# 3.3. 模型评估: 量化预测的质量
校验者:
[@飓风](https://github.com/apachecn/scikit-learn-doc-zh)
[@小瑶](https://github.com/apachecn/scikit-learn-doc-zh)
[@FAME](https://github.com/apachecn/scikit-learn-doc-zh)
[@v](https://github.com/apachecn/scikit-learn-doc-zh)
翻译者:
[@小瑶](https://github.com/apachecn/scikit-learn-doc-zh)
[@片刻](https://github.com/apachecn/scikit-learn-doc-zh)
[@那伊抹微笑](https://github.com/apachecn/scikit-learn-doc-zh)
有 3 种不同的 API 用于评估模型预测的质量:
- **Estimator score method(估计器得分的方法)**: Estimators(估计器)有一个 `score(得分)` 方法,为其解决的问题提供了默认的 evaluation criterion (评估标准)。 在这个页面上没有相关讨论,但是在每个 estimator (估计器)的文档中会有相关的讨论。
- **Scoring parameter(评分参数)**: Model-evaluation tools (模型评估工具)使用 [cross-validation](cross_validation.html#cross-validation) (如 [`model_selection.cross_val_score`](generated/sklearn.model_selection.cross_val_score.html#sklearn.model_selection.cross_val_score "sklearn.model_selection.cross_val_score") 和 [`model_selection.GridSearchCV`](generated/sklearn.model_selection.GridSearchCV.html#sklearn.model_selection.GridSearchCV "sklearn.model_selection.GridSearchCV")) 依靠 internal *scoring* strategy (内部 *scoring(得分)* 策略)。这在 [scoring 参数: 定义模型评估规则](#scoring-parameter) 部分讨论。
- **Metric functions(指标函数)**: `metrics` 模块实现了针对特定目的评估预测误差的函数。这些指标在以下部分部分详细介绍 [分类指标](#classification-metrics), [多标签排名指标](#multilabel-ranking-metrics), [回归指标](#regression-metrics) 和 [聚类指标](#clustering-metrics) 。
最后, [虚拟估计](#dummy-estimators) 用于获取随机预测的这些指标的基准值。
See also
对于 “pairwise(成对)” metrics(指标),*samples(样本)* 之间而不是 estimators (估计量)或者 predictions(预测值),请参阅 [成对的矩阵, 类别和核函数](metrics.html#metrics) 部分。
## 3.3.1. `scoring` 参数: 定义模型评估规则
Model selection (模型选择)和 evaluation (评估)使用工具,例如 [`model_selection.GridSearchCV`](generated/sklearn.model_selection.GridSearchCV.html#sklearn.model_selection.GridSearchCV "sklearn.model_selection.GridSearchCV") 和 [`model_selection.cross_val_score`](generated/sklearn.model_selection.cross_val_score.html#sklearn.model_selection.cross_val_score "sklearn.model_selection.cross_val_score") ,采用 `scoring` 参数来控制它们对 estimators evaluated (评估的估计量)应用的指标。
### 3.3.1.1. 常见场景: 预定义值
对于最常见的用例, 您可以使用 `scoring` 参数指定一个 scorer object (记分对象); 下表显示了所有可能的值。 所有 scorer objects (记分对象)遵循惯例 **higher return values are better than lower return values(较高的返回值优于较低的返回值)** 。因此,测量模型和数据之间距离的 metrics (度量),如 [`metrics.mean_squared_error`](generated/sklearn.metrics.mean_squared_error.html#sklearn.metrics.mean_squared_error "sklearn.metrics.mean_squared_error") 可用作返回 metric (指数)的 negated value (否定值)的 neg\_mean\_squared\_error 。
Scoring(得分)Function(函数)Comment(注解)**Classification(分类)** ‘accuracy’[`metrics.accuracy_score`](generated/sklearn.metrics.accuracy_score.html#sklearn.metrics.accuracy_score "sklearn.metrics.accuracy_score") ‘average\_precision’[`metrics.average_precision_score`](generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score "sklearn.metrics.average_precision_score") ‘f1’[`metrics.f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score")for binary targets(用于二进制目标)‘f1\_micro’[`metrics.f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score")micro-averaged(微平均)‘f1\_macro’[`metrics.f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score")macro-averaged(微平均)‘f1\_weighted’[`metrics.f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score")weighted average(加权平均)‘f1\_samples’[`metrics.f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score")by multilabel sample(通过 multilabel 样本)‘neg\_log\_loss’[`metrics.log_loss`](generated/sklearn.metrics.log_loss.html#sklearn.metrics.log_loss "sklearn.metrics.log_loss")requires `predict_proba` support(需要 `predict_proba` 支持)‘precision’ etc.[`metrics.precision_score`](generated/sklearn.metrics.precision_score.html#sklearn.metrics.precision_score "sklearn.metrics.precision_score")suffixes apply as with ‘f1’(后缀适用于 ‘f1’)‘recall’ etc.[`metrics.recall_score`](generated/sklearn.metrics.recall_score.html#sklearn.metrics.recall_score "sklearn.metrics.recall_score")suffixes apply as with ‘f1’(后缀适用于 ‘f1’)‘roc\_auc’[`metrics.roc_auc_score`](generated/sklearn.metrics.roc_auc_score.html#sklearn.metrics.roc_auc_score "sklearn.metrics.roc_auc_score") **Clustering(聚类)** ‘adjusted\_mutual\_info\_score’[`metrics.adjusted_mutual_info_score`](generated/sklearn.metrics.adjusted_mutual_info_score.html#sklearn.metrics.adjusted_mutual_info_score "sklearn.metrics.adjusted_mutual_info_score") ‘adjusted\_rand\_score’[`metrics.adjusted_rand_score`](generated/sklearn.metrics.adjusted_rand_score.html#sklearn.metrics.adjusted_rand_score "sklearn.metrics.adjusted_rand_score") ‘completeness\_score’[`metrics.completeness_score`](generated/sklearn.metrics.completeness_score.html#sklearn.metrics.completeness_score "sklearn.metrics.completeness_score") ‘fowlkes\_mallows\_score’[`metrics.fowlkes_mallows_score`](generated/sklearn.metrics.fowlkes_mallows_score.html#sklearn.metrics.fowlkes_mallows_score "sklearn.metrics.fowlkes_mallows_score") ‘homogeneity\_score’[`metrics.homogeneity_score`](generated/sklearn.metrics.homogeneity_score.html#sklearn.metrics.homogeneity_score "sklearn.metrics.homogeneity_score") ‘mutual\_info\_score’[`metrics.mutual_info_score`](generated/sklearn.metrics.mutual_info_score.html#sklearn.metrics.mutual_info_score "sklearn.metrics.mutual_info_score") ‘normalized\_mutual\_info\_score’[`metrics.normalized_mutual_info_score`](generated/sklearn.metrics.normalized_mutual_info_score.html#sklearn.metrics.normalized_mutual_info_score "sklearn.metrics.normalized_mutual_info_score") ‘v\_measure\_score’[`metrics.v_measure_score`](generated/sklearn.metrics.v_measure_score.html#sklearn.metrics.v_measure_score "sklearn.metrics.v_measure_score") **Regression(回归)** ‘explained\_variance’[`metrics.explained_variance_score`](generated/sklearn.metrics.explained_variance_score.html#sklearn.metrics.explained_variance_score "sklearn.metrics.explained_variance_score") ‘neg\_mean\_absolute\_error’[`metrics.mean_absolute_error`](generated/sklearn.metrics.mean_absolute_error.html#sklearn.metrics.mean_absolute_error "sklearn.metrics.mean_absolute_error") ‘neg\_mean\_squared\_error’[`metrics.mean_squared_error`](generated/sklearn.metrics.mean_squared_error.html#sklearn.metrics.mean_squared_error "sklearn.metrics.mean_squared_error") ‘neg\_mean\_squared\_log\_error’[`metrics.mean_squared_log_error`](generated/sklearn.metrics.mean_squared_log_error.html#sklearn.metrics.mean_squared_log_error "sklearn.metrics.mean_squared_log_error") ‘neg\_median\_absolute\_error’[`metrics.median_absolute_error`](generated/sklearn.metrics.median_absolute_error.html#sklearn.metrics.median_absolute_error "sklearn.metrics.median_absolute_error") ‘r2’[`metrics.r2_score`](generated/sklearn.metrics.r2_score.html#sklearn.metrics.r2_score "sklearn.metrics.r2_score") 使用案例:
```
>>> from sklearn import svm, datasets
>>> from sklearn.model_selection import cross_val_score
>>> iris = datasets.load_iris()
>>> X, y = iris.data, iris.target
>>> clf = svm.SVC(probability=True, random_state=0)
>>> cross_val_score(clf, X, y, scoring='neg_log_loss')
array([-0.07..., -0.16..., -0.06...])
>>> model = svm.SVC()
>>> cross_val_score(model, X, y, scoring='wrong_choice')
Traceback (most recent call last):
ValueError: 'wrong_choice' is not a valid scoring value. Valid options are ['accuracy', 'adjusted_mutual_info_score', 'adjusted_rand_score', 'average_precision', 'completeness_score', 'explained_variance', 'f1', 'f1_macro', 'f1_micro', 'f1_samples', 'f1_weighted', 'fowlkes_mallows_score', 'homogeneity_score', 'mutual_info_score', 'neg_log_loss', 'neg_mean_absolute_error', 'neg_mean_squared_error', 'neg_mean_squared_log_error', 'neg_median_absolute_error', 'normalized_mutual_info_score', 'precision', 'precision_macro', 'precision_micro', 'precision_samples', 'precision_weighted', 'r2', 'recall', 'recall_macro', 'recall_micro', 'recall_samples', 'recall_weighted', 'roc_auc', 'v_measure_score']
```
Note
ValueError exception 列出的值对应于以下部分描述的 functions measuring prediction accuracy (测量预测精度的函数)。 这些函数的 scorer objects (记分对象)存储在 dictionary `sklearn.metrics.SCORERS` 中。
### 3.3.1.2. 根据 metric 函数定义您的评分策略
模块 [`sklearn.metrics`](classes.html#module-sklearn.metrics "sklearn.metrics") 还公开了一组 measuring a prediction error (测量预测误差)的简单函数,给出了基础真实的数据和预测:
- 函数以 `_score` 结尾返回一个值来最大化,越高越好。
- 函数 `_error` 或 `_loss` 结尾返回一个值来 minimize (最小化),越低越好。当使用 [`make_scorer`](generated/sklearn.metrics.make_scorer.html#sklearn.metrics.make_scorer "sklearn.metrics.make_scorer") 转换成 scorer object (记分对象)时,将 `greater_is_better` 参数设置为 False(默认为 True; 请参阅下面的参数说明)。
可用于各种机器学习任务的 Metrics (指标)在下面详细介绍。
许多 metrics (指标)没有被用作 `scoring(得分)` 值的名称,有时是因为它们需要额外的参数,例如 [`fbeta_score`](generated/sklearn.metrics.fbeta_score.html#sklearn.metrics.fbeta_score "sklearn.metrics.fbeta_score") 。在这种情况下,您需要生成一个适当的 scoring object (评分对象)。生成 callable object for scoring (可评估对象进行评分)的最简单方法是使用 [`make_scorer`](generated/sklearn.metrics.make_scorer.html#sklearn.metrics.make_scorer "sklearn.metrics.make_scorer") 。该函数将 metrics (指数)转换为可用于可调用的 model evaluation (模型评估)。
一个典型的用例是从库中包含一个非默认值参数的 existing metric function (现有指数函数),例如 [`fbeta_score`](generated/sklearn.metrics.fbeta_score.html#sklearn.metrics.fbeta_score "sklearn.metrics.fbeta_score") 函数的 `beta` 参数:
```
>>> from sklearn.metrics import fbeta_score, make_scorer
>>> ftwo_scorer = make_scorer(fbeta_score, beta=2)
>>> from sklearn.model_selection import GridSearchCV
>>> from sklearn.svm import LinearSVC
>>> grid = GridSearchCV(LinearSVC(), param_grid={'C': [1, 10]}, scoring=ftwo_scorer)
```
第二个用例是使用 [`make_scorer`](generated/sklearn.metrics.make_scorer.html#sklearn.metrics.make_scorer "sklearn.metrics.make_scorer") 从简单的 python 函数构建一个完全 custom scorer object (自定义的记分对象),可以使用几个参数 :
- 你要使用的 python 函数(在下面的例子中是 `my_custom_loss_func`)
- python 函数是否返回一个分数 (`greater_is_better=True`, 默认值) 或者一个 loss (损失) (`greater_is_better=False`)。 如果是一个 loss (损失),scorer object (记分对象)的 python 函数的输出被 negated (否定),符合 cross validation convention (交叉验证约定),scorers 为更好的模型返回更高的值。
- 仅用于 classification metrics (分类指数): 您提供的 python 函数是否需要连续的 continuous decision certainties (判断确定性)(`needs_threshold=True`)。默认值为 False 。
- 任何其他参数,如 `beta` 或者 `labels` 在 函数 [`f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score") 。
以下是建立 custom scorers (自定义记分对象)的示例,并使用 `greater_is_better` 参数:
```
>>> import numpy as np
>>> def my_custom_loss_func(ground_truth, predictions):
... diff = np.abs(ground_truth - predictions).max()
... return np.log(1 + diff)
...
>>> # loss_func will negate the return value of my_custom_loss_func,
>>> # which will be np.log(2), 0.693, given the values for ground_truth
>>> # and predictions defined below.
