiclass import type_of_target

from yellowbrick.exceptions import ModelError
from yellowbrick.style.palettes import LINE_COLOR
from yellowbrick.exceptions import YellowbrickValueError
from yellowbrick.classifier.base import ClassificationScoreVisualizer


# Dictionary keys for ROCAUC
MACRO = "macro"
MICRO = "micro"

# Target Type Constants
BINARY = "binary"
MULTICLASS = "multiclass"


##########################################################################
## ROCAUC Visualizer
##########################################################################


[docs]class ROCAUC(ClassificationScoreVisualizer): """ Receiver Operating Characteristic (ROC) curves are a measure of a classifier's predictive quality that compares and visualizes the tradeoff between the models' sensitivity and specificity. The ROC curve displays the true positive rate on the Y axis and the false positive rate on the X axis on both a global average and per-class basis. The ideal point is therefore the top-left corner of the plot: false positives are zero and true positives are one. This leads to another metric, area under the curve (AUC), a computation of the relationship between false positives and true positives. The higher the AUC, the better the model generally is. However, it is also important to inspect the "steepness" of the curve, as this describes the maximization of the true positive rate while minimizing the false positive rate. Generalizing "steepness" usually leads to discussions about convexity, which we do not get into here. Parameters ---------- estimator : estimator A scikit-learn estimator that should be a classifier. If the model is not a classifier, an exception is raised. If the internal model is not fitted, it is fit when the visualizer is fitted, unless otherwise specified by ``is_fitted``. ax : matplotlib Axes, default: None The axes to plot the figure on. If not specified the current axes will be used (or generated if required). micro : bool, default: True Plot the micro-averages ROC curve, computed from the sum of all true positives and false positives across all classes. Micro is not defined for binary classification problems with estimators with only a decision_function method. macro : bool, default: True Plot the macro-averages ROC curve, which simply takes the average of curves across all classes. Macro is not defined for binary classification problems with estimators with only a decision_function method. per_class : bool, default: True Plot the ROC curves for each individual class. This should be set to false if only the macro or micro average curves are required. For true binary classifiers, setting per_class=False will plot the positive class ROC curve, and per_class=True will use ``1-P(1)`` to compute the curve of the negative class if only a decision_function method exists on the estimator. binary : bool, default: False This argument quickly resets the visualizer for true binary classification by updating the micro, macro, and per_class arguments to False (do not use in conjunction with those other arguments). Note that this is not a true hyperparameter to the visualizer, it just collects other parameters into a single, simpler argument. classes : list of str, defult: None The class labels to use for the legend ordered by the index of the sorted classes discovered in the ``fit()`` method. Specifying classes in this manner is used to change the class names to a more specific format or to label encoded integer classes. Some visualizers may also use this field to filter the visualization for specific classes. For more advanced usage specify an encoder rather than class labels. encoder : dict or LabelEncoder, default: None A mapping of classes to human readable labels. Often there is a mismatch between desired class labels and those contained in the target variable passed to ``fit()`` or ``score()``. The encoder disambiguates this mismatch ensuring that classes are labeled correctly in the visualization. is_fitted : bool or str, default="auto" Specify if the wrapped estimator is already fitted. If False, the estimator will be fit when the visualizer is fit, otherwise, the estimator will not be modified. If "auto" (default), a helper method will check if the estimator is fitted before fitting it again. force_model : bool, default: False Do not check to ensure that the underlying estimator is a classifier. This will prevent an exception when the visualizer is initialized but may result in unexpected or unintended behavior. kwargs : dict Keyword arguments passed to the visualizer base classes. Attributes ---------- classes_ : ndarray of shape (n_classes,) The class labels observed while fitting. class_count_ : ndarray