分类问题的评价指标:多分类【Precision、 micro-P、macro-P】、【Recall、micro-R、macro-R】、【F1、 micro-F1、macro-F1】

一、混淆矩阵

对于二分类的模型,预测结果与实际结果分别可以取0和1。我们用N和P代替0和1,T和F表示预测正确和错误。将他们两两组合,就形成了下图所示的混淆矩阵(注意:组合结果都是针对预测结果而言的)。

由于1和0是数字,阅读性不好,所以我们分别用P和N表示1和0两种结果。变换之后为PP,PN,NP,NN,阅读性也很差,我并不能轻易地看出来预测的正确性与否。因此,为了能够更清楚地分辨各种预测情况是否正确,我们将其中一个符号修改为T和F,以便于分辨出结果。

在这里插入图片描述

  • P(Positive):代表 1
  • N(Negative):代表 0
  • T(True):代表预测正确
  • F(False):代表预测错误

二、准确率、精确率、召回率、F1-Measure

在这里插入图片描述

  • 准确率(Accuracy):对于给定的测试数据集,分类器正确分类的样本数与总样本数之比。
    A c c u r a c y = T P + T N T P + T N + F P + F N = T P + T N 总 样 本 数 量 Accuracy=\cfrac{TP+TN}{TP+TN+FP+FN}=\cfrac{TP+TN}{总样本数量} Accuracy=TP+TN+FP+FNTP+TN=TP+TN
  • 精确率(Precision)**:精指分类正确的正样本个数(TP)占分类器判定为正样本的样本个数(TP+FP)的比例。
    P r e c i s i o n = T P T P + F P = 分 类 正 确 的 正 样 本 个 数 判 定 为 正 样 本 的 样 本 个 数 Precision=\cfrac{TP}{TP+FP}=\cfrac{分类正确的正样本个数}{判定为正样本的样本个数} Precision=TP+FPTP=
  • 召回率(Recall):召回率是指分类正确的正样本个数(TP)占真正的正样本个数(TP+FN)的比例。
    R e c a l l = T P T P + F N = 分 类 正 确 的 正 样 本 个 数 全 部 真 正 的 正 样 本 个 数 Recall=\cfrac{TP}{TP+FN}=\cfrac{分类正确的正样本个数}{全部真正的正样本个数} Recall=TP+FNTP=
  • F1-Measure值:就是精确率和召回率的调和平均值
    F 1 − M e a s u r e = 2 1 P r e c i s i o n + 1 R e c a l l = 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l \begin{aligned}F1-Measure=\cfrac{2}{\cfrac{1}{Precision}+\cfrac{1}{Recall}}=\cfrac{2×Precision×Recall}{Precision+Recall}\end{aligned} F1Measure=Precision1+Recall12=Precision+Recall2×Precision×Recall

每个评估指标都有其价值,但如果只从单一的评估指标出发去评估模型,往往会得出片面甚至错误的结论;只有通过一组互补的指标去评估模型,才能更好地发现并解决模型存在的问题,从而更好地解决实际业务场景中遇到的问题。

三、多分类评价指标-案例

假设有如下的数据

预测 真实
A A
A A
B A
C A
B B
B B
C B
B C
C C

可以看出,上表为一份样本量为9,类别数为3的含标注结果的三分类预测样本。TN对于准召的计算而言是不需要的,因此下面的表格中未统计该值。

1、按照定义计算Precision、Recall

1.1 对于类别A

TP = 2 FP = 0
FN = 2 TN = ~

P r e c i s i o n = T P T P + F P = 分 类 正 确 的 正 样 本 个 数 判 定 为 正 样 本 的 样 本 个 数 = 2 2 + 0 = 100 % = 1.0 Precision=\cfrac{TP}{TP+FP}=\cfrac{分类正确的正样本个数}{判定为正样本的样本个数}=\cfrac{2}{2+0}=100\%=1.0 Precision=TP+FPTP==2+02=100%=1.0

R e c a l l = T P T P + F N = 分 类 正 确 的 正 样 本 个 数 真 正 的 正 样 本 个 数 = 2 2 + 2 = 50 % = 0.5 Recall=\cfrac{TP}{TP+FN}=\cfrac{分类正确的正样本个数}{真正的正样本个数}=\cfrac{2}{2+2}=50\%=0.5 Recall=TP+FNTP==2+22=50%=0.5

