python – 在Scikit-learn中小于精度和召回的F1

我正在进行多级分类,具有不平衡的类别.

我注意到f1总是小于精度和召回的直接调和平均值,在某些情况下,f1甚至小于精度和召回率.

仅供参考,我将metrics.precision_score(y,pred)称为精度等等.

我知道微观/宏观平均值的差异,并通过使用precision_recall_fscore_support()的类别结果测试它们不是微观的.

不确定这是由于使用了宏观平均值还是其他一些原因？

更新了详细结果如下：

n_samples：75,n_features：250

MultinomialNB(alpha = 0.01,fit_prior = True)

2倍CV：

第一轮：

F1:        0.706029106029
Precision: 0.731531531532
Recall:    0.702702702703

         precision    recall  f1-score   support

      0       0.44      0.67      0.53         6
      1       0.80      0.50      0.62         8
      2       0.78      0.78      0.78        23

avg / total       0.73      0.70      0.71        37

第二轮：

F1:        0.787944219523
Precision: 0.841165413534
Recall:    0.815789473684

         precision    recall  f1-score   support

      0       1.00      0.29      0.44         7
      1       0.75      0.86      0.80         7
      2       0.82      0.96      0.88        24

avg / total       0.84      0.82      0.79        38

总体：

Overall f1-score:   0.74699 (+/- 0.02)
Overall precision:  0.78635 (+/- 0.03)
Overall recall:     0.75925 (+/- 0.03)

从Scholarpedia开始的微/平均值定义：

In multi-label classification, the simplest method for computing an
aggregate score across categories is to average the scores of all
binary task. The resulted scores are called macro-averaged recall,
precision, F1, etc. Another way of averaging is to sum over TP, FP,
TN, FN and N over all the categories first, and then compute each of
the above metrics. The resulted scores are called micro-averaged.
Macro-averaging gives an equal weight to each category, and is often
dominated by the system’s performance on rare categories (the
majority) in a power-law like distribution. Micro-averaging gives an
equal weight to each document, and is often dominated by the system’s
performance on most common categories.

目前在Github的#open issue,#83.

以下示例演示了Micro,Macro和加权(Scikit-learn中的当前)平均值可能有何不同：

y    = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2]
pred = [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 2, 0, 1, 2, 2, 2, 2]

混淆矩阵：

[[9 3 0]
 [3 5 1]
 [1 1 4]]

Wei Pre: 0.670655270655
Wei Rec: 0.666666666667
Wei F1 : 0.666801346801
Wei F5 : 0.668625356125

Mic Pre: 0.666666666667
Mic Rec: 0.666666666667
Mic F1 : 0.666666666667
Mic F5 : 0.666666666667

Mac Pre: 0.682621082621
Mac Rec: 0.657407407407
Mac F1 : 0.669777037588
Mac F5 : 0.677424801371

上面的F5是F0.5的简写…

解决方法:

你可以用以下输出更新你的问题：

>>> from sklearn.metrics import classification_report
>>> print classification_report(y_true, y_predicted)

这将显示每个类别的精确度和召回以及支持,从而帮助我们理解平均值如何工作并决定这是否是适当的行为.

标签：python,machine-learning,scikit-learn,scikits
来源： https://codeday.me/bug/20190704/1378466.html