of shape (n_classes,) Number of samples encountered for each class during fitting. score_ : float An evaluation metric of the classifier on test data produced when ``score()`` is called. This metric is between 0 and 1 -- higher scores are generally better. For classifiers, this score is usually accuracy, but if micro or macro is specified this returns an F1 score. target_type_ : string Specifies if the detected classification target was binary or multiclass. Notes ----- ROC curves are typically used in binary classification, and in fact the Scikit-Learn ``roc_curve`` metric is only able to perform metrics for binary classifiers. As a result it is necessary to binarize the output or to use one-vs-rest or one-vs-all strategies of classification. The visualizer does its best to handle multiple situations, but exceptions can arise from unexpected models or outputs. Another important point is the relationship of class labels specified on initialization to those drawn on the curves. The classes are not used to constrain ordering or filter curves; the ROC computation happens on the unique values specified in the target vector to the ``score`` method. To ensure the best quality visualization, do not use a LabelEncoder for this and do not pass in class labels. .. seealso:: http://scikit-learn.org/stable/auto_examples/model_selection/plot_roc.html .. todo:: Allow the class list to filter the curves on the visualization. Examples -------- >>> from yellowbrick.classifier import ROCAUC >>> from sklearn.linear_model import LogisticRegression >>> from sklearn.model_selection import train_test_split >>> data = load_data("occupancy") >>> features = ["temp", "relative humidity", "light", "C02", "humidity"] >>> X_train, X_test, y_train, y_test = train_test_split(X, y) >>> oz = ROCAUC(LogisticRegression()) >>> oz.fit(X_train, y_train) >>> oz.score(X_test, y_test) >>> oz.show() """ def __init__( self, estimator, ax=None, micro=True, macro=True, per_class=True, binary=False, classes=None, encoder=None, is_fitted="auto", force_model=False, **kwargs ): super(ROCAUC, self).__init__( estimator, ax=ax, classes=classes, encoder=encoder, is_fitted=is_fitted, force_model=force_model, **kwargs ) # Set the visual parameters for ROCAUC # NOTE: the binary flag breaks our API since it's really just a meta parameter # for micro, macro, and per_class. We knew this going in, but did it anyway. self.binary = binary if self.binary: self.micro = False self.macro = False self.per_class = False else: self.micro = micro self.macro = macro self.per_class = per_class
[docs] def fit(self, X, y=None): """ Fit the classification model. """ # The target determines what kind of estimator is fit ttype = type_of_target(y) if ttype.startswith(MULTICLASS): self.target_type_ = MULTICLASS elif ttype.startswith(BINARY): self.target_type_ = BINARY else: raise YellowbrickValueError( ( "{} does not support target type '{}', " "please provide a binary or multiclass single-output target" ).format(self.__class__.__name__, ttype) ) # Fit the model and return self return super(ROCAUC, self).fit(X, y)
[docs] def score(self, X, y=None): """ Generates the predicted target values using the Scikit-Learn estimator. Parameters ---------- X : ndarray or DataFrame of shape n x m A matrix of n instances with m features y : ndarray or Series of length n An array or series of target or class values Returns ------- score_ : float Global accuracy unless micro or macro scores are requested. """ # Call super to check if fitted and to compute self.score_ # NOTE: this sets score to the base score if neither macro nor micro super(ROCAUC, self).score(X, y) # Compute the predictions for the test data y_pred = self._get_y_scores(X) if self.target_type_ == BINARY: # For binary, per_class must be True to draw micro/macro curves if (self.micro or self.macro) and not self.per_class: raise ModelError( "no curves will be drawn; ", "set per_class=True or micro=False and macro=False.", ) # For binary, if predictions are returned in shape (n,), micro and macro # curves are not defined if (self.micro or self.macro) and len(y_pred.shape) == 1: raise ModelError( "no curves will be drawn; set binary=True.", ) if self.target_type_ == MULTICLASS: # If it's multiclass classification, at least one of micro, macro, or # per_class must be True if not self.micro and not self.macro and not self.per_class: raise YellowbrickValueError( "no curves will be drawn; specify micro, macro, or per_class" ) # Classes may be label encoded so only use what's in y to compute. # The self.classes_ attribute will be used as names for labels. classes = np.unique(y) n_classes = len(classes) # Store the false positive rate, true positive