>>> loss = make_scorer(my_custom_loss_func, greater_is_better=False)
>>> score = make_scorer(my_custom_loss_func, greater_is_better=True)
>>> ground_truth = [[1], [1]]
>>> predictions = [0, 1]
>>> from sklearn.dummy import DummyClassifier
>>> clf = DummyClassifier(strategy='most_frequent', random_state=0)
>>> clf = clf.fit(ground_truth, predictions)
>>> loss(clf,ground_truth, predictions)
-0.69...
>>> score(clf,ground_truth, predictions)
0.69...
```
### 3.3.1.3. 实现自己的记分对象
您可以通过从头开始构建自己的 scoring object (记分对象),而不使用 [`make_scorer`](generated/sklearn.metrics.make_scorer.html#sklearn.metrics.make_scorer "sklearn.metrics.make_scorer") factory 来生成更加灵活的 model scorers (模型记分对象)。 对于被叫做 scorer 来说,它需要符合以下两个规则所指定的协议:
- 可以使用参数 `(estimator, X, y)` 来调用它,其中 `estimator` 是要被评估的模型,`X` 是验证数据, `y` 是 `X` (在有监督情况下) 或 `None` (在无监督情况下) 已经被标注的真实数据目标。
- 它返回一个浮点数,用于对 `X` 进行量化 `estimator` 的预测质量,参考 `y` 。 再次,按照惯例,更高的数字更好,所以如果你的 scorer 返回 loss ,那么这个值应该被 negated 。
### 3.3.1.4. 使用多个指数评估
Scikit-learn 还允许在 `GridSearchCV`, `RandomizedSearchCV` 和 `cross_validate` 中评估 multiple metric (多个指数)。
为 `scoring` 参数指定多个评分指标有两种方法:
- As an iterable of string metrics(作为 string metrics 的迭代)::
```
>>> scoring = ['accuracy', 'precision']
```
- As a `dict` mapping the scorer name to the scoring function(作为 `dict` ,将 scorer 名称映射到 scoring 函数)::
```
>>> from sklearn.metrics import accuracy_score
>>> from sklearn.metrics import make_scorer
>>> scoring = {'accuracy': make_scorer(accuracy_score),
... 'prec': 'precision'}
```
请注意, dict 值可以是 scorer functions (记分函数)或者 predefined metric strings (预定义 metric 字符串)之一。
目前,只有那些返回 single score (单一分数)的 scorer functions (记分函数)才能在 dict 内传递。不允许返回多个值的 Scorer functions (Scorer 函数),并且需要一个 wrapper 才能返回 single metric(单个指标):
```
>>> from sklearn.model_selection import cross_validate
>>> from sklearn.metrics import confusion_matrix
>>> # A sample toy binary classification dataset
>>> X, y = datasets.make_classification(n_classes=2, random_state=0)
>>> svm = LinearSVC(random_state=0)
>>> def tp(y_true, y_pred): return confusion_matrix(y_true, y_pred)[0, 0]
>>> def tn(y_true, y_pred): return confusion_matrix(y_true, y_pred)[0, 0]
>>> def fp(y_true, y_pred): return confusion_matrix(y_true, y_pred)[1, 0]
>>> def fn(y_true, y_pred): return confusion_matrix(y_true, y_pred)[0, 1]
>>> scoring = {'tp' : make_scorer(tp), 'tn' : make_scorer(tn),
... 'fp' : make_scorer(fp), 'fn' : make_scorer(fn)}
>>> cv_results = cross_validate(svm.fit(X, y), X, y, scoring=scoring)
>>> # Getting the test set true positive scores
>>> print(cv_results['test_tp'])
[12 13 15]
>>> # Getting the test set false negative scores
>>> print(cv_results['test_fn'])
[5 4 1]
```
## 3.3.2. 分类指标
[`sklearn.metrics`](classes.html#module-sklearn.metrics "sklearn.metrics") 模块实现了几个 loss, score, 和 utility 函数来衡量 classification (分类)性能。 某些 metrics (指标)可能需要 positive class (正类),confidence values(置信度值)或 binary decisions values (二进制决策值)的概率估计。 大多数的实现允许每个样本通过 `sample_weight` 参数为 overall score (总分)提供 weighted contribution (加权贡献)。
其中一些仅限于二分类案例:
[`precision_recall_curve`](generated/sklearn.metrics.precision_recall_curve.html#sklearn.metrics.precision_recall_curve "sklearn.metrics.precision_recall_curve")(y\_true, probas\_pred)Compute precision-recall pairs for different probability thresholds[`roc_curve`](generated/sklearn.metrics.roc_curve.html#sklearn.metrics.roc_curve "sklearn.metrics.roc_curve")(y\_true, y\_score\[, pos\_label, …\])Compute Receiver operating characteristic (ROC)其他也可以在多分类案例中运行:
[`cohen_kappa_score`](generated/sklearn.metrics.cohen_kappa_score.html#sklearn.metrics.cohen_kappa_score "sklearn.metrics.cohen_kappa_score")(y1, y2\[, labels, weights, …\])Cohen’s kappa: a statistic that measures inter-annotator agreement.[`confusion_matrix`](generated/sklearn.metrics.confusion_matrix.html#sklearn.metrics.confusion_matrix "sklearn.metrics.confusion_matrix")(y\_true, y\_pred\[, labels, …\])Compute confusion matrix to evaluate the accuracy of a classification[`hinge_loss`](generated/sklearn.metrics.hinge_loss.html#sklearn.metrics.hinge_loss "sklearn.metrics.hinge_loss")(y\_true, pred\_decision\[, labels, …\])Average hinge loss (non-regularized)[`matthews_corrcoef`](generated/sklearn.metrics.matthews_corrcoef.html#sklearn.metrics.matthews_corrcoef "sklearn.metrics.matthews_corrcoef")(y\_true, y\_pred\[, …\])Compute the Matthews correlation coefficient (MCC)有些还可以在 multilabel case (多重案例)中工作:
[`accuracy_score`](generated/sklearn.metrics.accuracy_score.html#sklearn.metrics.accuracy_score "sklearn.metrics.accuracy_score")(y\_true, y\_pred\[, normalize, …\])Accuracy classification score.[`classification_report`](generated/sklearn.metrics.classification_report.html#sklearn.metrics.classification_report "sklearn.metrics.classification_report")(y\_true, y\_pred\[, …\])Build a text report showing the main classification metrics[`f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score")(y\_true, y\_pred\[, labels, …\])Compute the F1 score, also known as balanced F-score or F-measure[`fbeta_score`](generated/sklearn.metrics.fbeta_score.html#sklearn.metrics.fbeta_score "sklearn.metrics.fbeta_score")(y\_true, y\_pred, beta\[, labels, …\])Compute the F-beta score[`hamming_loss`](generated/sklearn.metrics.hamming_loss.html#sklearn.metrics.hamming_loss "sklearn.metrics.hamming_loss")(y\_true, y\_pred\[, labels, …\])Compute the average Hamming loss.[`jaccard_similarity_score`](generated/sklearn.metrics.jaccard_similarity_score.html#sklearn.metrics.jaccard_similarity_score "sklearn.metrics.jaccard_similarity_score")(y\_true, y\_pred\[, …\])Jaccard similarity coefficient score[`log_loss`](generated/sklearn.metrics.log_loss.html#sklearn.metrics.log_loss "sklearn.metrics.log_loss")(y\_true, y\_pred\[, eps, normalize, …\])Log loss, aka logistic loss or cross-entropy loss.[`precision_recall_fscore_support`](generated/sklearn.metrics.precision_recall_fscore_support.html#sklearn.metrics.precision_recall_fscore_support "sklearn.metrics.precision_recall_fscore_support")(y\_true, y\_pred)Compute precision, recall, F-measure and support for each class[`precision_score`](generated/sklearn.metrics.precision_score.html#sklearn.metrics.precision_score "sklearn.metrics.precision_score")(y\_true, y\_pred\[, labels, …\])Compute the precision[`recall_score`](generated/sklearn.metrics.recall_score.html#sklearn.metrics.recall_score "sklearn.metrics.recall_score")(y\_true, y\_pred\[, labels, …\])Compute the recall[`zero_one_loss`](generated/sklearn.metrics.zero_one_loss.html#sklearn.metrics.zero_one_loss "sklearn.metrics.zero_one_loss")(y\_true, y\_pred\[, normalize, …\])Zero-one classification loss.一些通常用于 ranking:
[`dcg_score`](generated/sklearn.metrics.dcg_score.html#sklearn.metrics.dcg_score "sklearn.metrics.dcg_score")(y\_true, y\_score\[, k\])Discounted cumulative gain (DCG) at rank K.[`ndcg_score`](generated/sklearn.metrics.ndcg_score.html#sklearn.metrics.ndcg_score "sklearn.metrics.ndcg_score")(y\_true, y\_score\[, k\])Normalized discounted cumulative gain (NDCG) at rank K.有些工作与 binary 和 multilabel (但不是多类)的问题:
[`average_precision_score`](generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score "sklearn.metrics.average_precision_score")(y\_true, y\_score\[, …\])Compute average precision (AP) from prediction scores[`roc_auc_score`](generated/sklearn.metrics.roc_auc_score.html#sklearn.metrics.roc_auc_score "sklearn.metrics.roc_auc_score")(y\_true, y\_score\[, average, …\])Compute Area Under the Curve (AUC) from prediction scores在以下小节中,我们将介绍每个这些功能,前面是一些关于通用 API 和 metric 定义的注释。
### 3.3.2.1. 从二分到多分类和 multilabel
一些 metrics 基本上是为 binary classification tasks (二分类任务)定义的 (例如 [`f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score"), [`roc_auc_score`](generated/sklearn.metrics.roc_auc_score.html#sklearn.metrics.roc_auc_score "sklearn.metrics.roc_auc_score")) 。在这些情况下,默认情况下仅评估 positive label (正标签),假设默认情况下,positive label (正类)标记为 `1` (尽管可以通过 `pos_label` 参数进行配置)。
将 binary metric (二分指标)扩展为 multiclass (多类)或 multilabel (多标签)问题时,数据将被视为二分问题的集合,每个类都有一个。 然后可以使用多种方法在整个类中 average binary metric calculations (平均二分指标计算),每种类在某些情况下可能会有用。 如果可用,您应该使用 `average` 参数来选择它们。
- `"macro(宏)"` 简单地计算 binary metrics (二分指标)的平均值,赋予每个类别相同的权重。在不常见的类别重要的问题上,macro-averaging (宏观平均)可能是突出表现的一种手段。另一方面,所有类别同样重要的假设通常是不真实的,因此 macro-averaging (宏观平均)将过度强调不频繁类的典型的低性能。
- `"weighted(加权)"` 通过计算其在真实数据样本中的存在来对每个类的 score 进行加权的 binary metrics (二分指标)的平均值来计算类不平衡。
- `"micro(微)"` 给每个 sample-class pair (样本类对)对 overall metric (总体指数)(sample-class 权重的结果除外) 等同的贡献。除了对每个类别的 metric 进行求和之外,这个总和构成每个类别度量的 dividends (除数)和 divisors (除数)计算一个整体商。 在 multilabel settings (多标签设置)中,Micro-averaging 可能是优先选择的,包括要忽略 majority class (多数类)的 multiclass classification (多类分类)。
- `"samples(样本)"` 仅适用于 multilabel problems (多标签问题)。它 does not calculate a per-class measure (不计算每个类别的 measure),而是计算 evaluation data (评估数据)中的每个样本的 true and predicted classes (真实和预测类别)的 metric (指标),并返回 (`sample_weight`-weighted) 加权平均。
- 选择 `average=None` 将返回一个 array 与每个类的 score 。
虽然将 multiclass data (多类数据)提供给 metric ,如 binary targets (二分类目标),作为 array of class labels (类标签的数组),multilabel data (多标签数据)被指定为 indicator matrix(指示符矩阵),其中 cell `[i, j]` 具有值 1,如果样本 `i` 具有标号 `j` ,否则为值 0 。
### 3.3.2.2. 精确度得分
[`accuracy_score`](generated/sklearn.metrics.accuracy_score.html#sklearn.metrics.accuracy_score "sklearn.metrics.accuracy_score") 函数计算 [accuracy](https://en.wikipedia.org/wiki/Accuracy_and_precision), 正确预测的分数(默认)或计数 (normalize=False)。
在 multilabel classification (多标签分类)中,函数返回 subset accuracy(子集精度)。如果样本的 entire set of predicted labels (整套预测标签)与真正的标签组合匹配,则子集精度为 1.0; 否则为 0.0 。
如果 ![\hat{y}_i](https://box.kancloud.cn/a71dac1a33f465fa5197c33da8720585_13x17.jpg) 是第 ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg) 个样本的预测值,![y_i](https://box.kancloud.cn/e79627211612bc56c4f7d926a93fbe8d_13x12.jpg) 是相应的真实值,则 ![n_\text{samples}](https://box.kancloud.cn/70c64e65bf204369a0c85d76a39ecf38_55x14.jpg) 上的正确预测的分数被定义为
![\texttt{accuracy}(y, \hat{y}) = \frac{1}{n_\text{samples}} \sum_{i=0}^{n_\text{samples}-1} 1(\hat{y}_i = y_i)](https://box.kancloud.cn/809171f366a6d6d848e9d06d2132880f_341x55.jpg)
其中 ![1(x)](https://box.kancloud.cn/89c489280f4c7e8dd0c3c421b0d4358b_31x18.jpg) 是 [indicator function(指示函数)](https://en.wikipedia.org/wiki/Indicator_function).