1.2 对于类别B

TP = 2 FP = 2
FN = 1 TN = ~

P r e c i s i o n = T P T P + F P = 分 类 正 确 的 正 样 本 个 数 判 定 为 正 样 本 的 样 本 个 数 = 2 2 + 2 = 50 % = 0.5 Precision=\cfrac{TP}{TP+FP}=\cfrac{分类正确的正样本个数}{判定为正样本的样本个数}=\cfrac{2}{2+2}=50\%=0.5 Precision=TP+FPTP==2+22=50%=0.5

R e c a l l = T P T P + F N = 分 类 正 确 的 正 样 本 个 数 真 正 的 正 样 本 个 数 = 2 2 + 1 = 67 % = 0.67 Recall=\cfrac{TP}{TP+FN}=\cfrac{分类正确的正样本个数}{真正的正样本个数}=\cfrac{2}{2+1}=67\%=0.67 Recall=TP+FNTP==2+12=67%=0.67

1.3 对于类别C

TP = 1 FP = 2
FN = 1 TN = ~

P r e c i s i o n = T P T P + F P = 分 类 正 确 的 正 样 本 个 数 判 定 为 正 样 本 的 样 本 个 数 = 1 1 + 2 = 33 % = 0.33 Precision=\cfrac{TP}{TP+FP}=\cfrac{分类正确的正样本个数}{判定为正样本的样本个数}=\cfrac{1}{1+2}=33\%=0.33 Precision=TP+FPTP==1+21=33%=0.33

R e c a l l = T P T P + F N = 分 类 正 确 的 正 样 本 个 数 真 正 的 正 样 本 个 数 = 1 1 + 1 = 50 % = 0.5 Recall=\cfrac{TP}{TP+FN}=\cfrac{分类正确的正样本个数}{真正的正样本个数}=\cfrac{1}{1+1}=50\%=0.5 Recall=TP+FNTP==1+11=50%=0.5

2、调用sklearn的api进行验证

from sklearn.metrics import classification_report
from sklearn.metrics import precision_score, recall_score, f1_score

true_lable = [0, 0, 0, 0, 1, 1, 1, 2, 2]
prediction = [0, 0, 1, 2, 1, 1, 2, 1, 2]


measure_result = classification_report(true_lable, prediction)
print('measure_result = \n', measure_result)

打印结果:

measure_result = 
               precision    recall  f1-score   support

           0       1.00      0.50      0.67         4
           1       0.50      0.67      0.57         3
           2       0.33      0.50      0.40         2

    accuracy                           0.56         9
   macro avg       0.61      0.56      0.55         9
weighted avg       0.69      0.56      0.58         9

四、Micro-F1、Macro-F1、weighted-F1

分类问题的评价指标:多分类【Precision、 micro-P、macro-P】、【Recall、micro-R、macro-R】、【F1、 micro-F1、macro-F1】_第1张图片

总的来说,微观F1(micro-F1)和宏观F1(macro-F1)都是F1合并后的结果,这两个F1都是用在多分类任务中的评价指标,是两种不一样的求F1均值的方式;micro-F1和macro-F1的计算方法有差异,得出来的结果也略有差异;

1、Micro-F1

Micro-F1 不需要区分类别,直接使用总体样本的准召计算f1 score。

  • 计算方法:先计算所有类别的总的Precision和Recall,然后计算出来的F1值即为micro-F1;

  • 使用场景:在计算公式中考虑到了每个类别的数量,所以适用于数据分布不平衡的情况;但同时因为考虑到数据的数量,所以在数据极度不平衡的情况下,数量较多数量的类会较大的影响到F1的值;

该样本的混淆矩阵如下:

TP = 5 FP = 4
FN = 2 TN = ~

P r e c i s i o n = T P T P + F P = 分 类 正 确 的 正 样 本 个 数 判 定 为 正 样 本 的 样 本 个 数 = 5 5 + 4 = 55.56 % = 0.5556 Precision=\cfrac{TP}{TP+FP}=\cfrac{分类正确的正样本个数}{判定为正样本的样本个数}=\cfrac{5}{5+4}=55.56\%=0.5556 Precision=TP+FPTP==5+45=55.56%=0.5556

