参考资料

• 关于理论推理部分参考刘建平老师的博客决策树算法原理
• python实现以及参数含义参考scikit-learn决策树算法类库使用小结

算法实现

导入包

• 最后一行用于设置graphviz插件的环境变量

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from sklearn import tree
from sklearn.datasets import load_iris
from sklearn import datasets
from sklearn.tree import DecisionTreeClassifier
import sys
import os
import warnings
warnings.filterwarnings('ignore')
os.environ["PATH"] += os.pathsep + 'D:/Python/Graphviz/bin'

数据准备

• 使用sklearn包中自带鸢尾花数据集。

# 载入鸢尾花数据集
iris = load_iris()
X = iris.data
y = iris.target
X.shape

模型建立与调参

clf = DecisionTreeClassifier()
from sklearn.model_selection import GridSearchCV
param_grid = [{'criterion':['gini','entropy'],'splitter':['best', 'random'],'max_depth':np.arange(5,21,3),'class_weight':['balanced'],'max_leaf_nodes':np.arange(2,6,1)}]
grid_search = GridSearchCV(clf, param_grid, cv = 10,scoring = 'f1_micro',return_train_score = True)
grid_search.fit(X,np.ravel(y))

模型最优参

grid_search.best_params_

输出：

{'class_weight': 'balanced','criterion': 'gini','max_depth': 11,'max_leaf_nodes': 5,'splitter': 'random'}

各参数下模型评估

cvres = grid_search.cv_results_
for accuracy,params in zip(cvres["mean_test_score"],cvres["params"]):print("{:.2}".format(accuracy),params)

输出：

0.67 {'class_weight': 'balanced', 'criterion': 'gini', 'max_depth': 5, 'max_leaf_nodes': 2, 'splitter': 'best'}
0.66 {'class_weight': 'balanced', 'criterion': 'gini', 'max_depth': 5, 'max_leaf_nodes': 2, 'splitter': 'random'}
0.95 {'class_weight': 'balanced', 'criterion': 'gini', 'max_depth': 5, 'max_leaf_nodes': 3, 'splitter': 'best'}
0.79 {'class_weight': 'balanced', 'criterion': 'gini', 'max_depth': 5, 'max_leaf_nodes': 3, 'splitter': 'random'}
...

保存模型训练结果

final_model = grid_search.best_estimator_
# 保存模型
with open("iris.dot", 'w') as f:f = tree.export_graphviz(final_model, out_file=f)

决策树可视化

import pydotplus 
from IPython.display import Image  
dot_data = tree.export_graphviz(final_model, out_file=None, feature_names=iris.feature_names,  class_names=iris.target_names,  filled=True, rounded=True,  special_characters=True)  
graph = pydotplus.graph_from_dot_data(dot_data)  
Image(graph.create_png()) 
#保存为pdf文件
#graph.write_pdf("DTtree.pdf") 

输出：