rate and curve info. self.fpr = dict() self.tpr = dict() self.roc_auc = dict() # If the decision is binary draw only ROC curve for the positive class if self.target_type_ is BINARY and not self.per_class: # In this case predict_proba returns an array of shape (n, 2) which # specifies the probabilities of both the negative and positive classes. if len(y_pred.shape) == 2 and y_pred.shape[1] == 2: self.fpr[BINARY], self.tpr[BINARY], _ = roc_curve(y, y_pred[:, 1]) else: # decision_function returns array of shape (n,), so plot it directly self.fpr[BINARY], self.tpr[BINARY], _ = roc_curve(y, y_pred) self.roc_auc[BINARY] = auc(self.fpr[BINARY], self.tpr[BINARY]) # Per-class binary decisions may have to have the negative class curve computed elif self.target_type_ is BINARY and self.per_class: # draw a curve for class 1 (the positive class) if len(y_pred.shape) == 2 and y_pred.shape[1] == 2: # predict_proba returns array of shape (n, 2), so use # probability of class 1 to compute ROC self.fpr[1], self.tpr[1], _ = roc_curve(y, y_pred[:, 1]) else: # decision_function returns array of shape (n,) self.fpr[1], self.tpr[1], _ = roc_curve(y, y_pred) self.roc_auc[1] = auc(self.fpr[1], self.tpr[1]) # draw a curve for class 0 (the negative class) if len(y_pred.shape) == 2 and y_pred.shape[1] == 2: # predict_proba returns array of shape (n, 2), so use # probability of class 0 to compute ROC self.fpr[0], self.tpr[0], _ = roc_curve(1 - y, y_pred[:, 0]) else: # decision_function returns array of shape (n,). # To draw a ROC curve for class 0 we swap the classes 0 and 1 in y # and reverse classifiers predictions y_pred. self.fpr[0], self.tpr[0], _ = roc_curve(1 - y, -y_pred) self.roc_auc[0] = auc(self.fpr[0], self.tpr[0]) else: # Otherwise compute the ROC curve and ROC area for each class for i, c in enumerate(classes): self.fpr[i], self.tpr[i], _ = roc_curve(y, y_pred[:, i], pos_label=c) self.roc_auc[i] = auc(self.fpr[i], self.tpr[i]) # Compute micro average if self.micro: self._score_micro_average(y, y_pred, classes, n_classes) # Compute macro average if self.macro: self._score_macro_average(n_classes) # Draw the Curves self.draw() # Set score to micro average if specified if self.micro: self.score_ = self.roc_auc[MICRO] # Set score to macro average if not micro if self.macro: self.score_ = self.roc_auc[MACRO] return self.score_
[docs] def draw(self): """ Renders ROC-AUC plot. Called internally by score, possibly more than once Returns ------- ax : the axis with the plotted figure """ colors = self.class_colors_[0 : len(self.classes_)] n_classes = len(colors) # If it's a binary decision, plot the single ROC curve if self.target_type_ == BINARY and not self.per_class: self.ax.plot( self.fpr[BINARY], self.tpr[BINARY], label="ROC for binary decision, AUC = {:0.2f}".format( self.roc_auc[BINARY] ), ) # If per-class plotting is requested, plot ROC curves for each class if self.per_class: for i, color in zip(range(n_classes), colors): self.ax.plot( self.fpr[i], self.tpr[i], color=color, label="ROC of class {}, AUC = {:0.2f}".format( self.classes_[i], self.roc_auc[i] ), ) # If requested, plot the ROC curve for the micro average if self.micro: self.ax.plot( self.fpr[MICRO], self.tpr[MICRO], linestyle="--", color=self.class_colors_[len(self.classes_) - 1], label="micro-average ROC curve, AUC = {:0.2f}".format( self.roc_auc["micro"] ), ) # If requested, plot the ROC curve for the macro average if self.macro: self.ax.plot( self.fpr[MACRO], self.tpr[MACRO], linestyle="--", color=self.class_colors_[len(self.classes_) - 1], label="macro-average ROC curve, AUC = {:0.2f}".format( self.roc_auc["macro"] ), ) # Plot the line of no discrimination to compare the curve to. self.ax.plot([0, 1], [0, 1], linestyle=":", c=LINE_COLOR) return self.ax
[docs] def finalize(self, **kwargs): """ Sets a title and axis labels of the figures and ensures the axis limits are scaled between the valid ROCAUC score values. Parameters ---------- kwargs: generic keyword arguments. Notes ----- Generally this method is called from show and not directly by the user. """ # Set the title and add the legend self.set_title("ROC Curves for {}".format(self.name)) self.ax.legend(loc="lower right", frameon=True) # Set the limits for the ROC/AUC (always between 0 and 1) self.ax.set_xlim([0.0, 1.0]) self.ax.set_ylim([0.0, 1.0]) # Set x and y axis labels self.ax.set_ylabel("True Positive Rate") self.ax.set_xlabel("False Positive Rate")
