【文本分类】多意图分类评估指标

【文本分类】多意图分类评估指标

主要分为两类：label based measures和example based measures。

label based measures

就是针对每一个分类，都进行一次计算，最后再用一种average方法把多个分类统一起来。

假设有这么一组数据，

expected   predicted
A, C        A, B
C           C
A, B, C     B, C

用sklearn MultiLabelBinarizer 进行转化

expected    predicted
1 0 1       1 1 0
0 0 1       0 0 1
1 1 1       0 1 1

对classA来说，

TP = 1（真实和预测都是1）

FP = 0（真实0，预测1）

TN = 1（真实0，预测0）

FN = 1（真实1，预测0）

TN   FP           1   0
FN   TP           1   1

precision = TP / (TP + FP) = 1 / (1+0) = 1
  
recall = TP / (TP + FN) = 1 / (1+1) = 0.5
  
f1-score = 2*p*r / (p+r) = 0.667

class B

TN   FP           1   1
FN   TP           0   1

Precision = 0.5

Recall = 1.0

F1-score = 0.667

class C

TN   FP           0   0
FN   TP           1   2

Precision = 1.0

Recall = 0.667

F1-score = 0.8

• macro average

  Precision (macro avg)
  = (Precision of A + Precision of B + Precision of C) / 3
  = 0.833
  

• micro average (preferred)

  Precision (micro avg)
  = sum(TP) / (sum(TP) + sum(FP))
  = 1+1+2 / ((1+1+2) + (0+1+0))
  = 0.8
  

• weighted average

  Precision(weighted avg)
  = [(Precision of A * support A) + 
  (Precision of B * support B) + 
  (Precision of C * support C)] 
  / (support A + support B + support C)
  = (1*2 + 0.5*1 + 1*3) / 6
  = 0.9166
  

• sample average
  第一行，真实 AC，预测 AB，precision 1/2 → 两个预测值中有一个是正确的

  第二行，真实 C，预测 C，precision 1

  第三行，真实 ABC，预测 BC，precision 1 → 预测的都是对的

  (1/2 + 1 + 1) / 3 = 5/6 = 0.833
  

• classification_report
  直接用classification_report
  

example based measures

计算每对真实与预测标签的average difference

• hamming loss

  预测错了的label占总label的比例

• subset accuracy

  也叫exact match ratio

  最严格的评估方法，真实和预测label必须完全一致，否则为0。这种方法忽略了部分正确的情况，在scikit-learn中的accuracy_score就是subset accuracy。

• example-based accuracy

  预测正确的label占总label（预测为1和真实为1）的比例

• example-based precision
  预测正确的label占总预测label的比例

参考来源

https://towardsdatascience.com/evaluating-multi-label-classifiers-a31be83da6ea Evaluating Multi-label Classifiers

https://towardsdatascience.com/journey-to-the-center-of-multi-label-classification-384c40229bff Deep dive into multi-label classification…! (With detailed Case Study)

https://medium.datadriveninvestor.com/a-survey-of-evaluation-metrics-for-multilabel-classification-bb16e8cd41cd Evaluation Metrics for Multi-Label Classification