R e c a l l = T P T P + F N = 分 类 正 确 的 正 样 本 个 数 真 正 的 正 样 本 个 数 = 5 5 + 4 = 55.56 % = 0.5556 Recall=\cfrac{TP}{TP+FN}=\cfrac{分类正确的正样本个数}{真正的正样本个数}=\cfrac{5}{5+4}=55.56\%=0.5556 Recall=TP+FNTP==5+45=55.56%=0.5556

F 1 − M e a s u r e = 2 1 P r e c i s i o n + 1 R e c a l l = 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l = 2 × 0.5556 × 0.5556 0.5556 + 0.5556 = 0.5556 \begin{aligned}F1-Measure=\cfrac{2}{\cfrac{1}{Precision}+\cfrac{1}{Recall}}=\cfrac{2×Precision×Recall}{Precision+Recall}=\cfrac{2×0.5556×0.5556}{0.5556+0.5556}=0.5556\end{aligned} F1Measure=Precision1+Recall12=Precision+Recall2×Precision×Recall=0.5556+0.55562×0.5556×0.5556=0.5556

2、Macro-F1

不同于micro f1,macro f1需要先计算出每一个类别的准召及其f1 score,然后通过求均值得到在整个样本上的f1 score。

  • 计算方法:将所有类别的Precision和Recall求平均,然后计算F1值作为macro-F1;
  • 使用场景:没有考虑到数据的数量,所以会平等的看待每一类(因为每一类的precision和recall都在0-1之间),会相对受高precision和高recall类的影响较大;

类别A的​:
F 1 − A = 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l = 2 × 1 × 0.5 1 + 0.5 = 0.6667 \begin{aligned}F1-A=\cfrac{2×Precision×Recall}{Precision+Recall}=\cfrac{2×1×0.5}{1+0.5}=0.6667\end{aligned} F1A=Precision+Recall2×Precision×Recall=1+0.52×1×0.5=0.6667

类别B的​:
F 1 − B = 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l = 2 × 0.5 × 0.67 0.5 + 0.67 = 0.57265 \begin{aligned}F1-B=\cfrac{2×Precision×Recall}{Precision+Recall}=\cfrac{2×0.5×0.67}{0.5+0.67}=0.57265\end{aligned} F1B=Precision+Recall2×Precision×Recall=0.5+0.672×0.5×0.67=0.57265

类别C的​:
F 1 − C = 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l = 2 × 0.33 × 0.5 0.33 + 0.5 = 0.39759 \begin{aligned}F1-C=\cfrac{2×Precision×Recall}{Precision+Recall}=\cfrac{2×0.33×0.5}{0.33+0.5}=0.39759\end{aligned} F1C=Precision+Recall2×Precision×Recall=0.33+0.52×0.33×0.5=0.39759

Macro-F1为上面三者的平均值:
M a c r o − F 1 = F 1 − A + F 1 − B + F 1 − C 3 = 0.6667 + 0.57265 + 0.39759 3 = 0.546 \begin{aligned}Macro-F1=\cfrac{F1-A + F1-B + F1-C}{3}=\cfrac{0.6667 + 0.57265 + 0.39759}{3}=0.546\end{aligned} MacroF1=3F1A+F1B+F1C=30.6667+0.57265+0.39759=0.546

3、weighted-F1

除了micro-F1和macro-F1,还有weighted-F1,是一个将F1-score乘以该类的比例之后相加的结果,也可以看做是macro-F1的变体吧。

weighted-F1和macro-F1的区别在于:macro-F1对每一类都赋予了相同的权重,而weighted-F1则根据每一类的比例分别赋予不同的权重。

五、指标的选择问题

“我们看到,对于 Macro 来说, 小类别相当程度上拉高了 Precision 的值,而实际上, 并没有那么多样本被正确分类,考虑到实际的环境中,真实样本分布和训练样本分布相同的情况下,这种指标明显是有问题的, 小类别起到的作用太大,以至于大样本的分类情况不佳。 而对于 Micro 来说,其考虑到了这种样本不均衡的问题, 因此在这种情况下相对较佳。