def _get_y_scores(self, X): """ The ``roc_curve`` metric requires target scores that can either be the probability estimates of the positive class, confidence values or non- thresholded measure of decisions (as returned by "decision_function"). This method computes the scores by resolving the estimator methods that retreive these values. .. todo:: implement confidence values metric. Parameters ---------- X : ndarray or DataFrame of shape n x m A matrix of n instances with m features -- generally the test data that is associated with y_true values. """ # The resolution order of scoring functions attrs = ("predict_proba", "decision_function") # Return the first resolved function for attr in attrs: try: method = getattr(self.estimator, attr, None) if method: return method(X) except AttributeError: # Some Scikit-Learn estimators have both probability and # decision functions but override __getattr__ and raise an # AttributeError on access. # Note that because of the ordering of our attrs above, # estimators with both will *only* ever use probability. continue # If we've gotten this far, raise an error raise ModelError( "ROCAUC requires estimators with predict_proba or " "decision_function methods." ) def _score_micro_average(self, y, y_pred, classes, n_classes): """ Compute the micro average scores for the ROCAUC curves. """ # Convert y to binarized array for micro and macro scores y = label_binarize(y, classes=classes) if n_classes == 2: y = np.hstack((1 - y, y)) # Compute micro-average self.fpr[MICRO], self.tpr[MICRO], _ = roc_curve(y.ravel(), y_pred.ravel()) self.roc_auc[MICRO] = auc(self.fpr[MICRO], self.tpr[MICRO]) def _score_macro_average(self, n_classes): """ Compute the macro average scores for the ROCAUC curves. """ # Gather all FPRs all_fpr = np.unique(np.concatenate([self.fpr[i] for i in range(n_classes)])) avg_tpr = np.zeros_like(all_fpr) # Compute the averages per class for i in range(n_classes): avg_tpr += np.interp(all_fpr, self.fpr[i], self.tpr[i]) # Finalize the average avg_tpr /= n_classes # Store the macro averages self.fpr[MACRO] = all_fpr self.tpr[MACRO] = avg_tpr self.roc_auc[MACRO] = auc(self.fpr[MACRO], self.tpr[MACRO])
########################################################################## ## Quick method for ROCAUC ##########################################################################
[docs]def roc_auc( estimator, X_train, y_train, X_test=None, y_test=None, ax=None, micro=True, macro=True, per_class=True, binary=False, classes=None, encoder=None, is_fitted="auto", force_model=False, show=True, **kwargs ): """ROCAUC Receiver Operating Characteristic (ROC) curves are a measure of a classifier's predictive quality that compares and visualizes the tradeoff between the models' sensitivity and specificity. The ROC curve displays the true positive rate on the Y axis and the false positive rate on the X axis on both a global average and per-class basis. The ideal point is therefore the top-left corner of the plot: false positives are zero and true positives are one. This leads to another metric, area under the curve (AUC), a computation of the relationship between false positives and true positives. The higher the AUC, the better the model generally is. However, it is also important to inspect the "steepness" of the curve, as this describes the maximization of the true positive rate while minimizing the false positive rate. Generalizing "steepness" usually leads to discussions about convexity, which we do not get into here. Parameters ---------- estimator : estimator A scikit-learn estimator that should be a classifier. If the model is not a classifier, an exception is raised. If the internal model is not fitted, it is fit when the visualizer is fitted, unless otherwise specified by ``is_fitted``. X_train : array-like, 2D The table of instance data or independent variables that describe the outcome of the dependent variable, y. Used to fit the visualizer and also to score the visualizer if test splits are not specified. y_train : array-like, 2D The vector of target data or the dependent variable predicted by X. Used to fit the visualizer and also to score the visualizer if test splits not specified. X_test: array-like, 2D, default: None The table of instance data or independent variables that describe the outcome of the dependent variable, y. Used to score the visualizer if specified. y_test: array-like, 1D, default: None The vector of target data or the dependent variable predicted by X. Used to score the visualizer if specified. ax : matplotlib Axes, default: None The axes to plot the figure on. If not specified the current axes will be used (or generated if required). test_size : float, default=0.2 The percentage of the data to reserve as test data. random_state : int or None, default=None The value to seed the random number generator for shuffling data. micro : bool, default: True Plot the micro-averages ROC curve, computed from the sum of