```
>>> import numpy as np
>>> from sklearn.metrics import accuracy_score
>>> y_pred = [0, 2, 1, 3]
>>> y_true = [0, 1, 2, 3]
>>> accuracy_score(y_true, y_pred)
0.5
>>> accuracy_score(y_true, y_pred, normalize=False)
2
```
In the multilabel case with binary label indicators(在具有二分标签指示符的多标签情况下):
```
>>> accuracy_score(np.array([[0, 1], [1, 1]]), np.ones((2, 2)))
0.5
```
示例:
- 参阅 [Test with permutations the significance of a classification score](../auto_examples/feature_selection/plot_permutation_test_for_classification.html#sphx-glr-auto-examples-feature-selection-plot-permutation-test-for-classification-py)例如使用数据集排列的 accuracy score (精度分数)。
### 3.3.2.3. Cohen’s kappa
函数 [`cohen_kappa_score`](generated/sklearn.metrics.cohen_kappa_score.html#sklearn.metrics.cohen_kappa_score "sklearn.metrics.cohen_kappa_score") 计算 [Cohen’s kappa](https://en.wikipedia.org/wiki/Cohen%27s_kappa) statistic(统计)。 这个 measure (措施)旨在比较不同人工标注者的标签,而不是 classifier (分类器)与 ground truth (真实数据)。
kappa score (参阅 docstring )是 -1 和 1 之间的数字。 .8 以上的 scores 通常被认为是很好的 agreement (协议); 0 或者 更低表示没有 agreement (实际上是 random labels (随机标签))。
Kappa scores 可以计算 binary or multiclass (二分或者多分类)问题,但不能用于 multilabel problems (多标签问题)(除了手动计算 per-label score (每个标签分数)),而不是两个以上的 annotators (注释器)。
```
>>> from sklearn.metrics import cohen_kappa_score
>>> y_true = [2, 0, 2, 2, 0, 1]
>>> y_pred = [0, 0, 2, 2, 0, 2]
>>> cohen_kappa_score(y_true, y_pred)
0.4285714285714286
```
### 3.3.2.4. 混淆矩阵
[`confusion_matrix`](generated/sklearn.metrics.confusion_matrix.html#sklearn.metrics.confusion_matrix "sklearn.metrics.confusion_matrix") 函数通过计算 [confusion matrix(混淆矩阵)](https://en.wikipedia.org/wiki/Confusion_matrix) 来 evaluates classification accuracy (评估分类的准确性)。
根据定义,confusion matrix (混淆矩阵)中的 entry(条目) ![i, j](https://box.kancloud.cn/d2f39b1461f90c8e29c7740511a49fbb_22x17.jpg),是实际上在 group ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg) 中的 observations (观察数),但预测在 group ![j](https://box.kancloud.cn/7c1f52c4cc615183dbd16304bc8c1e94_9x16.jpg) 中。这里是一个示例:
```
>>> from sklearn.metrics import confusion_matrix
>>> y_true = [2, 0, 2, 2, 0, 1]
>>> y_pred = [0, 0, 2, 2, 0, 2]
>>> confusion_matrix(y_true, y_pred)
array([[2, 0, 0],
[0, 0, 1],
[1, 0, 2]])
```
这是一个这样的 confusion matrix (混淆矩阵)的可视化表示 (这个数字来自于 [Confusion matrix](../auto_examples/model_selection/plot_confusion_matrix.html#sphx-glr-auto-examples-model-selection-plot-confusion-matrix-py)):
[![http://sklearn.apachecn.org/cn/0.19.0/_images/sphx_glr_plot_confusion_matrix_0011.png](https://box.kancloud.cn/2aeb2d020c8bce222c39a58e3747112f_566x424.jpg)](../auto_examples/model_selection/plot_confusion_matrix.html)对于 binary problems (二分类问题),我们可以得到 true negatives(真 negatives), false positives(假 positives), false negatives(假 negatives) 和 true positives(真 positives) 的数量如下:
```
>>> y_true = [0, 0, 0, 1, 1, 1, 1, 1]
>>> y_pred = [0, 1, 0, 1, 0, 1, 0, 1]
>>> tn, fp, fn, tp = confusion_matrix(y_true, y_pred).ravel()
>>> tn, fp, fn, tp
(2, 1, 2, 3)
```
示例:
- 参阅 [Confusion matrix](../auto_examples/model_selection/plot_confusion_matrix.html#sphx-glr-auto-examples-model-selection-plot-confusion-matrix-py)例如使用 confusion matrix (混淆矩阵)来评估 classifier (分类器)的输出质量。
- 参阅 [Recognizing hand-written digits](../auto_examples/classification/plot_digits_classification.html#sphx-glr-auto-examples-classification-plot-digits-classification-py)例如使用 confusion matrix (混淆矩阵)来分类手写数字。
- 参阅 [Classification of text documents using sparse features](../auto_examples/text/document_classification_20newsgroups.html#sphx-glr-auto-examples-text-document-classification-20newsgroups-py)例如使用 confusion matrix (混淆矩阵)对文本文档进行分类。
### 3.3.2.5. 分类报告
[`classification_report`](generated/sklearn.metrics.classification_report.html#sklearn.metrics.classification_report "sklearn.metrics.classification_report") 函数构建一个显示 main classification metrics (主分类指标)的文本报告。这是一个小例子,其中包含自定义的 `target_names` 和 inferred labels (推断标签):
```
>>> from sklearn.metrics import classification_report
>>> y_true = [0, 1, 2, 2, 0]
>>> y_pred = [0, 0, 2, 1, 0]
>>> target_names = ['class 0', 'class 1', 'class 2']
>>> print(classification_report(y_true, y_pred, target_names=target_names))
precision recall f1-score support
class 0 0.67 1.00 0.80 2
class 1 0.00 0.00 0.00 1
class 2 1.00 0.50 0.67 2
avg / total 0.67 0.60 0.59 5
```
示例:
- 参阅 [Recognizing hand-written digits](../auto_examples/classification/plot_digits_classification.html#sphx-glr-auto-examples-classification-plot-digits-classification-py)作为手写数字的分类报告的使用示例。
- 参阅 [Classification of text documents using sparse features](../auto_examples/text/document_classification_20newsgroups.html#sphx-glr-auto-examples-text-document-classification-20newsgroups-py)作为文本文档的分类报告使用的示例。
- 参阅 [Parameter estimation using grid search with cross-validation](../auto_examples/model_selection/plot_grid_search_digits.html#sphx-glr-auto-examples-model-selection-plot-grid-search-digits-py)例如使用 grid search with nested cross-validation (嵌套交叉验证进行网格搜索)的分类报告。
### 3.3.2.6. 汉明损失
[`hamming_loss`](generated/sklearn.metrics.hamming_loss.html#sklearn.metrics.hamming_loss "sklearn.metrics.hamming_loss") 计算两组样本之间的 average Hamming loss (平均汉明损失)或者 [Hamming distance(汉明距离)](https://en.wikipedia.org/wiki/Hamming_distance) 。
如果 ![\hat{y}_j](https://box.kancloud.cn/861bda2ee215ffca90db3a0d09904009_15x19.jpg) 是给定样本的第 ![j](https://box.kancloud.cn/7c1f52c4cc615183dbd16304bc8c1e94_9x16.jpg) 个标签的预测值,则 ![y_j](https://box.kancloud.cn/677cd283282d3a460d2aa6e3eb63ac0c_15x14.jpg) 是相应的真实值,而 ![n_\text{labels}](https://box.kancloud.cn/da46b8c2d704c4e72eb3d96ac2a94be8_43x12.jpg) 是 classes or labels (类或者标签)的数量,则两个样本之间的 Hamming loss (汉明损失) ![L_{Hamming}](https://box.kancloud.cn/8ab12c0818ad4fd4021b518fafe01faa_72x18.jpg) 定义为:
![L_{Hamming}(y, \hat{y}) = \frac{1}{n_\text{labels}} \sum_{j=0}^{n_\text{labels} - 1} 1(\hat{y}_j \not= y_j)](https://box.kancloud.cn/cacfcbf9798a368e1b7b9700a55759cd_321x56.jpg)
其中 ![1(x)](https://box.kancloud.cn/89c489280f4c7e8dd0c3c421b0d4358b_31x18.jpg) 是 [indicator function(指标函数)](https://en.wikipedia.org/wiki/Indicator_function).
```
>>> from sklearn.metrics import hamming_loss
>>> y_pred = [1, 2, 3, 4]
>>> y_true = [2, 2, 3, 4]
>>> hamming_loss(y_true, y_pred)
0.25
```
在具有 binary label indicators (二分标签指示符)的 multilabel (多标签)情况下:
```
>>> hamming_loss(np.array([[0, 1], [1, 1]]), np.zeros((2, 2)))
0.75
```
Note
在 multiclass classification (多类分类)中, Hamming loss (汉明损失)对应于 `y_true` 和 `y_pred` 之间的 Hamming distance(汉明距离),它类似于 [零一损失](#zero-one-loss) 函数。然而, zero-one loss penalizes (0-1损失惩罚)不严格匹配真实集合的预测集,Hamming loss (汉明损失)惩罚 individual labels (独立标签)。因此,Hamming loss(汉明损失)高于 zero-one loss(0-1 损失),总是在 0 和 1 之间,包括 0 和 1;预测真正的标签的正确的 subset or superset (子集或超集)将给出 0 和 1 之间的 Hamming loss(汉明损失)。
### 3.3.2.7. Jaccard 相似系数 score
[`jaccard_similarity_score`](generated/sklearn.metrics.jaccard_similarity_score.html#sklearn.metrics.jaccard_similarity_score "sklearn.metrics.jaccard_similarity_score") 函数计算 pairs of label sets (标签组对)之间的 [Jaccard similarity coefficients](https://en.wikipedia.org/wiki/Jaccard_index) 也称作 Jaccard index 的平均值(默认)或总和。
将第 ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg) 个样本的 Jaccard similarity coefficient 与 被标注过的真实数据的标签集 ![y_i](https://box.kancloud.cn/e79627211612bc56c4f7d926a93fbe8d_13x12.jpg) 和 predicted label set (预测标签集):math:hat{y}\_i 定义为
![J(y_i, \hat{y}_i) = \frac{|y_i \cap \hat{y}_i|}{|y_i \cup \hat{y}_i|}.](https://box.kancloud.cn/c1610da1d48f42a710e13784d9a0b237_150x44.jpg)
在 binary and multiclass classification (二分和多类分类)中,Jaccard similarity coefficient score 等于 classification accuracy(分类精度)。
```
>>> import numpy as np
>>> from sklearn.metrics import jaccard_similarity_score
>>> y_pred = [0, 2, 1, 3]
>>> y_true = [0, 1, 2, 3]
>>> jaccard_similarity_score(y_true, y_pred)
0.5
>>> jaccard_similarity_score(y_true, y_pred, normalize=False)
2
```
在具有 binary label indicators (二分标签指示符)的 multilabel (多标签)情况下:
```
>>> jaccard_similarity_score(np.array([[0, 1], [1, 1]]), np.ones((2, 2)))
0.75
```
### 3.3.2.8. 精准,召回和 F-measures
直观地来理解,[precision](https://en.wikipedia.org/wiki/Precision_and_recall#Precision) 是 the ability of the classifier not to label as positive a sample that is negative (classifier (分类器)的标签不能被标记为正的样本为负的能力),并且 [recall](https://en.wikipedia.org/wiki/Precision_and_recall#Recall) 是 classifier (分类器)查找所有 positive samples (正样本)的能力。
[F-measure](https://en.wikipedia.org/wiki/F1_score) (![F_\beta](https://box.kancloud.cn/aebb383fddf232175c9166b889cb08d2_20x18.jpg) 和 ![F_1](https://box.kancloud.cn/fa296a713faafbad8659a3b82cc39f07_18x16.jpg) measures) 可以解释为 precision (精度)和 recall (召回)的 weighted harmonic mean (加权调和平均值)。 ![F_\beta](https://box.kancloud.cn/aebb383fddf232175c9166b889cb08d2_20x18.jpg) measure 值达到其最佳值 1 ,其最差分数为 0 。与 ![\beta = 1](https://box.kancloud.cn/5aec3b14495f3f06c786a0fda0ba733f_43x16.jpg), ![F_\beta](https://box.kancloud.cn/aebb383fddf232175c9166b889cb08d2_20x18.jpg) 和 ![F_1](https://box.kancloud.cn/fa296a713faafbad8659a3b82cc39f07_18x16.jpg) 是等价的, recall (召回)和 precision (精度)同样重要。
[`precision_recall_curve`](generated/sklearn.metrics.precision_recall_curve.html#sklearn.metrics.precision_recall_curve "sklearn.metrics.precision_recall_curve") 通过改变 decision threshold (决策阈值)从 ground truth label (被标记的真实数据标签) 和 score given by the classifier (分类器给出的分数)计算 precision-recall curve (精确召回曲线)。
[`average_precision_score`](generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score "sklearn.metrics.average_precision_score") 函数根据 prediction scores (预测分数)计算出 average precision (AP)(平均精度)。该分数对应于 precision-recall curve (精确召回曲线)下的面积。该值在 0 和 1 之间,并且越高越好。通过 random predictions (随机预测), AP 是 fraction of positive samples (正样本的分数)。
几个函数可以让您 analyze the precision (分析精度),recall(召回) 和 F-measures 得分:
