如何在 scikit-learn 中为 OneVsRestClassifier 设置类权重？答案

【问题标题】：How to set class weights for OneVsRestClassifier in scikit-learn?如何在 scikit-learn 中为 OneVsRestClassifier 设置类权重？
【发布时间】：2013-04-10 21:19:58
【问题描述】：

我需要一个 SVM 作为多标签分类器，所以我决定使用 OneVsRestClassifier 包装器。然而问题出现了，训练集变得高度不平衡：对于给定的类，负样本比正样本多得多。这可以通过 class_weight 参数来解决，但是如果我在封装在 OneVsRestClassifier 中的分类器中使用它，则会出现错误：

from sklearn.svm import LinearSVC
from sklearn.multiclass import OneVsRestClassifier

weights = {'ham': 1, 'eggs': 2}
svm = OneVsRestClassifier(LinearSVC(class_weight=weights))

X = [[1, 2], [3, 4], [5, 6], [7, 8], [9, 0]]
Y = [['ham'], [], ['eggs', 'spam'], ['spam'], ['eggs']]

svm.fit(X, Y)

Traceback (most recent call last):
  File "", line 1, in 
  File "/usr/local/lib/python2.7/site-packages/sklearn/multiclass.py", line 197, in fit
    n_jobs=self.n_jobs)
  File "/usr/local/lib/python2.7/site-packages/sklearn/multiclass.py", line 87, in fit_ovr
    for i in range(Y.shape[1]))
  File "/usr/local/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.py", line 514, in __call__
    self.dispatch(function, args, kwargs)
  File "/usr/local/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.py", line 311, in dispatch
    job = ImmediateApply(func, args, kwargs)
  File "/usr/local/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.py", line 135, in __init__
    self.results = func(*args, **kwargs)
  File "/usr/local/lib/python2.7/site-packages/sklearn/multiclass.py", line 56, in _fit_binary
    estimator.fit(X, y)
  File "/usr/local/lib/python2.7/site-packages/sklearn/svm/base.py", line 681, in fit
    self.classes_, y)
  File "/usr/local/lib/python2.7/site-packages/sklearn/utils/class_weight.py", line 49, in compute_class_weight
    if classes[i] != c:
IndexError: index 2 is out of bounds for axis 0 with size 2

【问题讨论】：

标签： scikit-learn

【解决方案1】：

问题在于 LinearSVC 需要二进制类 [0, 1]。因此，为非二进制类（'ham'、'egg' 甚至 [0,1,2]）赋予权重会失败。但是您可以改用“自动”权重，它会通过选择适当的权重来自动“平衡”您的课程。然后它也适用于您的多类 OneVsRest 分类器。

svm = OneVsRestClassifier(LinearSVC(class_weight='auto'))

X = [[1, 2], [3, 4], [5, 4]]
Y = [0,1,2]

svm.fit(X, Y)

【讨论】：

行得通，谢谢！我没有意识到有一个“自动”选项。
有没有办法在 OneVsRestClassifier 中为每个标签分别提供 sample_weight ？由于 GradientBoostingClassifier 没有 class_weight 选项，所以不平衡数据需要指定 sample_weight。