总的来说, 如果你的类别比较均衡,则随便; 如果你认为大样本的类别应该占据更重要的位置, 使用Micro; 如果你认为小样本也应该占据重要的位置,则使用 Macro; 如果 Micro << Macro , 则意味着在大样本类别中出现了严重的分类错误; 如果 Macro << Micro , 则意味着小样本类别中出现了严重的分类错误。

为了解决 Macro 无法衡量样本均衡问题,一个很好的方法是求加权的 Macro, 因此 Weighed F1 出现了。”

六、代码

1、数据01

true_lable = [0, 0, 0, 0, 1, 1, 1, 2, 2]
prediction = [0, 0, 1, 2, 1, 1, 2, 1, 2]
from sklearn.metrics import classification_report
from sklearn.metrics import precision_score, recall_score, f1_score

true_lable = [0, 0, 0, 0, 1, 1, 1, 2, 2]
prediction = [0, 0, 1, 2, 1, 1, 2, 1, 2]


measure_result = classification_report(true_lable, prediction)
print('measure_result = \n', measure_result)

print("----------------------------- precision(精确率)-----------------------------")
precision_score_average_None = precision_score(true_lable, prediction, average=None)
precision_score_average_micro = precision_score(true_lable, prediction, average='micro')
precision_score_average_macro = precision_score(true_lable, prediction, average='macro')
precision_score_average_weighted = precision_score(true_lable, prediction, average='weighted')
print('precision_score_average_None = ', precision_score_average_None)
print('precision_score_average_micro = ', precision_score_average_micro)
print('precision_score_average_macro = ', precision_score_average_macro)
print('precision_score_average_weighted = ', precision_score_average_weighted)

print("\n\n----------------------------- recall(召回率)-----------------------------")
recall_score_average_None = recall_score(true_lable, prediction, average=None)
recall_score_average_micro = recall_score(true_lable, prediction, average='micro')
recall_score_average_macro = recall_score(true_lable, prediction, average='macro')
recall_score_average_weighted = recall_score(true_lable, prediction, average='weighted')
print('recall_score_average_None = ', recall_score_average_None)
print('recall_score_average_micro = ', recall_score_average_micro)
print('recall_score_average_macro = ', recall_score_average_macro)
print('recall_score_average_weighted = ', recall_score_average_weighted)

print("\n\n----------------------------- F1-value-----------------------------")
f1_score_average_None = f1_score(true_lable, prediction, average=None)
f1_score_average_micro = f1_score(true_lable, prediction, average='micro')
f1_score_average_macro = f1_score(true_lable, prediction, average='macro')
f1_score_average_weighted = f1_score(true_lable, prediction, average='weighted')
print('f1_score_average_None = ', f1_score_average_None)
print('f1_score_average_micro = ', f1_score_average_micro)
print('f1_score_average_macro = ', f1_score_average_macro)
print('f1_score_average_weighted = ', f1_score_average_weighted)

打印结果:

measure_result = 
               precision    recall  f1-score   support

           0       1.00      0.50      0.67         4
           1       0.50      0.67      0.57         3
           2       0.33      0.50      0.40         2

    accuracy                           0.56         9
   macro avg       0.61      0.56      0.55         9
weighted avg       0.69      0.56      0.58         9

----------------------------- precision(精确率)-----------------------------
precision_score_average_None =  [1.         0.5        0.33333333]
precision_score_average_micro =  0.5555555555555556
precision_score_average_macro =  0.611111111111111
precision_score_average_weighted =  0.6851851851851852


----------------------------- recall(召回率)-----------------------------
recall_score_average_None =  [0.5        0.66666667 0.5       ]
recall_score_average_micro =  0.5555555555555556
recall_score_average_macro =  0.5555555555555555
recall_score_average_weighted =  0.5555555555555556


----------------------------- F1-value-----------------------------
f1_score_average_None =  [0.66666667 0.57142857 0.4       ]
f1_score_average_micro =  0.5555555555555556
f1_score_average_macro =  0.546031746031746
f1_score_average_weighted =  0.5756613756613757