[`average_precision_score`](generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score "sklearn.metrics.average_precision_score")(y\_true, y\_score\[, …\])Compute average precision (AP) from prediction scores[`f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score")(y\_true, y\_pred\[, labels, …\])Compute the F1 score, also known as balanced F-score or F-measure[`fbeta_score`](generated/sklearn.metrics.fbeta_score.html#sklearn.metrics.fbeta_score "sklearn.metrics.fbeta_score")(y\_true, y\_pred, beta\[, labels, …\])Compute the F-beta score[`precision_recall_curve`](generated/sklearn.metrics.precision_recall_curve.html#sklearn.metrics.precision_recall_curve "sklearn.metrics.precision_recall_curve")(y\_true, probas\_pred)Compute precision-recall pairs for different probability thresholds[`precision_recall_fscore_support`](generated/sklearn.metrics.precision_recall_fscore_support.html#sklearn.metrics.precision_recall_fscore_support "sklearn.metrics.precision_recall_fscore_support")(y\_true, y\_pred)Compute precision, recall, F-measure and support for each class[`precision_score`](generated/sklearn.metrics.precision_score.html#sklearn.metrics.precision_score "sklearn.metrics.precision_score")(y\_true, y\_pred\[, labels, …\])Compute the precision[`recall_score`](generated/sklearn.metrics.recall_score.html#sklearn.metrics.recall_score "sklearn.metrics.recall_score")(y\_true, y\_pred\[, labels, …\])Compute the recall请注意,[`precision_recall_curve`](generated/sklearn.metrics.precision_recall_curve.html#sklearn.metrics.precision_recall_curve "sklearn.metrics.precision_recall_curve") 函数仅限于 binary case (二分情况)。 [`average_precision_score`](generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score "sklearn.metrics.average_precision_score") 函数只适用于 binary classification and multilabel indicator format (二分类和多标签指示器格式)。
示例:
- 参阅 [Classification of text documents using sparse features](../auto_examples/text/document_classification_20newsgroups.html#sphx-glr-auto-examples-text-document-classification-20newsgroups-py)例如 [`f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score") 用于分类文本文档的用法。
- 参阅 [Parameter estimation using grid search with cross-validation](../auto_examples/model_selection/plot_grid_search_digits.html#sphx-glr-auto-examples-model-selection-plot-grid-search-digits-py)例如 [`precision_score`](generated/sklearn.metrics.precision_score.html#sklearn.metrics.precision_score "sklearn.metrics.precision_score") 和 [`recall_score`](generated/sklearn.metrics.recall_score.html#sklearn.metrics.recall_score "sklearn.metrics.recall_score") 用于 using grid search with nested cross-validation (使用嵌套交叉验证的网格搜索)来估计参数。
- 参阅 [Precision-Recall](../auto_examples/model_selection/plot_precision_recall.html#sphx-glr-auto-examples-model-selection-plot-precision-recall-py)例如 [`precision_recall_curve`](generated/sklearn.metrics.precision_recall_curve.html#sklearn.metrics.precision_recall_curve "sklearn.metrics.precision_recall_curve") 用于 evaluate classifier output quality(评估分类器输出质量)。
#### 3.3.2.8.1. 二分类
在二分类任务中,术语 ‘’positive(正)’’ 和 ‘’negative(负)’’ 是指 classifier’s prediction (分类器的预测),术语 ‘’true(真)’’ 和 ‘’false(假)’’ 是指该预测是否对应于 external judgment (外部判断)(有时被称为 ‘’observation(观测值)’‘)。给出这些定义,我们可以指定下表:
Actual class (observation)Predicted class (expectation)tp (true positive) Correct resultfp (false positive) Unexpected resultfn (false negative) Missing resulttn (true negative) Correct absence of result在这种情况下,我们可以定义 precision(精度), recall(召回) 和 F-measure 的概念:
![\text{precision} = \frac{tp}{tp + fp},](https://box.kancloud.cn/35516d0d1bf6b5398e4e880de7b90125_157x41.jpg)
![\text{recall} = \frac{tp}{tp + fn},](https://box.kancloud.cn/b2177648bc64936d464c43526f394ac3_132x41.jpg)
![F_\beta = (1 + \beta^2) \frac{\text{precision} \times \text{recall}}{\beta^2 \text{precision} + \text{recall}}.](https://box.kancloud.cn/ab1e348b2b224dec4bd59566d68baada_266x42.jpg)
以下是 binary classification (二分类)中的一些小例子:
```
>>> from sklearn import metrics
>>> y_pred = [0, 1, 0, 0]
>>> y_true = [0, 1, 0, 1]
>>> metrics.precision_score(y_true, y_pred)
1.0
>>> metrics.recall_score(y_true, y_pred)
0.5
>>> metrics.f1_score(y_true, y_pred)
0.66...
>>> metrics.fbeta_score(y_true, y_pred, beta=0.5)
0.83...
>>> metrics.fbeta_score(y_true, y_pred, beta=1)
0.66...
>>> metrics.fbeta_score(y_true, y_pred, beta=2)
0.55...
>>> metrics.precision_recall_fscore_support(y_true, y_pred, beta=0.5)
(array([ 0.66..., 1. ]), array([ 1. , 0.5]), array([ 0.71..., 0.83...]), array([2, 2]...))
>>> import numpy as np
>>> from sklearn.metrics import precision_recall_curve
>>> from sklearn.metrics import average_precision_score
>>> y_true = np.array([0, 0, 1, 1])
>>> y_scores = np.array([0.1, 0.4, 0.35, 0.8])
>>> precision, recall, threshold = precision_recall_curve(y_true, y_scores)
>>> precision
array([ 0.66..., 0.5 , 1. , 1. ])
>>> recall
array([ 1. , 0.5, 0.5, 0. ])
>>> threshold
array([ 0.35, 0.4 , 0.8 ])
>>> average_precision_score(y_true, y_scores)
0.83...
```
#### 3.3.2.8.2. 多类和多标签分类
在 multiclass and multilabel classification task(多类和多标签分类任务)中,precision(精度), recall(召回), and F-measures 的概念可以独立地应用于每个标签。 有以下几种方法 combine results across labels (将结果跨越标签组合),由 `average` 参数指定为 [`average_precision_score`](generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score "sklearn.metrics.average_precision_score") (仅用于 multilabel), [`f1_score`](generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score "sklearn.metrics.f1_score"), [`fbeta_score`](generated/sklearn.metrics.fbeta_score.html#sklearn.metrics.fbeta_score "sklearn.metrics.fbeta_score"), [`precision_recall_fscore_support`](generated/sklearn.metrics.precision_recall_fscore_support.html#sklearn.metrics.precision_recall_fscore_support "sklearn.metrics.precision_recall_fscore_support"), [`precision_score`](generated/sklearn.metrics.precision_score.html#sklearn.metrics.precision_score "sklearn.metrics.precision_score") 和 [`recall_score`](generated/sklearn.metrics.recall_score.html#sklearn.metrics.recall_score "sklearn.metrics.recall_score") 函数,如上 [above](#average) 所述。请注意,对于在包含所有标签的多类设置中进行 “micro”-averaging (”微”平均),将产生相等的 precision(精度), recall(召回)和 ![F](https://box.kancloud.cn/3108b1c6b81a7a313256a1d3d96339cb_14x12.jpg) ,而 “weighted(加权)” averaging(平均)可能会产生 precision(精度)和 recall(召回)之间的 F-score 。
为了使这一点更加明确,请考虑以下 notation (符号):
- ![y](https://box.kancloud.cn/0255a09d3dccb9843dcf063bbeec303f_9x12.jpg)*predicted(预测)* ![(sample, label)](https://box.kancloud.cn/9774608efebee2864b745e6385a89457_113x18.jpg) 对
- ![\hat{y}](https://box.kancloud.cn/277d247a09c0ccb4240fe50a4806934e_9x17.jpg)*true(真)* ![(sample, label)](https://box.kancloud.cn/9774608efebee2864b745e6385a89457_113x18.jpg) 对
- ![L](https://box.kancloud.cn/932e52dfeb15d15287c07f0b899113b1_12x12.jpg) labels 集合
- ![S](https://box.kancloud.cn/3654ba253cea374c1cf48d1877e4bf6c_12x12.jpg) samples 集合
- ![y_s](https://box.kancloud.cn/dd31a23b6c2dbeefc1f85acd7e3a8d50_15x12.jpg)![y](https://box.kancloud.cn/0255a09d3dccb9843dcf063bbeec303f_9x12.jpg) 的子集与样本 ![s](https://box.kancloud.cn/2f503365de225730371850cf4efa70b8_8x8.jpg), 即 ![y_s := \left\{(s', l) \in y | s' = s\right\}](https://box.kancloud.cn/934d26a5b6b4d4a1b66208825f7e09d3_184x19.jpg)
- ![y_l](https://box.kancloud.cn/0096fa3ea351944ac5ecbbc83435d4a3_13x12.jpg)![y](https://box.kancloud.cn/0255a09d3dccb9843dcf063bbeec303f_9x12.jpg) 的子集与 label ![l](https://box.kancloud.cn/aee0e789bbfa3644393125081b3c7fe3_5x13.jpg)
- 类似的, ![\hat{y}_s](https://box.kancloud.cn/60703a24106632dea3e2367d6ac26998_15x17.jpg) 和 ![\hat{y}_l](https://box.kancloud.cn/dc92a9e5cfa3bc302fce1d30f9cb21db_13x17.jpg) 是 ![\hat{y}](https://box.kancloud.cn/277d247a09c0ccb4240fe50a4806934e_9x17.jpg) 的子集
- ![P(A, B) := \frac{\left| A \cap B \right|}{\left|A\right|}](https://box.kancloud.cn/b171fae44f547082158d7064c6b1ed86_131x27.jpg)
- ![R(A, B) := \frac{\left| A \cap B \right|}{\left|B\right|}](https://box.kancloud.cn/3c415239cb39c5c44ed525902b44d7f3_131x27.jpg)(Conventions (公约)在处理 ![B = \emptyset](https://box.kancloud.cn/25c189db7d8e2e08c207908cfca8b154_48x17.jpg) 有所不同; 这个实现使用 ![R(A, B):=0](https://box.kancloud.cn/04d7cdbf733a3efb5f1b97606339878e_102x18.jpg), 与 ![P](https://box.kancloud.cn/08277e04611b27b30b29f99ba0830d27_14x12.jpg) 类似.)
- ![F_\beta(A, B) := \left(1 + \beta^2\right) \frac{P(A, B) \times R(A, B)}{\beta^2 P(A, B) + R(A, B)}](https://box.kancloud.cn/58682d26e2c0f321451be5c51a71b083_279x29.jpg)
然后将 metrics (指标)定义为:
`average`PrecisionRecallF\_beta`"micro"`![P(y, \hat{y})](https://box.kancloud.cn/31bdafc622943bf7bbf046a585f234a1_54x18.jpg)![R(y, \hat{y})](https://box.kancloud.cn/77291189d1c126bde670c7573836df8a_54x18.jpg)![F_\beta(y, \hat{y})](https://box.kancloud.cn/3e1bb97dd8c354b59ccd006fbbc0ffbe_60x20.jpg)`"samples"`![\frac{1}{\left|S\right|} \sum_{s \in S} P(y_s, \hat{y}_s)](https://box.kancloud.cn/ffc47952b16d9ce0fc5a2a1072fc2cc1_133x26.jpg)![\frac{1}{\left|S\right|} \sum_{s \in S} R(y_s, \hat{y}_s)](https://box.kancloud.cn/d950c91e223d0880b9c067aa2c613191_133x26.jpg)![\frac{1}{\left|S\right|} \sum_{s \in S} F_\beta(y_s, \hat{y}_s)](https://box.kancloud.cn/6bc2f50f2a4574345196a579e2e75679_139x26.jpg)`"macro"`![\frac{1}{\left|L\right|} \sum_{l \in L} P(y_l, \hat{y}_l)](https://box.kancloud.cn/9dbcb3265dcb4917f65feaa4d657ee11_127x26.jpg)![\frac{1}{\left|L\right|} \sum_{l \in L} R(y_l, \hat{y}_l)](https://box.kancloud.cn/12cf8b4c87020129b9c559a18090aeaa_127x26.jpg)![\frac{1}{\left|L\right|} \sum_{l \in L} F_\beta(y_l, \hat{y}_l)](https://box.kancloud.cn/4275f37b87c67faea17b7717f1e212be_134x26.jpg)`"weighted"`![\frac{1}{\sum_{l \in L} \left|\hat{y}_l\right|} \sum_{l \in L} \left|\hat{y}_l\right| P(y_l, \hat{y}_l)](https://box.kancloud.cn/15564b3dbe432feb2db5382696528958_189x28.jpg)![\frac{1}{\sum_{l \in L} \left|\hat{y}_l\right|} \sum_{l \in L} \left|\hat{y}_l\right| R(y_l, \hat{y}_l)](https://box.kancloud.cn/01627e11e94d2a84370fa4059fe7ad50_189x28.jpg)![\frac{1}{\sum_{l \in L} \left|\hat{y}_l\right|} \sum_{l \in L} \left|\hat{y}_l\right| F_\beta(y_l, \hat{y}_l)](https://box.kancloud.cn/de9bc098c0220fc46108d35a282ad6a7_196x28.jpg)`None`![\langle P(y_l, \hat{y}_l) | l \in L \rangle](https://box.kancloud.cn/794db845db45cb15c4b585eb2a04cf41_119x19.jpg)![\langle R(y_l, \hat{y}_l) | l \in L \rangle](https://box.kancloud.cn/abe1b0edf7b287e3822bc59dfeead69c_119x19.jpg)![\langle F_\beta(y_l, \hat{y}_l) | l \in L \rangle](https://box.kancloud.cn/bfc9917fb837e2d72b3407d605eb6280_126x20.jpg)
```
>>> from sklearn import metrics
>>> y_true = [0, 1, 2, 0, 1, 2]
>>> y_pred = [0, 2, 1, 0, 0, 1]
>>> metrics.precision_score(y_true, y_pred, average='macro')
0.22...
