决策树

步骤

• 1.计算不纯度
• 2.选取不纯度最高的特征进行分支
• 3.计算不纯度
• 4.继续划分

from sklearn import tree
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import cross_val_score
from sklearn.model_selection import train_test_split
from sklearn.decomposition import PCA
from imblearn.over_sampling import RandomOverSampler
from sklearn.preprocessing import KBinsDiscretizer
from imblearn.over_sampling import SMOTE
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
# 导入自己写的工具类
from my_tools import *
# 忽略warning
import warnings
warnings.filterwarnings("ignore")

加载数据

jibing_res = pd.read_excel("./jibing_feature_res_final.xlsx")
jibing = pd.read_excel("./jibing_feature_final.xlsx")

jibing.head()

	左右	是否外伤	症状持续时间	明显夜间痛	年龄	高血压	高血脂	2型糖尿病	吸烟与否	饮酒与否	...	腺苷脱氨酶ADA	果糖胺	肌酸激酶	α-L-盐藻糖苷酶	乳酸	淀粉酶	同型半胱氨酸	铁	总铁结合力	血型
0	0	0	3	0	65	1	0	0	0	0	...	10.0	1.32	48.0	12.0	1.9	49.0	9.9	12.3	43.5	3
1	1	1	2	0	62	1	0	0	0	0	...	10.0	1.67	77.0	16.0	1.4	81.0	9.2	16.9	55.5	0
2	1	0	4	1	55	0	0	0	0	0	...	15.0	1.86	78.0	22.0	1.9	89.0	9.9	7.0	51.4	0
3	1	0	3	0	60	0	0	0	0	0	...	16.0	1.68	92.0	12.0	1.4	69.0	9.3	15.8	53.0	0
4	0	1	3	0	61	0	0	0	0	0	...	13.0	1.60	58.0	14.0	1.7	153.0	8.1	13.2	45.9	0

5 rows × 60 columns

尝试特征筛选，发现 f1-score 太低

from sklearn.feature_selection import SelectKBest
from sklearn.feature_selection import chi2
from imblearn.over_sampling import SMOTE
from sklearn.feature_selection import mutual_info_classif
f1_list = []
best_k = -1
best_score = -1
set_font()
for i in range(1,60):
#     sampler = RandomOverSampler(sampling_strategy=0.2, random_state=42)
    smote = SMOTE(sampling_strategy=1, random_state=42)
    selector = SelectKBest(mutual_info_classif, k=i)
    jibing_ = selector.fit_transform(jibing, jibing_res)
    Xtrain,Xtest,Ytrain,Ytest = train_test_split(jibing_,jibing_res,test_size=0.3,random_state=42)
#     Xtrain, Ytrain = sampler.fit_resample(Xtrain,Ytrain)
    Xtrain, Ytrain = smote.fit_resample(Xtrain,Ytrain)
    clf = tree.DecisionTreeClassifier(random_state=42)
    clf.fit(Xtrain, Ytrain)
    y_pre = clf.predict(Xtest)
    metrics_ = res_metrics(Ytest,y_pre,"调参")
    f1_list.append(metrics_["f1-score"])
    if best_score < metrics_["f1-score"]:
        best_k = i
        best_score = metrics_["f1-score"]
zhexiantu(range(1,60),f1_list,"f1 - 特征筛选")

分箱，寻找最佳的分箱参数

• 方法有等频，等间隔，kmeans三种
• 还有就是要确定分箱的数量
• 使用一种自定义的网格搜索方法

best_method = "s"
best_num = -1
best_score = 0
for method in ["uniform","quantile","kmeans"]:
    for num in np.linspace(3,100,10,dtype = int):
        jibing_res = pd.read_excel("./jibing_feature_res_final.xlsx")
        jibing = pd.read_excel("./jibing_feature_final.xlsx")
        col = jibing.columns.tolist()
        col = col[10:59]
        col.append("年龄")
        est = KBinsDiscretizer(n_bins=num, encode='ordinal', strategy=method)
        est.fit(jibing[col])
        jibing[col] = est.transform(jibing[col])
        
        sampler = SMOTE(sampling_strategy=1, random_state=42)
#         sampler = RandomOverSampler(sampling_strategy=1, random_state=42)
        Xtrain,Xtest,Ytrain,Ytest = train_test_split(jibing,jibing_res,test_size=0.3,random_state=42)
        Xtrain, Ytrain = sampler.fit_resample(Xtrain,Ytrain)
        
        clf = tree.DecisionTreeClassifier(random_state=42)
        clf.fit(Xtrain,Ytrain)
        y_pre = clf.predict(Xtest)
        
        metrics_ = res_metrics(Ytest,y_pre,"调参")
        if metrics_["f1-score"] > best_score:
            best_num = num
            best_method = method
            best_score = metrics_["f1-score"]

最佳的分箱方法是quantile，最佳的分箱数目为89

print("best_score={}\nbest_num={}\nbest_method={}".format(best_score,best_num,best_method))

best_score=0.48395599537538303
best_num=89
best_method=quantile

连续型数据分箱后的结果

jibing.head()

