【机器学习】python机器学习使用scikit-learn对模型进行微调：按特征贡献大小保留最重要k个特征的transform

原创已于 2022-08-19 05:43:07 修改

· 456 阅读

1 ·

版权

文章标签：

#python #机器学习 #scikit-learn

于 2022-08-18 21:30:00 首次发布

python 同时被 3 个专栏收录

89 篇文章

订阅专栏

机器学习

36 篇文章

订阅专栏

scikit-learn

27 篇文章

订阅专栏

from sklearn.base import BaseEstimator, TransformerMixin

def indices_of_top_k(arr, k):
    #np.argpartition ()将传入的数组arr分成 两部分 ，即：排在第k位置前面的数都小于k位置，排在第k位置后面的值都大于k位置
    #np.argpartition(np.array(arr), -k)保证最后k个数大于k位置的数,[-k:]取最后的k个值，这个组合求数组内最大的k个数
    #该函数返回的是下标，再按下标进行排序
    return np.sort(np.argpartition(np.array(arr), -k)[-k:])

class TopFeatureSelector(BaseEstimator, TransformerMixin):
    #feature_importances为特证重要度，越大重要度越高
    def __init__(self, feature_importances, k):
        self.feature_importances = feature_importances
        self.k = k
    def fit(self, X, y=None):
        self.feature_indices_ = indices_of_top_k(self.feature_importances, self.k)
        return self
    def transform(self, X):
        return X[:, self.feature_indices_]

2.用例：
这里以树模型进行举例：feature_importances_ 参数目前只出现在决策树或以决策树为基础的评估器中：
数据准备：

import os

HOUSING_PATH = os.path.join("datasets", "housing")
import pandas as pd

def load_housing_data(housing_path=HOUSING_PATH):
    csv_path = os.path.join(housing_path, "housing.csv")
    return pd.read_csv(csv_path)
housing=load_housing_data()
housing2=housing.copy()
import numpy as np

#预处理前去掉带文字的指定列
from sklearn.preprocessing import MinMaxScaler,StandardScaler,OneHotEncoder

from sklearn.impute import SimpleImputer
housing_num = housing2.drop("ocean_proximity", axis=1)
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
num_pipeline = Pipeline([
        ('imputer', SimpleImputer(strategy="median")),
        ('std_scaler', StandardScaler())
    ])



from sklearn.compose import ColumnTransformer
#返回所有列名
num_attribs = list(housing_num)

cat_attribs = ["ocean_proximity"]
#找出待独热编码列的最大分类数，不然在进行测试集划分处理时，
#容易造成独热向量因测试集构成不同而列数不一致的情况
categories=housing2['ocean_proximity'].unique()
full_pipeline = ColumnTransformer([
        ("num", num_pipeline, num_attribs),
        ("cat", OneHotEncoder(categories=[categories]), cat_attribs),
    ])
#抽样后的数据，去除预测目标列，并拿出对应目标列准备数据训练
housing_labels = housing2["median_house_value"]

housing_prepared = full_pipeline.fit_transform(housing2)
from sklearn.model_selection import train_test_split
train_set, test_set,train_sety, test_sety = train_test_split(housing_prepared,housing_labels, test_size=0.1, random_state=42)

预测得到特征贡献值，并用本文特征筛选器进行筛选：

from sklearn.model_selection import RandomizedSearchCV
from sklearn.ensemble import RandomForestRegressor
import matplotlib.pyplot as plt
import numpy as np

param_distribs = {
    #对于搜索范围是list的超参数，在给定的list中等概率采样
        'kernel': ['linear', 'rbf'],
        'C': reciprocal(20, 200),
        'gamma': expon(scale=1.0),
    }
forest_reg = RandomForestRegressor(random_state=42)
forest_reg.fit(train_set[:1000], train_sety[:1000])
print(forest_reg.feature_importances_)
importantArray=forest_reg.feature_importances_

from sklearn.base import BaseEstimator, TransformerMixin

def indices_of_top_k(arr, k):
    #np.argpartition ()将传入的数组arr分成 两部分 ，即：排在第k位置前面的数都小于k位置，排在第k位置后面的值都大于k位置
    #np.argpartition(np.array(arr), -k)保证最后k个数大于k位置的数,[-k:]取最后的k个值，这个组合求数组内最大的k个数
    #该函数返回的是下标，再按下标进行排序
    return np.sort(np.argpartition(np.array(arr), -k)[-k:])

class TopFeatureSelector(BaseEstimator, TransformerMixin):
    #feature_importances为特证重要度，越大重要度越高
    def __init__(self, feature_importances, k):
        self.feature_importances = feature_importances
        self.k = k
    def fit(self, X, y=None):
        self.feature_indices_ = indices_of_top_k(self.feature_importances, self.k)
        return self
    def transform(self, X):
        return X[:, self.feature_indices_]

#有了重要特征后重新训练：
full_pipeline2=  Pipeline([
        ('prepared', full_pipeline),
        ("filter",TopFeatureSelector(importantArray,5))
    ])

housing_prepared = full_pipeline2.fit_transform(housing2)
print(housing_prepared.shape)

输出为：

原始特征贡献度列表为：[1.02515161e-05 1.61509216e-05 1.76575395e-05 1.02708858e-05
 1.12920206e-05 1.58529503e-05 8.76720983e-06 1.89549348e-05
 9.99882943e-01 1.34032475e-06 2.87741167e-06 1.97157680e-06
 1.66973611e-06 0.00000000e+00]
最终特征矩阵大小为(20640, 5)