>>> metrics.recall_score(y_true, y_pred, average='micro')
...
0.33...
>>> metrics.f1_score(y_true, y_pred, average='weighted')
0.26...
>>> metrics.fbeta_score(y_true, y_pred, average='macro', beta=0.5)
0.23...
>>> metrics.precision_recall_fscore_support(y_true, y_pred, beta=0.5, average=None)
...
(array([ 0.66..., 0. , 0. ]), array([ 1., 0., 0.]), array([ 0.71..., 0. , 0. ]), array([2, 2, 2]...))
```
For multiclass classification with a “negative class”, it is possible to exclude some labels:
```
>>> metrics.recall_score(y_true, y_pred, labels=[1, 2], average='micro')
... # excluding 0, no labels were correctly recalled
0.0
```
Similarly, labels not present in the data sample may be accounted for in macro-averaging.
```
>>> metrics.precision_score(y_true, y_pred, labels=[0, 1, 2, 3], average='macro')
...
0.166...
```
### 3.3.2.9. Hinge loss
[`hinge_loss`](generated/sklearn.metrics.hinge_loss.html#sklearn.metrics.hinge_loss "sklearn.metrics.hinge_loss") 函数使用 [hinge loss](https://en.wikipedia.org/wiki/Hinge_loss) 计算模型和数据之间的 average distance (平均距离),这是一种只考虑 prediction errors (预测误差)的 one-sided metric (单向指标)。(Hinge loss 用于最大边界分类器,如支持向量机)
如果标签用 +1 和 -1 编码,则 ![y](https://box.kancloud.cn/0255a09d3dccb9843dcf063bbeec303f_9x12.jpg): 是真实值,并且 ![w](https://box.kancloud.cn/0635104a899a1b8951f0b8da2816a950_13x8.jpg) 是由 `decision_function` 输出的 predicted decisions (预测决策),则 hinge loss 定义为:
![L_\text{Hinge}(y, w) = \max\left\{1 - wy, 0\right\} = \left|1 - wy\right|_+](https://box.kancloud.cn/7945964990e3ffe30480d4126a51c97b_338x21.jpg)
如果有两个以上的标签, [`hinge_loss`](generated/sklearn.metrics.hinge_loss.html#sklearn.metrics.hinge_loss "sklearn.metrics.hinge_loss") 由于 Crammer & Singer 而使用了 multiclass variant (多类型变体)。 [Here](http://jmlr.csail.mit.edu/papers/volume2/crammer01a/crammer01a.pdf) 是描述它的论文。
如果 ![y_w](https://box.kancloud.cn/f59a5c945b996a714d788df68ff437c9_18x12.jpg) 是真实标签的 predicted decision (预测决策),并且 ![y_t](https://box.kancloud.cn/9323f4f7e5c7cc18a7f74854ba96f757_14x12.jpg) 是所有其他标签的预测决策的最大值,其中预测决策由 decision function (决策函数)输出,则 multiclass hinge loss 定义如下:
![L_\text{Hinge}(y_w, y_t) = \max\left\{1 + y_t - y_w, 0\right\}](https://box.kancloud.cn/ad08bde480d9c85f362ae931509e9f47_282x20.jpg)
这里是一个小例子,演示了在 binary class (二类)问题中使用了具有 svm classifier (svm 的分类器)的 [`hinge_loss`](generated/sklearn.metrics.hinge_loss.html#sklearn.metrics.hinge_loss "sklearn.metrics.hinge_loss") 函数:
```
>>> from sklearn import svm
>>> from sklearn.metrics import hinge_loss
>>> X = [[0], [1]]
>>> y = [-1, 1]
>>> est = svm.LinearSVC(random_state=0)
>>> est.fit(X, y)
LinearSVC(C=1.0, class_weight=None, dual=True, fit_intercept=True,
intercept_scaling=1, loss='squared_hinge', max_iter=1000,
multi_class='ovr', penalty='l2', random_state=0, tol=0.0001,
verbose=0)
>>> pred_decision = est.decision_function([[-2], [3], [0.5]])
>>> pred_decision
array([-2.18..., 2.36..., 0.09...])
>>> hinge_loss([-1, 1, 1], pred_decision)
0.3...
```
这里是一个示例,演示了在 multiclass problem (多类问题)中使用了具有 svm 分类器的 [`hinge_loss`](generated/sklearn.metrics.hinge_loss.html#sklearn.metrics.hinge_loss "sklearn.metrics.hinge_loss") 函数:
```
>>> X = np.array([[0], [1], [2], [3]])
>>> Y = np.array([0, 1, 2, 3])
>>> labels = np.array([0, 1, 2, 3])
>>> est = svm.LinearSVC()
>>> est.fit(X, Y)
LinearSVC(C=1.0, class_weight=None, dual=True, fit_intercept=True,
intercept_scaling=1, loss='squared_hinge', max_iter=1000,
multi_class='ovr', penalty='l2', random_state=None, tol=0.0001,
verbose=0)
>>> pred_decision = est.decision_function([[-1], [2], [3]])
>>> y_true = [0, 2, 3]
>>> hinge_loss(y_true, pred_decision, labels)
0.56...
```
### 3.3.2.10. Log 损失
Log loss,又被称为 logistic regression loss(logistic 回归损失)或者 cross-entropy loss(交叉熵损失) 定义在 probability estimates (概率估计)。它通常用于 (multinomial) logistic regression ((多项式)logistic 回归)和 neural networks (神经网络)以及 expectation-maximization (期望最大化)的一些变体中,并且可用于评估分类器的 probability outputs (概率输出)(`predict_proba`)而不是其 discrete predictions (离散预测)。
对于具有真实标签 ![y \in \{0,1\}](https://box.kancloud.cn/75f499527495bb78ae7059f7b06dace1_75x19.jpg) 的 binary classification (二分类)和 probability estimate (概率估计) ![p = \operatorname{Pr}(y = 1)](https://box.kancloud.cn/06010300e140579db2ce53cdb9684b17_109x18.jpg), 每个样本的 log loss 是给定的分类器的 negative log-likelihood 真正的标签:
![L_{\log}(y, p) = -\log \operatorname{Pr}(y|p) = -(y \log (p) + (1 - y) \log (1 - p))](https://box.kancloud.cn/839a8713ade921d22bc68371c07ec19f_460x20.jpg)
这扩展到 multiclass case (多类案例)如下。 让一组样本的真实标签被编码为 1-of-K binary indicator matrix ![Y](https://box.kancloud.cn/f016166d5c910a983468b95f8fb3a11e_14x12.jpg), 即 如果样本 ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg) 具有取自一组 ![K](https://box.kancloud.cn/cdbf8b35090576059f2b14a851c73b86_16x12.jpg) 个标签的标签 ![k](https://box.kancloud.cn/300675e73ace6bf4c352cfbb633f0199_9x13.jpg) ,则 ![y_{i,k} = 1](https://box.kancloud.cn/a09df6dff8352bafdef46a2689937570_57x18.jpg) 。令 ![P](https://box.kancloud.cn/08277e04611b27b30b29f99ba0830d27_14x12.jpg) 为 matrix of probability estimates (概率估计矩阵), ![p_{i,k} = \operatorname{Pr}(t_{i,k} = 1)](https://box.kancloud.cn/58f2753deaabdbe91357330520f2273f_137x20.jpg) 。那么整套的 log loss 就是
![L_{\log}(Y, P) = -\log \operatorname{Pr}(Y|P) = - \frac{1}{N} \sum_{i=0}^{N-1} \sum_{k=0}^{K-1} y_{i,k} \log p_{i,k}](https://box.kancloud.cn/2448c2f352a06ded30953a56b537885c_412x55.jpg)
为了看这这里如何 generalizes (推广)上面给出的 binary log loss (二分 log loss),请注意,在 binary case (二分情况下),![p_{i,0} = 1 - p_{i,1}](https://box.kancloud.cn/3e09dbd17aed7c847240eec39536209b_104x18.jpg) 和 ![y_{i,0} = 1 - y_{i,1}](https://box.kancloud.cn/21012ac5279bc983e8559254ba107797_102x18.jpg) ,因此扩展 ![y_{i,k} \in \{0,1\}](https://box.kancloud.cn/ffbfd10600b986e967aa461d09c77741_90x20.jpg) 的 inner sum (内部和),给出 binary log loss (二分 log loss)。
[`log_loss`](generated/sklearn.metrics.log_loss.html#sklearn.metrics.log_loss "sklearn.metrics.log_loss") 函数计算出一个 a list of ground-truth labels (已标注的真实数据的标签的列表)和一个 probability matrix (概率矩阵) 的 log loss,由 estimator (估计器)的 `predict_proba` 方法返回。
```
>>> from sklearn.metrics import log_loss
>>> y_true = [0, 0, 1, 1]
>>> y_pred = [[.9, .1], [.8, .2], [.3, .7], [.01, .99]]
>>> log_loss(y_true, y_pred)
0.1738...
```
`y_pred` 中的第一个 `[.9, .1]` 表示第一个样本具有标签 0 的 90% 概率。log loss 是非负数。
### 3.3.2.11. 马修斯相关系数
[`matthews_corrcoef`](generated/sklearn.metrics.matthews_corrcoef.html#sklearn.metrics.matthews_corrcoef "sklearn.metrics.matthews_corrcoef") 函数用于计算 binary classes (二分类)的 [Matthew’s correlation coefficient (MCC)](https://en.wikipedia.org/wiki/Matthews_correlation_coefficient) 引用自 Wikipedia:
> “Matthews correlation coefficient(马修斯相关系数)用于机器学习,作为 binary (two-class) classifications (二分类)分类质量的度量。它考虑到 true and false positives and negatives (真和假的 positives 和 negatives),通常被认为是可以使用的 balanced measure(平衡措施),即使 classes are of very different sizes (类别大小不同)。MCC 本质上是 -1 和 +1 之间的相关系数值。系数 +1 表示完美预测,0 表示平均随机预测, -1 表示反向预测。statistic (统计量)也称为 phi coefficient (phi)系数。”
在 binary (two-class) (二分类)情况下,![tp](https://box.kancloud.cn/3bc2f4b08dfd9e12a871651663b77f05_15x16.jpg), ![tn](https://box.kancloud.cn/1a61c0b8663984aa130af62f0355004b_17x12.jpg), ![fp](https://box.kancloud.cn/3a735bc9d1d0d0a559cc8b668f8a0a8d_20x16.jpg) 和 ![fn](https://box.kancloud.cn/c58d49e179940b9451fad80ba9889407_22x16.jpg) 分别是 true positives, true negatives, false positives 和 false negatives 的数量,MCC 定义为
![MCC = \frac{tp \times tn - fp \times fn}{\sqrt{(tp + fp)(tp + fn)(tn + fp)(tn + fn)}}.](https://box.kancloud.cn/8e130557bd3736e69b6491d0a8918690_389x46.jpg)
在 multiclass case (多类的情况)下, Matthews correlation coefficient(马修斯相关系数) 可以根据 ![K](https://box.kancloud.cn/cdbf8b35090576059f2b14a851c73b86_16x12.jpg) classes (类)的 [`confusion_matrix`](generated/sklearn.metrics.confusion_matrix.html#sklearn.metrics.confusion_matrix "sklearn.metrics.confusion_matrix") ![C](https://box.kancloud.cn/95378f1036b3ba9a15a5f33f8521b6f2_14x12.jpg) 定义 [defined](http://rk.kvl.dk/introduction/index.html) 。为了简化定义,考虑以下中间变量:
- ![t_k=\sum_{i}^{K} C_{ik}](https://box.kancloud.cn/412b32fa1a93162b2cd88c462229851b_97x24.jpg) 真正发生了 ![k](https://box.kancloud.cn/300675e73ace6bf4c352cfbb633f0199_9x13.jpg) 类的次数,
- ![p_k=\sum_{i}^{K} C_{ki}](https://box.kancloud.cn/43a1b032ee1bc365425add998e6adef1_101x24.jpg)![k](https://box.kancloud.cn/300675e73ace6bf4c352cfbb633f0199_9x13.jpg) 类被预测的次数,
- ![c=\sum_{k}^{K} C_{kk}](https://box.kancloud.cn/8bdbb2f7a55b60deaaac83f7ec2f5b44_94x24.jpg) 正确预测的样本总数,
- ![s=\sum_{i}^{K} \sum_{j}^{K} C_{ij}](https://box.kancloud.cn/5c840b3b50f7688d5fd3d5bcd10341b6_125x27.jpg) 样本总数.
然后 multiclass MCC 定义为:
![MCC = \frac{ c \times s - \sum_{k}^{K} p_k \times t_k }{\sqrt{ (s^2 - \sum_{k}^{K} p_k^2) \times (s^2 - \sum_{k}^{K} t_k^2) }}](https://box.kancloud.cn/45f9f42d0f6da4e0516c11514d394a0f_319x63.jpg)
当有两个以上的标签时, MCC 的值将不再在 -1 和 +1 之间。相反,根据已经标注的真实数据的数量和分布情况,最小值将介于 -1 和 0 之间。最大值始终为 +1 。
这是一个小例子,说明了使用 [`matthews_corrcoef`](generated/sklearn.metrics.matthews_corrcoef.html#sklearn.metrics.matthews_corrcoef "sklearn.metrics.matthews_corrcoef") 函数:
```
>>> from sklearn.metrics import matthews_corrcoef
>>> y_true = [+1, +1, +1, -1]
>>> y_pred = [+1, -1, +1, +1]
>>> matthews_corrcoef(y_true, y_pred)
-0.33...