Process finished with exit code 0

2、数据02

true_lable = [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3]
prediction = [3, 0, 0, 0, 0, 0, 0, 0, 2, 3, 3, 1, 1, 1, 1, 1, 1, 3, 1, 2, 2, 2, 2, 2, 3, 0, 3, 3, 3, 3]
from sklearn.metrics import classification_report
from sklearn.metrics import precision_score, recall_score, f1_score

true_lable = [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3]
prediction = [3, 0, 0, 0, 0, 0, 0, 0, 2, 3, 3, 1, 1, 1, 1, 1, 1, 3, 1, 2, 2, 2, 2, 2, 3, 0, 3, 3, 3, 3]

measure_result = classification_report(true_lable, prediction)
print('measure_result = \n', measure_result)

print("----------------------------- precision(精确率)-----------------------------")
precision_score_average_None = precision_score(true_lable, prediction, average=None)
precision_score_average_micro = precision_score(true_lable, prediction, average='micro')
precision_score_average_macro = precision_score(true_lable, prediction, average='macro')
precision_score_average_weighted = precision_score(true_lable, prediction, average='weighted')
print('precision_score_average_None = ', precision_score_average_None)
print('precision_score_average_micro = ', precision_score_average_micro)
print('precision_score_average_macro = ', precision_score_average_macro)
print('precision_score_average_weighted = ', precision_score_average_weighted)

print("\n\n----------------------------- recall(召回率)-----------------------------")
recall_score_average_None = recall_score(true_lable, prediction, average=None)
recall_score_average_micro = recall_score(true_lable, prediction, average='micro')
recall_score_average_macro = recall_score(true_lable, prediction, average='macro')
recall_score_average_weighted = recall_score(true_lable, prediction, average='weighted')
print('recall_score_average_None = ', recall_score_average_None)
print('recall_score_average_micro = ', recall_score_average_micro)
print('recall_score_average_macro = ', recall_score_average_macro)
print('recall_score_average_weighted = ', recall_score_average_weighted)

print("\n\n----------------------------- F1-value-----------------------------")
f1_score_average_None = f1_score(true_lable, prediction, average=None)
f1_score_average_micro = f1_score(true_lable, prediction, average='micro')
f1_score_average_macro = f1_score(true_lable, prediction, average='macro')
f1_score_average_weighted = f1_score(true_lable, prediction, average='weighted')
print('f1_score_average_None = ', f1_score_average_None)
print('f1_score_average_micro = ', f1_score_average_micro)
print('f1_score_average_macro = ', f1_score_average_macro)
print('f1_score_average_weighted = ', f1_score_average_weighted)

打印结果:

measure_result = 
               precision    recall  f1-score   support

           0       0.88      0.78      0.82         9
           1       0.86      0.75      0.80         8
           2       0.83      0.71      0.77         7
           3       0.56      0.83      0.67         6

    accuracy                           0.77        30
   macro avg       0.78      0.77      0.76        30
weighted avg       0.80      0.77      0.77        30

----------------------------- precision(精确率)-----------------------------
precision_score_average_None =  [0.875      0.85714286 0.83333333 0.55555556]
precision_score_average_micro =  0.7666666666666667
precision_score_average_macro =  0.7802579365079365
precision_score_average_weighted =  0.7966269841269841


----------------------------- recall(召回率)-----------------------------
recall_score_average_None =  [0.77777778 0.75       0.71428571 0.83333333]
recall_score_average_micro =  0.7666666666666667
recall_score_average_macro =  0.7688492063492064
recall_score_average_weighted =  0.7666666666666667


----------------------------- F1-value-----------------------------
f1_score_average_None =  [0.82352941 0.8        0.76923077 0.66666667]
f1_score_average_micro =  0.7666666666666667
f1_score_average_macro =  0.7648567119155354
f1_score_average_weighted =  0.7732126696832579

Process finished with exit code 0



参考资料:
Macro-F1 Score与Micro-F1 Score
分类问题的几个评价指标(Precision、Recall、F1-Score、Micro-F1、Macro-F1)
分类问题中的各种评价指标——precision,recall,F1-score,macro-F1,micro-F1

你可能感兴趣的:(#,NLP应用/分类,分类,机器学习,人工智能)