如何在scikit learn的管道中实现RandomUnderSampler？

Question

我有一个scikit learn的pipeline，用于缩放数字特征和编码分类特征。一切都很正常，直到我尝试使用imblearn的RandomUnderSampler。我的目标是实现下采样步骤，因为我的数据集非常不平衡，比例为1：1000。

我确保使用了imblearn而不是sklearn的Pipeline方法。以下是我尝试过的代码。

如果不使用下采样器方法，则代码数据可以正常工作（使用sklearn pipeline）。

from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.preprocessing import OneHotEncoder, StandardScaler
from imblearn.pipeline import make_pipeline as make_pipeline_imb
from imblearn.pipeline import Pipeline as Pipeline_imb

from sklearn.base import BaseEstimator, TransformerMixin
class TypeSelector(BaseEstimator, TransformerMixin):
    def __init__(self, dtype):
        self.dtype = dtype
    def fit(self, X, y=None):
        return self
    def transform(self, X):
        assert isinstance(X, pd.DataFrame)
        return X.select_dtypes(include=[self.dtype])

transformer = Pipeline([
    # Union numeric, categoricals and boolean
    ('features', FeatureUnion(n_jobs=1, transformer_list=[
         # Select bolean features                                                  
        ('boolean', Pipeline([
            ('selector', TypeSelector('bool')),
        ])),
         # Select and scale numericals
        ('numericals', Pipeline([
            ('selector', TypeSelector(np.number)),
            ('scaler', StandardScaler()),
        ])),
         # Select and encode categoricals
        ('categoricals', Pipeline([
            ('selector', TypeSelector('category')),
            ('encoder', OneHotEncoder(handle_unknown='ignore')),
        ])) 
    ])),
])
pipe = Pipeline([('prep', transformer), 
                 ('clf', RandomForestClassifier(n_estimators=500, class_weight='balanced'))
                 ])

使用imblearn管道时，使用下采样方法的代码无法正常工作。

from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.preprocessing import OneHotEncoder, StandardScaler
from imblearn.pipeline import make_pipeline as make_pipeline_imb
from imblearn.pipeline import Pipeline as Pipeline_imb

from sklearn.base import BaseEstimator, TransformerMixin
class TypeSelector(BaseEstimator, TransformerMixin):
    def __init__(self, dtype):
        self.dtype = dtype
    def fit(self, X, y=None):
        return self
    def transform(self, X):
        assert isinstance(X, pd.DataFrame)
        return X.select_dtypes(include=[self.dtype])

transformer = Pipeline_imb([
    # Union numeric, categoricals and boolean
    ('features', FeatureUnion(n_jobs=1, transformer_list=[
         # Select bolean features                                                  
        ('boolean', Pipeline_imb([
            ('selector', TypeSelector('bool')),
        ])),
         # Select and scale numericals
        ('numericals', Pipeline_imb([
            ('selector', TypeSelector(np.number)),
            ('scaler', StandardScaler()),
        ])),
         # Select and encode categoricals
        ('categoricals', Pipeline_imb([
            ('selector', TypeSelector('category')),
            ('encoder', OneHotEncoder(handle_unknown='ignore')),
        ])) 
    ])),  
])
pipe = Pipeline_imb([
                 ('sampler', RandomUnderSampler(0.1)),
                 ('prep', transformer), 
                 ('clf', RandomForestClassifier(n_estimators=500, class_weight='balanced'))
                 ])

这是我收到的错误信息：

/usr/local/lib/python3.6/dist-packages/sklearn/pipeline.py in __init__(self, steps, memory, verbose)
    133     def __init__(self, steps, memory=None, verbose=False):
    134         self.steps = steps
--> 135         self._validate_steps()
    136         self.memory = memory
    137         self.verbose = verbose

/usr/local/lib/python3.6/dist-packages/imblearn/pipeline.py in _validate_steps(self)
    144             if isinstance(t, pipeline.Pipeline):
    145                 raise TypeError(
--> 146                     "All intermediate steps of the chain should not be"
    147                     " Pipelines")
    148 

TypeError: All intermediate steps of the chain should not be Pipelines

- Ale M.

---

我找不到其他方法，只能通过将嵌套管道中的步骤取出并堆叠到单个 imblearn.pipeline.Pipeline 中来展平管道。 - ayorgo

Answer 1

如果你查看imblearn/pipeline.py这里中的代码，会发现在函数_validate_steps下，它们将检查transformers中的每个项目是否存在scikit的Pipeline实例化的转换器 (isinstance(t, pipeline.Pipeline))。

从你的代码来看，transformers是：

1. RandomUnderSampler
2. transformer

同时，在你的代码中使用Pipeline_imb是多余的，因为Pipeline_imb类继承了scikit的Pipeline。

以上是我的调整建议。

transformer = FeatureUnion(n_jobs=1, transformer_list=[
     # Select bolean features                                                  
    ('selector1', TypeSelector('bool'),
     # Select and scale numericals
    ('selector2', TypeSelector(np.number)),
    ('scaler', StandardScaler()),
     # Select and encode categoricals
    ('selector3', TypeSelector('category')),
    ('encoder', OneHotEncoder(handle_unknown='ignore'))
])

pipe = Pipeline_imb([
    ('sampler', RandomUnderSampler(0.1)),
    ('prep', transformer), 
    ('clf', RandomForestClassifier(n_estimators=500, class_weight='balanced'))
])

- Bryan