```
### 3.3.2.12. Receiver operating characteristic (ROC)
函数 [`roc_curve`](generated/sklearn.metrics.roc_curve.html#sklearn.metrics.roc_curve "sklearn.metrics.roc_curve") 计算 [receiver operating characteristic curve, or ROC curve](https://en.wikipedia.org/wiki/Receiver_operating_characteristic). 引用 Wikipedia :
> “A receiver operating characteristic (ROC), 或者简单的 ROC 曲线,是一个图形图,说明了 binary classifier (二分分类器)系统的性能,因为 discrimination threshold (鉴别阈值)是变化的。它是通过在不同的阈值设置下,从 true positives out of the positives (TPR = true positive 比例) 与 false positives out of the negatives (FPR = false positive 比例) 绘制 true positive 的比例来创建的。 TPR 也称为 sensitivity(灵敏度),FPR 是减去 specificity(特异性) 或 true negative 比例。”
该函数需要真正的 binar value (二分值)和 target scores(目标分数),这可以是 positive class 的 probability estimates (概率估计),confidence values(置信度值)或 binary decisions(二分决策)。 这是一个如何使用 [`roc_curve`](generated/sklearn.metrics.roc_curve.html#sklearn.metrics.roc_curve "sklearn.metrics.roc_curve") 函数的小例子:
```
>>> import numpy as np
>>> from sklearn.metrics import roc_curve
>>> y = np.array([1, 1, 2, 2])
>>> scores = np.array([0.1, 0.4, 0.35, 0.8])
>>> fpr, tpr, thresholds = roc_curve(y, scores, pos_label=2)
>>> fpr
array([ 0. , 0.5, 0.5, 1. ])
>>> tpr
array([ 0.5, 0.5, 1. , 1. ])
>>> thresholds
array([ 0.8 , 0.4 , 0.35, 0.1 ])
```
该图显示了这样的 ROC 曲线的示例:
[![http://sklearn.apachecn.org/cn/0.19.0/_images/sphx_glr_plot_roc_0011.png](https://box.kancloud.cn/5c2bf5f5aef5777e6f29a985946a7461_566x424.jpg)](../auto_examples/model_selection/plot_roc.html)[`roc_auc_score`](generated/sklearn.metrics.roc_auc_score.html#sklearn.metrics.roc_auc_score "sklearn.metrics.roc_auc_score") 函数计算 receiver operating characteristic (ROC) 曲线下的面积,也由 AUC 和 AUROC 表示。通过计算 roc 曲线下的面积,曲线信息总结为一个数字。 有关更多的信息,请参阅 [Wikipedia article on AUC](https://en.wikipedia.org/wiki/Receiver_operating_characteristic#Area_under_the_curve) .
```
>>> import numpy as np
>>> from sklearn.metrics import roc_auc_score
>>> y_true = np.array([0, 0, 1, 1])
>>> y_scores = np.array([0.1, 0.4, 0.35, 0.8])
>>> roc_auc_score(y_true, y_scores)
0.75
```
在 multi-label classification (多标签分类)中, [`roc_auc_score`](generated/sklearn.metrics.roc_auc_score.html#sklearn.metrics.roc_auc_score "sklearn.metrics.roc_auc_score") 函数通过在标签上进行平均来扩展 [above](#average) .
与诸如 subset accuracy (子集精确度),Hamming loss(汉明损失)或 F1 score 的 metrics(指标)相比, ROC 不需要优化每个标签的阈值。[`roc_auc_score`](generated/sklearn.metrics.roc_auc_score.html#sklearn.metrics.roc_auc_score "sklearn.metrics.roc_auc_score") 函数也可以用于 multi-class classification (多类分类),如果预测的输出被 binarized (二分化)。
[![http://sklearn.apachecn.org/cn/0.19.0/_images/sphx_glr_plot_roc_0021.png](https://box.kancloud.cn/559d19a1c834cd7073eeafc432137492_566x424.jpg)](../auto_examples/model_selection/plot_roc.html)示例:
- 参阅 [Receiver Operating Characteristic (ROC)](../auto_examples/model_selection/plot_roc.html#sphx-glr-auto-examples-model-selection-plot-roc-py)例如使用 ROC 来评估分类器输出的质量。
- 参阅 [Receiver Operating Characteristic (ROC) with cross validation](../auto_examples/model_selection/plot_roc_crossval.html#sphx-glr-auto-examples-model-selection-plot-roc-crossval-py)例如使用 ROC 来评估分类器输出质量,使用 cross-validation (交叉验证)。
- 参阅 [Species distribution modeling](../auto_examples/applications/plot_species_distribution_modeling.html#sphx-glr-auto-examples-applications-plot-species-distribution-modeling-py)例如使用 ROC 来 model species distribution 模拟物种分布。
### 3.3.2.13. 零一损失
[`zero_one_loss`](generated/sklearn.metrics.zero_one_loss.html#sklearn.metrics.zero_one_loss "sklearn.metrics.zero_one_loss") 函数通过 ![n_{\text{samples}}](https://box.kancloud.cn/70c64e65bf204369a0c85d76a39ecf38_55x14.jpg) 计算 0-1 classification loss (![L_{0-1}](https://box.kancloud.cn/57bcb4509034930f7cda2cb6e33c49ef_34x16.jpg)) 的 sum (和)或 average (平均值)。默认情况下,函数在样本上 normalizes (标准化)。要获得 ![L_{0-1}](https://box.kancloud.cn/57bcb4509034930f7cda2cb6e33c49ef_34x16.jpg) 的总和,将 `normalize` 设置为 `False`。
在 multilabel classification (多标签分类)中,如果零标签与标签严格匹配,则 [`zero_one_loss`](generated/sklearn.metrics.zero_one_loss.html#sklearn.metrics.zero_one_loss "sklearn.metrics.zero_one_loss") 将一个子集作为一个子集,如果有任何错误,则为零。默认情况下,函数返回不完全预测子集的百分比。为了得到这样的子集的计数,将 `normalize` 设置为 `False` 。
如果 ![\hat{y}_i](https://box.kancloud.cn/a71dac1a33f465fa5197c33da8720585_13x17.jpg) 是第 ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg) 个样本的预测值,![y_i](https://box.kancloud.cn/e79627211612bc56c4f7d926a93fbe8d_13x12.jpg) 是相应的真实值,则 0-1 loss ![L_{0-1}](https://box.kancloud.cn/57bcb4509034930f7cda2cb6e33c49ef_34x16.jpg) 定义为:
![L_{0-1}(y_i, \hat{y}_i) = 1(\hat{y}_i \not= y_i)](https://box.kancloud.cn/acf23a7229cdc980ad48988870da5430_183x18.jpg)
其中 ![1(x)](https://box.kancloud.cn/89c489280f4c7e8dd0c3c421b0d4358b_31x18.jpg) 是 [indicator function](https://en.wikipedia.org/wiki/Indicator_function).
```
>>> from sklearn.metrics import zero_one_loss
>>> y_pred = [1, 2, 3, 4]
>>> y_true = [2, 2, 3, 4]
>>> zero_one_loss(y_true, y_pred)
0.25
>>> zero_one_loss(y_true, y_pred, normalize=False)
1
```
在具有 binary label indicators (二分标签指示符)的 multilabel (多标签)情况下,第一个标签集 \[0,1\] 有错误:
```
>>> zero_one_loss(np.array([[0, 1], [1, 1]]), np.ones((2, 2)))
0.5
>>> zero_one_loss(np.array([[0, 1], [1, 1]]), np.ones((2, 2)), normalize=False)
1
```
示例:
- 参阅 [Recursive feature elimination with cross-validation](../auto_examples/feature_selection/plot_rfe_with_cross_validation.html#sphx-glr-auto-examples-feature-selection-plot-rfe-with-cross-validation-py)例如 zero one loss 使用以通过 cross-validation (交叉验证)执行递归特征消除。
### 3.3.2.14. Brier 分数损失
[`brier_score_loss`](generated/sklearn.metrics.brier_score_loss.html#sklearn.metrics.brier_score_loss "sklearn.metrics.brier_score_loss") 函数计算二进制类的 [Brier 分数](https://en.wikipedia.org/wiki/Brier_score) 。引用维基百科:
> “Brier 分数是一个特有的分数函数,用于衡量概率预测的准确性。它适用于预测必须将概率分配给一组相互排斥的离散结果的任务。”
该函数返回的是 实际结果与可能结果 的预测概率之间均方差的得分。 实际结果必须为1或0(真或假),而实际结果的预测概率可以是0到1之间的值。
Brier 分数损失也在0到1之间,分数越低(均方差越小),预测越准确。它可以被认为是对一组概率预测的 “校准” 的度量。
![BS = \frac{1}{N} \sum_{t=1}^{N}(f_t - o_t)^2](https://box.kancloud.cn/6156e9795596f7573d196022ae976075_171x54.jpg)
其中: ![N](https://box.kancloud.cn/08e4021d29ea7df2884794031c0a46ab_16x12.jpg) 是预测的总数, ![f_t](https://box.kancloud.cn/7f898c2f14804992e3c9a9f1c0325b5d_14x16.jpg) 是实际结果 ![o_t](https://box.kancloud.cn/f26397db34f90ba646578829418a63bd_14x11.jpg) 的预测概率。
这是一个使用这个函数的小例子:
```
>>> import numpy as np
>>> from sklearn.metrics import brier_score_loss
>>> y_true = np.array([0, 1, 1, 0])
>>> y_true_categorical = np.array(["spam", "ham", "ham", "spam"])
>>> y_prob = np.array([0.1, 0.9, 0.8, 0.4])
>>> y_pred = np.array([0, 1, 1, 0])
>>> brier_score_loss(y_true, y_prob)
0.055
>>> brier_score_loss(y_true, 1-y_prob, pos_label=0)
0.055
>>> brier_score_loss(y_true_categorical, y_prob, pos_label="ham")
0.055
>>> brier_score_loss(y_true, y_prob > 0.5)
0.0
```
示例:
- 请参阅分类器的概率校准 [Probability calibration of classifiers](../auto_examples/calibration/plot_calibration.html#sphx-glr-auto-examples-calibration-plot-calibration-py) ,通过 Brier 分数损失使用示例 来执行分类器的概率校准。
参考文献:
- 1. Brier, [以概率表示的预测验证](http://docs.lib.noaa.gov/rescue/mwr/078/mwr-078-01-0001.pdf) , 月度天气评估78.1(1950)
## 3.3.3. 多标签排名指标
在多分类学习中,每个样本可以具有与其相关联的任何数量的真实标签。目标是给予高分,更好地评价真实标签。
### 3.3.3.1. 覆盖误差
[`coverage_error`](generated/sklearn.metrics.coverage_error.html#sklearn.metrics.coverage_error "sklearn.metrics.coverage_error") 函数计算必须包含在最终预测中的标签的平均数,以便预测所有真正的标签。 如果您想知道有多少 top 评分标签,您必须通过平均来预测,而不会丢失任何真正的标签,这很有用。 因此,此指标的最佳价值是真正标签的平均数量。
Note
我们的实现的分数比 Tsoumakas 等人在2010年的分数大1。 这扩展了它来处理一个具有0个真实标签实例的退化情况。
正式地,给定真实标签 ![y \in \left\{0, 1\right\}^{n_\text{samples} \times n_\text{labels}}](https://box.kancloud.cn/99496fa61a296fc96d962538f62d613b_168x20.jpg) 的二进制指示矩阵和与每个标签 ![\hat{f} \in \mathbb{R}^{n_\text{samples} \times n_\text{labels}}](https://box.kancloud.cn/830203b1fefe4cf7ee211f050f60a73c_138x21.jpg) 相关联的分数,覆盖范围被定义为
![coverage(y, \hat{f}) = \frac{1}{n_{\text{samples}}} \sum_{i=0}^{n_{\text{samples}} - 1} \max_{j:y_{ij} = 1} \text{rank}_{ij}](https://box.kancloud.cn/0fb6115e9f25fcfbd4bab2844e3c8dc6_355x55.jpg)
与 ![\text{rank}_{ij} = \left|\left\{k: \hat{f}_{ik} \geq \hat{f}_{ij} \right\}\right|](https://box.kancloud.cn/a9131e4c2e30430d83cde05ac878ac37_195x35.jpg) 。给定等级定义,通过给出将被分配给所有绑定值的最大等级, `y_scores` 中的关系会被破坏。
这是一个使用这个函数的小例子:
```
>>> import numpy as np
>>> from sklearn.metrics import coverage_error
>>> y_true = np.array([[1, 0, 0], [0, 0, 1]])
>>> y_score = np.array([[0.75, 0.5, 1], [1, 0.2, 0.1]])
>>> coverage_error(y_true, y_score)
2.5
```
### 3.3.3.2. 标签排名平均精度
[`label_ranking_average_precision_score`](generated/sklearn.metrics.label_ranking_average_precision_score.html#sklearn.metrics.label_ranking_average_precision_score "sklearn.metrics.label_ranking_average_precision_score") 函数实现标签排名平均精度(LRAP)。 该度量值与 [`average_precision_score`](generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score "sklearn.metrics.average_precision_score") 函数相关联,但是基于标签排名的概念,而不是精确度和召回。
标签排名平均精度(LRAP)是分配给每个样本的每个真实标签的平均值,真实对总标签与较低分数的比率。 如果能够为每个样本相关标签提供更好的排名,这个指标就会产生更好的分数。 获得的得分总是严格大于0,最佳值为1。如果每个样本只有一个相关标签,则标签排名平均精度等于 [平均倒数等级](https://en.wikipedia.org/wiki/Mean_reciprocal_rank) 。
正式地,给定真实标签 ![y \in \mathcal{R}^{n_\text{samples} \times n_\text{labels}}](https://box.kancloud.cn/f7d7f6dafae1465920170b71d2e8ed7e_139x17.jpg) 的二进制指示矩阵和与每个标签 ![\hat{f} \in \mathcal{R}^{n_\text{samples} \times n_\text{labels}}](https://box.kancloud.cn/2b3bb2acf5548a8c7844591812f9a5de_141x21.jpg) 相关联的得分,平均精度被定义为
![LRAP(y, \hat{f}) = \frac{1}{n_{\text{samples}}} \sum_{i=0}^{n_{\text{samples}} - 1} \frac{1}{|y_i|} \sum_{j:y_{ij} = 1} \frac{|\mathcal{L}_{ij}|}{\text{rank}_{ij}}](https://box.kancloud.cn/549da3b83be31beef9fe61381af749ce_373x59.jpg)
与 ![\mathcal{L}_{ij} = \left\{k: y_{ik} = 1, \hat{f}_{ik} \geq \hat{f}_{ij} \right\}](https://box.kancloud.cn/a16c29007cc06ecc6337c03518dcd0bd_223x34.jpg), ![\text{rank}_{ij} = \left|\left\{k: \hat{f}_{ik} \geq \hat{f}_{ij} \right\}\right|](https://box.kancloud.cn/a9131e4c2e30430d83cde05ac878ac37_195x35.jpg) 和 ![|\cdot|](https://box.kancloud.cn/d4590e3f2b051ecb5ee082686a3277fb_19x19.jpg) 是集合的 l0 范数或基数。
这是一个使用这个函数的小例子:
```
>>> import numpy as np
>>> from sklearn.metrics import label_ranking_average_precision_score
>>> y_true = np.array([[1, 0, 0], [0, 0, 1]])
>>> y_score = np.array([[0.75, 0.5, 1], [1, 0.2, 0.1]])
>>> label_ranking_average_precision_score(y_true, y_score)
0.416...