all true positives and false positives across all classes. Micro is not defined for binary classification problems with estimators with only a decision_function method. macro : bool, default: True Plot the macro-averages ROC curve, which simply takes the average of curves across all classes. Macro is not defined for binary classification problems with estimators with only a decision_function method. per_class : bool, default: True Plot the ROC curves for each individual class. This should be set to false if only the macro or micro average curves are required. For true binary classifiers, setting per_class=False will plot the positive class ROC curve, and per_class=True will use ``1-P(1)`` to compute the curve of the negative class if only a decision_function method exists on the estimator. binary : bool, default: False This argument quickly resets the visualizer for true binary classification by updating the micro, macro, and per_class arguments to False (do not use in conjunction with those other arguments). Note that this is not a true hyperparameter to the visualizer, it just collects other parameters into a single, simpler argument. classes : list of str, defult: None The class labels to use for the legend ordered by the index of the sorted classes discovered in the ``fit()`` method. Specifying classes in this manner is used to change the class names to a more specific format or to label encoded integer classes. Some visualizers may also use this field to filter the visualization for specific classes. For more advanced usage specify an encoder rather than class labels. encoder : dict or LabelEncoder, default: None A mapping of classes to human readable labels. Often there is a mismatch between desired class labels and those contained in the target variable passed to ``fit()`` or ``score()``. The encoder disambiguates this mismatch ensuring that classes are labeled correctly in the visualization. is_fitted : bool or str, default="auto" Specify if the wrapped estimator is already fitted. If False, the estimator will be fit when the visualizer is fit, otherwise, the estimator will not be modified. If "auto" (default), a helper method will check if the estimator is fitted before fitting it again. force_model : bool, default: False Do not check to ensure that the underlying estimator is a classifier. This will prevent an exception when the visualizer is initialized but may result in unexpected or unintended behavior. show: bool, default: True If True, calls ``show()``, which in turn calls ``plt.show()`` however you cannot call ``plt.savefig`` from this signature, nor ``clear_figure``. If False, simply calls ``finalize()`` kwargs : dict Keyword arguments passed to the visualizer base classes. Notes ----- ROC curves are typically used in binary classification, and in fact the Scikit-Learn ``roc_curve`` metric is only able to perform metrics for binary classifiers. As a result it is necessary to binarize the output or to use one-vs-rest or one-vs-all strategies of classification. The visualizer does its best to handle multiple situations, but exceptions can arise from unexpected models or outputs. Another important point is the relationship of class labels specified on initialization to those drawn on the curves. The classes are not used to constrain ordering or filter curves; the ROC computation happens on the unique values specified in the target vector to the ``score`` method. To ensure the best quality visualization, do not use a LabelEncoder for this and do not pass in class labels. .. seealso:: https://bit.ly/2IORWO2 .. todo:: Allow the class list to filter the curves on the visualization. Examples -------- >>> from yellowbrick.classifier import ROCAUC >>> from sklearn.linear_model import LogisticRegression >>> data = load_data("occupancy") >>> features = ["temp", "relative humidity", "light", "C02", "humidity"] >>> X = data[features].values >>> y = data.occupancy.values >>> roc_auc(LogisticRegression(), X, y) Returns ------- viz : ROCAUC Returns the fitted, finalized visualizer object """ # Instantiate the visualizer visualizer = ROCAUC( estimator=estimator, ax=ax, micro=micro, macro=macro, per_class=per_class, binary=binary, classes=classes, encoder=encoder, is_fitted=is_fitted, force_model=force_model, **kwargs ) # Fit and transform the visualizer (calls draw) visualizer.fit(X_train, y_train, **kwargs) # Scores the visualizer with X_test and y_test if provided, # X_train, y_train if not provided if X_test is not None and y_test is not None: visualizer.score(X_test, y_test) else: visualizer.score(X_train, y_train) if show: visualizer.show() else: visualizer.finalize() # Return the visualizer return visualizer