OneHotEncoder未转换前序转换器生成的新列，如何解决？

ColumnTransformer的各个转换器是并行独立处理原始输入数据的：

• date转换器处理原始的visitStartTime列，生成day_of_week、hour等新特征
• ohe转换器仅处理原始数据中dtype_include='object'的列，不会处理date转换器生成的新分类列
  两者结果直接拼接，导致新生成的day_of_week等分类列未被编码。

以下两种方式均可解决问题：

方案1：先统一生成时间特征，再全局处理所有分类列 #

先通过时间转换生成全部衍生特征，再用ColumnTransformer对所有分类特征（包括新生成的）编码，数值特征直接保留。

import pandas as pd
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder
from sklearn.compose import make_column_selector as selector

time_col = ['visitStartTime']

class TimeTransformer:
    def fit(self, X, y=None):
        return self
    
    def transform(self, X):
        # 复制输入避免修改原始数据
        X_copy = X.copy()
        for column in X_copy.columns:
            X_copy['time'] = pd.to_datetime(X_copy[column], unit='s', origin='unix')
            X_copy['day_of_week'] = X_copy['time'].dt.strftime('%A')
            X_copy['hour'] = X_copy['time'].dt.hour
            X_copy['day'] = X_copy['time'].dt.day
            X_copy['month'] = X_copy['time'].dt.month
            X_copy['year'] = X_copy['time'].dt.year
            X_copy = X_copy.drop(['time', column], axis=1)
        return X_copy

# 第一步：处理时间列生成衍生特征
time_transformer = TimeTransformer()

# 分类特征编码规则（覆盖新生成的day_of_week和原始object列）
ohe_transformer = Pipeline(steps=[
    ('ohe', OneHotEncoder(sparse_output=False))
])

# 数值特征直接保留（包括hour、day等时间衍生列和原始数值列）
num_transformer = Pipeline(steps=[])

# 构建完整预处理管道
preprocessor = Pipeline(steps=[
    ('time_process', time_transformer),
    ('feature_process', ColumnTransformer(transformers=[
        ('cat', ohe_transformer, selector(dtype_include='object')),
        ('num', num_transformer, selector(dtype_exclude='object'))
    ], remainder='passthrough'))
])

# 拟合转换数据
j = preprocessor.fit_transform(X_train)

方案2：在时间处理子管道内单独编码分类特征 #

针对时间列构建子管道，先生成特征，再对其中的分类列编码、数值列保留；同时处理原始分类列的编码，最后合并所有结果。

import pandas as pd
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder
from sklearn.compose import make_column_selector as selector

time_col = ['visitStartTime']

class TimeTransformer:
    def fit(self, X, y=None):
        return self
    
    def transform(self, X):
        X_copy = X.copy()
        for column in X_copy.columns:
            X_copy['time'] = pd.to_datetime(X_copy[column], unit='s', origin='unix')
            X_copy['day_of_week'] = X_copy['time'].dt.strftime('%A')
            X_copy['hour'] = X_copy['time'].dt.hour
            X_copy['day'] = X_copy['time'].dt.day
            X_copy['month'] = X_copy['time'].dt.month
            X_copy['year'] = X_copy['time'].dt.year
            X_copy = X_copy.drop(['time', column], axis=1)
        return X_copy

# 时间特征子管道：生成特征后，单独编码分类列、保留数值列
time_feature_process = ColumnTransformer(transformers=[
    ('time_cat', OneHotEncoder(sparse_output=False), ['day_of_week']),
    ('time_num', 'passthrough', ['hour', 'day', 'month', 'year'])
])

time_transformer_pipeline = Pipeline(steps=[
    ('time', TimeTransformer()),
    ('time_feature_encode', time_feature_process)
])

# 原始分类列编码管道
ohe_transformer = Pipeline(steps=[
    ('ohe', OneHotEncoder(sparse_output=False))
])

# 合并所有处理管道
preprocessor = ColumnTransformer(transformers=[
    ('date', time_transformer_pipeline, time_col),
    ('ohe', ohe_transformer, selector(dtype_include='object'))
], remainder='passthrough')

j = preprocessor.fit_transform(X_train)

• 在TimeTransformer的transform方法中复制输入数据，避免修改原始数据集。
• sklearn 1.2+版本用OneHotEncoder(sparse_output=False)，旧版本替换为sparse=False，方便查看输出结果。

内容的提问来源于stack exchange，提问作者gkl kmr