```
### 3.3.3.3. 排序损失
[`label_ranking_loss`](generated/sklearn.metrics.label_ranking_loss.html#sklearn.metrics.label_ranking_loss "sklearn.metrics.label_ranking_loss") 函数计算在样本上平均排序错误的标签对数量的排序损失,即真实标签的分数低于假标签,由虚假和真实标签的倒数加权。最低可实现的排名损失为零。
正式地,给定真相标签 ![y \in \left\{0, 1\right\}^{n_\text{samples} \times n_\text{labels}}](https://box.kancloud.cn/99496fa61a296fc96d962538f62d613b_168x20.jpg) 的二进制指示矩阵和与每个标签 ![\hat{f} \in \mathbb{R}^{n_\text{samples} \times n_\text{labels}}](https://box.kancloud.cn/830203b1fefe4cf7ee211f050f60a73c_138x21.jpg) 相关联的得分,排序损失被定义为
![\text{ranking\_loss}(y, \hat{f}) = \frac{1}{n_{\text{samples}}} \sum_{i=0}^{n_{\text{samples}} - 1} \frac{1}{|y_i|(n_\text{labels} - |y_i|)} \left|\left\{(k, l): \hat{f}_{ik} < \hat{f}_{il}, y_{ik} = 1, y_{il} = 0 \right\}\right|](https://box.kancloud.cn/0df989f2a7bdea33f129725711a67e84_566x44.jpg)
其中 ![|\cdot|](https://box.kancloud.cn/d4590e3f2b051ecb5ee082686a3277fb_19x19.jpg) 是 ![\ell_0](https://box.kancloud.cn/8f9f212fe4aff91bfceccf86970f3036_14x16.jpg) 范数或集合的基数。
这是一个使用这个函数的小例子:
```
>>> import numpy as np
>>> from sklearn.metrics import label_ranking_loss
>>> y_true = np.array([[1, 0, 0], [0, 0, 1]])
>>> y_score = np.array([[0.75, 0.5, 1], [1, 0.2, 0.1]])
>>> label_ranking_loss(y_true, y_score)
0.75...
>>> # With the following prediction, we have perfect and minimal loss
>>> y_score = np.array([[1.0, 0.1, 0.2], [0.1, 0.2, 0.9]])
>>> label_ranking_loss(y_true, y_score)
0.0
```
参考文献:
- Tsoumakas, G., Katakis, I., & Vlahavas, I. (2010). 挖掘多标签数据。在数据挖掘和知识发现手册(第667-685页)。美国 Springer.
## 3.3.4. 回归指标
该 [`sklearn.metrics`](classes.html#module-sklearn.metrics "sklearn.metrics") 模块实现了一些 loss, score 以及 utility 函数以测量 regression(回归)的性能. 其中一些已经被加强以处理多个输出的场景: [`mean_squared_error`](generated/sklearn.metrics.mean_squared_error.html#sklearn.metrics.mean_squared_error "sklearn.metrics.mean_squared_error"), [`mean_absolute_error`](generated/sklearn.metrics.mean_absolute_error.html#sklearn.metrics.mean_absolute_error "sklearn.metrics.mean_absolute_error"), [`explained_variance_score`](generated/sklearn.metrics.explained_variance_score.html#sklearn.metrics.explained_variance_score "sklearn.metrics.explained_variance_score") 和 [`r2_score`](generated/sklearn.metrics.r2_score.html#sklearn.metrics.r2_score "sklearn.metrics.r2_score").
这些函数有 `multioutput` 这样一个 keyword(关键的)参数, 它指定每一个目标的 score(得分)或 loss(损失)的平均值的方式. 默认是 `'uniform_average'`, 其指定了输出时一致的权重均值. 如果一个 `ndarray` 的 shape `(n_outputs,)` 被传递, 则其中的 entries(条目)将被解释为权重,并返回相应的加权平均值. 如果 `multioutput` 指定了 `'raw_values'` , 则所有未改变的部分 score(得分)或 loss(损失)将以 `(n_outputs,)` 形式的数组返回.
该 [`r2_score`](generated/sklearn.metrics.r2_score.html#sklearn.metrics.r2_score "sklearn.metrics.r2_score") 和 [`explained_variance_score`](generated/sklearn.metrics.explained_variance_score.html#sklearn.metrics.explained_variance_score "sklearn.metrics.explained_variance_score") 函数接受一个额外的值 `'variance_weighted'` 用于 `multioutput` 参数. 该选项通过相应目标变量的方差使得每个单独的 score 进行加权. 该设置量化了全局捕获的未缩放方差. 如果目标变量的大小不一样, 则该 score 更好地解释了较高的方差变量. `multioutput='variance_weighted'` 是 [`r2_score`](generated/sklearn.metrics.r2_score.html#sklearn.metrics.r2_score "sklearn.metrics.r2_score") 的默认值以向后兼容. 以后该值会被改成 `uniform_average`.
### 3.3.4.1. 解释方差得分
该 [`explained_variance_score`](generated/sklearn.metrics.explained_variance_score.html#sklearn.metrics.explained_variance_score "sklearn.metrics.explained_variance_score") 函数计算了 [explained variance regression score(解释的方差回归得分)](https://en.wikipedia.org/wiki/Explained_variation).
如果 ![\hat{y}](https://box.kancloud.cn/277d247a09c0ccb4240fe50a4806934e_9x17.jpg) 是预估的目标输出, ![y](https://box.kancloud.cn/0255a09d3dccb9843dcf063bbeec303f_9x12.jpg) 是相应(正确的)目标输出, 并且 ![Var](https://box.kancloud.cn/40c5a0b6c189fe9d7bb20b2bf4f5c652_32x12.jpg) is [方差](https://en.wikipedia.org/wiki/Variance), 标准差的平方, 那么解释的方差预估如下:
![\texttt{explained\_{}variance}(y, \hat{y}) = 1 - \frac{Var\{ y - \hat{y}\}}{Var\{y\}}](https://box.kancloud.cn/2bc78541e33cc29bff5b0a3f1f141401_356x44.jpg)
最好的得分是 1.0, 值越低越差.
下面是一下有关 [`explained_variance_score`](generated/sklearn.metrics.explained_variance_score.html#sklearn.metrics.explained_variance_score "sklearn.metrics.explained_variance_score") 函数使用的一些例子:
```
>>> from sklearn.metrics import explained_variance_score
>>> y_true = [3, -0.5, 2, 7]
>>> y_pred = [2.5, 0.0, 2, 8]
>>> explained_variance_score(y_true, y_pred)
0.957...
>>> y_true = [[0.5, 1], [-1, 1], [7, -6]]
>>> y_pred = [[0, 2], [-1, 2], [8, -5]]
>>> explained_variance_score(y_true, y_pred, multioutput='raw_values')
...
array([ 0.967..., 1. ])
>>> explained_variance_score(y_true, y_pred, multioutput=[0.3, 0.7])
...
0.990...
```
### 3.3.4.2. 平均绝对误差
该 [`mean_absolute_error`](generated/sklearn.metrics.mean_absolute_error.html#sklearn.metrics.mean_absolute_error "sklearn.metrics.mean_absolute_error") 函数计算了 [平均绝对误差](https://en.wikipedia.org/wiki/Mean_absolute_error), 一个对应绝对误差损失预期值或者 ![l1](https://box.kancloud.cn/3f2bf3b1e65ec3cf41a273465a2cb464_14x14.jpg)-norm 损失的风险度量.
如果 ![\hat{y}_i](https://box.kancloud.cn/a71dac1a33f465fa5197c33da8720585_13x17.jpg) 是 ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg)-th 样本的预测值, 并且 ![y_i](https://box.kancloud.cn/e79627211612bc56c4f7d926a93fbe8d_13x12.jpg) 是对应的真实值, 则平均绝对误差 (MAE) 预估的 ![n_{\text{samples}}](https://box.kancloud.cn/70c64e65bf204369a0c85d76a39ecf38_55x14.jpg) 定义如下
![\text{MAE}(y, \hat{y}) = \frac{1}{n_{\text{samples}}} \sum_{i=0}^{n_{\text{samples}}-1} \left| y_i - \hat{y}_i \right|.](https://box.kancloud.cn/237ebedaff706b133a7cf063cc26a923_302x55.jpg)
下面是一个有关 [`mean_absolute_error`](generated/sklearn.metrics.mean_absolute_error.html#sklearn.metrics.mean_absolute_error "sklearn.metrics.mean_absolute_error") 函数用法的小例子:
```
>>> from sklearn.metrics import mean_absolute_error
>>> y_true = [3, -0.5, 2, 7]
>>> y_pred = [2.5, 0.0, 2, 8]
>>> mean_absolute_error(y_true, y_pred)
0.5
>>> y_true = [[0.5, 1], [-1, 1], [7, -6]]
>>> y_pred = [[0, 2], [-1, 2], [8, -5]]
>>> mean_absolute_error(y_true, y_pred)
0.75
>>> mean_absolute_error(y_true, y_pred, multioutput='raw_values')
array([ 0.5, 1. ])
>>> mean_absolute_error(y_true, y_pred, multioutput=[0.3, 0.7])
...
0.849...
```
### 3.3.4.3. 均方误差
该 [`mean_squared_error`](generated/sklearn.metrics.mean_squared_error.html#sklearn.metrics.mean_squared_error "sklearn.metrics.mean_squared_error") 函数计算了 [均方误差](https://en.wikipedia.org/wiki/Mean_squared_error), 一个对应于平方(二次)误差或损失的预期值的风险度量.
如果 ![\hat{y}_i](https://box.kancloud.cn/a71dac1a33f465fa5197c33da8720585_13x17.jpg) 是 ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg)-th 样本的预测值, 并且 ![y_i](https://box.kancloud.cn/e79627211612bc56c4f7d926a93fbe8d_13x12.jpg) 是对应的真实值, 则均方误差(MSE)预估的 ![n_{\text{samples}}](https://box.kancloud.cn/70c64e65bf204369a0c85d76a39ecf38_55x14.jpg) 定义如下
![\text{MSE}(y, \hat{y}) = \frac{1}{n_\text{samples}} \sum_{i=0}^{n_\text{samples} - 1} (y_i - \hat{y}_i)^2.](https://box.kancloud.cn/15e8ddbde0468048a7e3cb5ff72178b5_304x55.jpg)
下面是一个有关 [`mean_squared_error`](generated/sklearn.metrics.mean_squared_error.html#sklearn.metrics.mean_squared_error "sklearn.metrics.mean_squared_error") 函数用法的小例子:
```
>>> from sklearn.metrics import mean_squared_error
>>> y_true = [3, -0.5, 2, 7]
>>> y_pred = [2.5, 0.0, 2, 8]
>>> mean_squared_error(y_true, y_pred)
0.375
>>> y_true = [[0.5, 1], [-1, 1], [7, -6]]
>>> y_pred = [[0, 2], [-1, 2], [8, -5]]
>>> mean_squared_error(y_true, y_pred)
0.7083...
```
Examples:
- 点击 [Gradient Boosting regression](../auto_examples/ensemble/plot_gradient_boosting_regression.html#sphx-glr-auto-examples-ensemble-plot-gradient-boosting-regression-py)查看均方误差用于梯度上升(gradient boosting)回归的使用例子。
### 3.3.4.4. 均方误差对数
该 [`mean_squared_log_error`](generated/sklearn.metrics.mean_squared_log_error.html#sklearn.metrics.mean_squared_log_error "sklearn.metrics.mean_squared_log_error") 函数计算了一个对应平方对数(二次)误差或损失的预估值风险度量.
如果 ![\hat{y}_i](https://box.kancloud.cn/a71dac1a33f465fa5197c33da8720585_13x17.jpg) 是 ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg)-th 样本的预测值, 并且 ![y_i](https://box.kancloud.cn/e79627211612bc56c4f7d926a93fbe8d_13x12.jpg) 是对应的真实值, 则均方误差对数(MSLE)预估的 ![n_{\text{samples}}](https://box.kancloud.cn/70c64e65bf204369a0c85d76a39ecf38_55x14.jpg) 定义如下
![\text{MSLE}(y, \hat{y}) = \frac{1}{n_\text{samples}} \sum_{i=0}^{n_\text{samples} - 1} (\log_e (1 + y_i) - \log_e (1 + \hat{y}_i) )^2.](https://box.kancloud.cn/dcbceb4d5febafaf34d96cea44ddb566_465x55.jpg)
其中 ![\log_e (x)](https://box.kancloud.cn/2f7277025fecc7b33c608e12a780e647_53x19.jpg) 表示 ![x](https://box.kancloud.cn/8b09310f09e864e3a4f6d92b559afe29_10x8.jpg) 的自然对数. 当目标具有指数增长的趋势时, 该指标最适合使用, 例如人口数量, 跨年度商品的平均销售额等. 请注意, 该指标会对低于预测的估计值进行估计.
下面是一个有关 [`mean_squared_log_error`](generated/sklearn.metrics.mean_squared_log_error.html#sklearn.metrics.mean_squared_log_error "sklearn.metrics.mean_squared_log_error") 函数用法的小例子:
```
>>> from sklearn.metrics import mean_squared_log_error
>>> y_true = [3, 5, 2.5, 7]
>>> y_pred = [2.5, 5, 4, 8]
>>> mean_squared_log_error(y_true, y_pred)
0.039...
>>> y_true = [[0.5, 1], [1, 2], [7, 6]]
>>> y_pred = [[0.5, 2], [1, 2.5], [8, 8]]
>>> mean_squared_log_error(y_true, y_pred)
0.044...
```
### 3.3.4.5. 中位绝对误差
该 [`median_absolute_error`](generated/sklearn.metrics.median_absolute_error.html#sklearn.metrics.median_absolute_error "sklearn.metrics.median_absolute_error") 函数尤其有趣, 因为它的离群值很强. 通过取目标和预测之间的所有绝对差值的中值来计算损失.
如果 ![\hat{y}_i](https://box.kancloud.cn/a71dac1a33f465fa5197c33da8720585_13x17.jpg) 是 ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg)-th 样本的预测值, 并且 ![y_i](https://box.kancloud.cn/e79627211612bc56c4f7d926a93fbe8d_13x12.jpg) 是对应的真实值, 则中位绝对误差(MedAE)预估的 ![n_{\text{samples}}](https://box.kancloud.cn/70c64e65bf204369a0c85d76a39ecf38_55x14.jpg) 定义如下
![\text{MedAE}(y, \hat{y}) = \text{median}(\mid y_1 - \hat{y}_1 \mid, \ldots, \mid y_n - \hat{y}_n \mid).](https://box.kancloud.cn/cd4fd08eb6c0c1dc2ec0e8a1b527b1c6_391x19.jpg)
该 [`median_absolute_error`](generated/sklearn.metrics.median_absolute_error.html#sklearn.metrics.median_absolute_error "sklearn.metrics.median_absolute_error") 函数不支持多输出.
下面是一个有关 [`median_absolute_error`](generated/sklearn.metrics.median_absolute_error.html#sklearn.metrics.median_absolute_error "sklearn.metrics.median_absolute_error") 函数用法的小例子:
```
>>> from sklearn.metrics import median_absolute_error
>>> y_true = [3, -0.5, 2, 7]
>>> y_pred = [2.5, 0.0, 2, 8]
>>> median_absolute_error(y_true, y_pred)
0.5
```
### 3.3.4.6. R² score, 可决系数
该 [`r2_score`](generated/sklearn.metrics.r2_score.html#sklearn.metrics.r2_score "sklearn.metrics.r2_score") 函数计算了 computes R², 即 [可决系数](https://en.wikipedia.org/wiki/Coefficient_of_determination). 它提供了将来样本如何可能被模型预测的估量. 最佳分数为 1.0, 可以为负数(因为模型可能会更糟). 总是预测 y 的预期值,不考虑输入特征的常数模型将得到 R^2 得分为 0.0.
如果 ![\hat{y}_i](https://box.kancloud.cn/a71dac1a33f465fa5197c33da8720585_13x17.jpg) 是 ![i](https://box.kancloud.cn/9f004454a20e19932b2a071a380ff8ff_6x13.jpg)-th 样本的预测值, 并且 ![y_i](https://box.kancloud.cn/e79627211612bc56c4f7d926a93fbe8d_13x12.jpg) 是对应的真实值, 则 R² 得分预估的 ![n_{\text{samples}}](https://box.kancloud.cn/70c64e65bf204369a0c85d76a39ecf38_55x14.jpg) 定义如下
![R^2(y, \hat{y}) = 1 - \frac{\sum_{i=0}^{n_{\text{samples}} - 1} (y_i - \hat{y}_i)^2}{\sum_{i=0}^{n_\text{samples} - 1} (y_i - \bar{y})^2}](https://box.kancloud.cn/c1c0b11f63ad138415e3c7661d24fd26_273x53.jpg)
其中 ![\bar{y} = \frac{1}{n_{\text{samples}}} \sum_{i=0}^{n_{\text{samples}} - 1} y_i](https://box.kancloud.cn/34472c0a4a8d151c01b20f772feadd89_186x31.jpg).
下面是一个有关 [`r2_score`](generated/sklearn.metrics.r2_score.html#sklearn.metrics.r2_score "sklearn.metrics.r2_score") 函数用法的小例子:
```
>>> from sklearn.metrics import r2_score
>>> y_true = [3, -0.5, 2, 7]
>>> y_pred = [2.5, 0.0, 2, 8]
>>> r2_score(y_true, y_pred)
0.948...
>>> y_true = [[0.5, 1], [-1, 1], [7, -6]]
>>> y_pred = [[0, 2], [-1, 2], [8, -5]]
>>> r2_score(y_true, y_pred, multioutput='variance_weighted')
...
0.938...
>>> y_true = [[0.5, 1], [-1, 1], [7, -6]]
>>> y_pred = [[0, 2], [-1, 2], [8, -5]]
>>> r2_score(y_true, y_pred, multioutput='uniform_average')
...
0.936...
>>> r2_score(y_true, y_pred, multioutput='raw_values')
...
array([ 0.965..., 0.908...])
>>> r2_score(y_true, y_pred, multioutput=[0.3, 0.7])
...
0.925...
```
示例:
- 点击 [Lasso and Elastic Net for Sparse Signals](../auto_examples/linear_model/plot_lasso_and_elasticnet.html#sphx-glr-auto-examples-linear-model-plot-lasso-and-elasticnet-py)查看关于R²用于评估在Lasso and Elastic Net on sparse signals上的使用.
## 3.3.5. 聚类指标
该 [`sklearn.metrics`](classes.html#module-sklearn.metrics "sklearn.metrics") 模块实现了一些 loss, score 和 utility 函数. 更多信息请参阅 [聚类性能度量](clustering.html#clustering-evaluation) 部分, 例如聚类, 以及用于二分聚类的 [Biclustering 评测](biclustering.html#biclustering-evaluation).
## 3.3.6. 虚拟估计
在进行监督学习的过程中,简单的 sanity check(理性检查)包括将人的估计与简单的经验法则进行比较. [`DummyClassifier`](generated/sklearn.dummy.DummyClassifier.html#sklearn.dummy.DummyClassifier "sklearn.dummy.DummyClassifier") 实现了几种简单的分类策略:
- `stratified` 通过在训练集类分布方面来生成随机预测.
- `most_frequent` 总是预测训练集中最常见的标签.
- `prior` always predicts the class that maximizes the class prior (like `most_frequent`) and ``predict_proba` returns the class prior.
- `uniform` 随机产生预测.
- `constant` 总是预测用户提供的常量标签.A major motivation of this method is F1-scoring, when the positive class is in the minority. 这种方法的主要动机是 F1-scoring, 当 positive class(正类)较少时.
请注意, 这些所有的策略, `predict` 方法彻底的忽略了输入数据!
为了说明 [`DummyClassifier`](generated/sklearn.dummy.DummyClassifier.html#sklearn.dummy.DummyClassifier "sklearn.dummy.DummyClassifier"), 首先让我们创建一个 imbalanced dataset:
```
>>> from sklearn.datasets import load_iris
>>> from sklearn.model_selection import train_test_split
>>> iris = load_iris()
>>> X, y = iris.data, iris.target
>>> y[y != 1] = -1
>>> X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)
```
接下来, 让我们比较一下 `SVC` 和 `most_frequent` 的准确性.
```
>>> from sklearn.dummy import DummyClassifier
>>> from sklearn.svm import SVC
>>> clf = SVC(kernel='linear', C=1).fit(X_train, y_train)
>>> clf.score(X_test, y_test)
0.63...
>>> clf = DummyClassifier(strategy='most_frequent',random_state=0)
>>> clf.fit(X_train, y_train)
DummyClassifier(constant=None, random_state=0, strategy='most_frequent')
>>> clf.score(X_test, y_test)
0.57...
```
我们看到 `SVC` 没有比一个 dummy classifier(虚拟分类器)好很多. 现在, 让我们来更改一下 kernel:
```
>>> clf = SVC(kernel='rbf', C=1).fit(X_train, y_train)
>>> clf.score(X_test, y_test)
0.97...
```
我们看到准确率提升到将近 100%. 建议采用交叉验证策略, 以更好地估计精度, 如果不是太耗 CPU 的话. 更多信息请参阅 [交叉验证:评估估算器的表现](cross_validation.html#cross-validation) 部分. 此外,如果要优化参数空间,强烈建议您使用适当的方法; 更多详情请参阅 [调整估计器的超参数](grid_search.html#grid-search) 部分.
通常来说,当分类器的准确度太接近随机情况时,这可能意味着出现了一些问题: 特征没有帮助, 超参数没有正确调整, class 不平衡造成分类器有问题等…
[`DummyRegressor`](generated/sklearn.dummy.DummyRegressor.html#sklearn.dummy.DummyRegressor "sklearn.dummy.DummyRegressor") 还实现了四个简单的经验法则来进行回归:
- `mean` 总是预测训练目标的平均值.
- `median` 总是预测训练目标的中位数.
- `quantile` 总是预测用户提供的训练目标的 quantile(分位数).
- `constant` 总是预测由用户提供的常数值.
在以上所有的策略中, `predict` 方法完全忽略了输入数据.
- scikit-learn 0.19 中文文档
- 用户指南
- 1. 监督学习
- 1.1. 广义线性模型
- 1.2. 线性和二次判别分析
- 1.3. 内核岭回归
- 1.4. 支持向量机
- 1.5. 随机梯度下降
- 1.6. 最近邻
- 1.7. 高斯过程
- 1.8. 交叉分解
- 1.9. 朴素贝叶斯
- 1.10. 决策树
- 1.11. 集成方法
- 1.12. 多类和多标签算法
- 1.13. 特征选择
- 1.14. 半监督学习
- 1.15. 等式回归
- 1.16. 概率校准
- 1.17. 神经网络模型(有监督)
- 2. 无监督学习
- 2.1. 高斯混合模型
- 2.2. 流形学习
- 2.3. 聚类
- 2.4. 双聚类
- 2.5. 分解成分中的信号(矩阵分解问题)
- 2.6. 协方差估计
- 2.7. 经验协方差
- 2.8. 收敛协方差
- 2.9. 稀疏逆协方差
- 2.10. Robust 协方差估计
- 2.11. 新奇和异常值检测
- 2.12. 密度估计
- 2.13. 神经网络模型(无监督)
- 3. 模型选择和评估
- 3.1. 交叉验证:评估估算器的表现
- 3.2. 调整估计器的超参数
- 3.3. 模型评估: 量化预测的质量
- 3.4. 模型持久化
- 3.5. 验证曲线: 绘制分数以评估模型
- 4. 数据集转换
- 4.1. Pipeline(管道)和 FeatureUnion(特征联合): 合并的评估器
- 4.2. 特征提取
- 4.3. 预处理数据
- 4.4. 无监督降维
- 4.5. 随机投影
- 4.6. 内核近似
- 4.7. 成对的矩阵, 类别和核函数
- 4.8. 预测目标 (y) 的转换
- 5. 数据集加载工具
- 6. 大规模计算的策略: 更大量的数据
- 7. 计算性能
- 教程
- 使用 scikit-learn 介绍机器学习
- 关于科学数据处理的统计学习教程
- 机器学习: scikit-learn 中的设置以及预估对象
- 监督学习:从高维观察预测输出变量
- 模型选择:选择估计量及其参数
- 无监督学习: 寻求数据表示
- 把它们放在一起
- 寻求帮助
- 处理文本数据
- 选择正确的评估器(estimator)
- 外部资源,视频和谈话