	左右	是否外伤	症状持续时间	明显夜间痛	年龄	高血压	高血脂	2型糖尿病	吸烟与否	饮酒与否	...	腺苷脱氨酶ADA	果糖胺	肌酸激酶	α-L-盐藻糖苷酶	乳酸	淀粉酶	同型半胱氨酸	铁	总铁结合力	血型
0	0	0	3	0	25.0	1	0	0	0	0	...	6.0	1.0	6.0	4.0	15.0	16.0	29.0	16.0	9.0	3
1	1	1	2	0	22.0	1	0	0	0	0	...	6.0	21.0	32.0	8.0	9.0	49.0	21.0	42.0	48.0	0
2	1	0	4	1	15.0	0	0	0	0	0	...	11.0	40.0	33.0	17.0	15.0	56.0	29.0	1.0	32.0	0
3	1	0	3	0	20.0	0	0	0	0	0	...	12.0	22.0	46.0	4.0	9.0	39.0	22.0	35.0	38.0	0
4	0	1	3	0	21.0	0	0	0	0	0	...	9.0	16.0	13.0	6.0	13.0	68.0	10.0	20.0	14.0	0

5 rows × 60 columns

jibing_res = pd.read_excel("./jibing_feature_res_final.xlsx")
jibing = pd.read_excel("./jibing_feature_final.xlsx")
col = jibing.columns.tolist()
col = col[10:59]
col.append("年龄")
est = KBinsDiscretizer(n_bins=89, encode='ordinal', strategy="quantile")
est.fit(jibing[col])
jibing[col] = est.transform(jibing[col])

稍微高了一些，但没高太多

sampler = SMOTE(sampling_strategy=1, random_state=42)
Xtrain,Xtest,Ytrain,Ytest = train_test_split(jibing,jibing_res,test_size=0.3,random_state=42)
Xtrain, Ytrain = sampler.fit_resample(Xtrain,Ytrain)

clf = tree.DecisionTreeClassifier(random_state=42)
clf.fit(Xtrain,Ytrain)
y_pre = clf.predict(Xtest)

metrics_ = res_metrics(Ytest,y_pre,"分箱后f1")

######################分箱后f1#######################
+--------------------+--------------------+---------------------+
|     precision      |       recall       |          f1         |
+--------------------+--------------------+---------------------+
| 0.8112885044964844 | 0.3448275862068966 | 0.48395599537538303 |
+--------------------+--------------------+---------------------+

特征选择

from sklearn.feature_selection import SelectKBest
from sklearn.feature_selection import chi2
from imblearn.over_sampling import SMOTE
from sklearn.feature_selection import mutual_info_classif
f1_list = []
best_k = -1
best_score = -1
set_font()
for i in range(1,60):
#     sampler = RandomOverSampler(sampling_strategy=0.2, random_state=42)
    smote = SMOTE(sampling_strategy=1, random_state=42)
    selector = SelectKBest(mutual_info_classif, k=i)
    jibing_ = selector.fit_transform(jibing, jibing_res)
    Xtrain,Xtest,Ytrain,Ytest = train_test_split(jibing_,jibing_res,test_size=0.3,random_state=42)
#     Xtrain, Ytrain = sampler.fit_resample(Xtrain,Ytrain)
    Xtrain, Ytrain = smote.fit_resample(Xtrain,Ytrain)
    clf = tree.DecisionTreeClassifier(random_state=42)
    clf.fit(Xtrain, Ytrain)
    y_pre = clf.predict(Xtest)
    metrics_ = res_metrics(Ytest,y_pre,"调参")
    f1_list.append(metrics_["f1-score"])
    if best_score < metrics_["f1-score"]:
        best_k = i
        best_score = metrics_["f1-score"]
zhexiantu(range(1,60),f1_list,"f1 - 特征筛选")

最佳的选择方案是选前4个特征

best_k

4

smote = SMOTE(sampling_strategy=1, random_state=42)
selector = SelectKBest(mutual_info_classif, k=4)
jibing_ = selector.fit_transform(jibing, jibing_res)
Xtrain,Xtest,Ytrain,Ytest = train_test_split(jibing_,jibing_res,test_size=0.3,random_state=42)

#     Xtrain, Ytrain = sampler.fit_resample(Xtrain,Ytrain)
Xtrain, Ytrain = smote.fit_resample(Xtrain,Ytrain)
clf = tree.DecisionTreeClassifier(random_state=42)
clf.fit(Xtrain, Ytrain)

y_pre = clf.predict(Xtest)
metrics_ = res_metrics(Ytest,y_pre,"best_k - f1")

###################best_k - f1####################
+--------------------+--------------------+---------------------+
|     precision      |       recall       |          f1         |
+--------------------+--------------------+---------------------+
| 0.8060573452934983 | 0.3275862068965517 | 0.46584884247907565 |
+--------------------+--------------------+---------------------+

PCA 降维

效果并不好

from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
f1_list = []
for i in range(1,3):
    clf = tree.DecisionTreeClassifier(random_state=42)
    pca = PCA(n_components=i,random_state=42)
    Xtrain_ = pca.fit_transform(Xtrain,Ytrain)
    clf.fit(Xtrain_, Ytrain)
    Xtest_ = pca.fit_transform(Xtest)
    y_pre = clf.predict(Xtest_)
    metrics_ = res_metrics(Ytest,y_pre,"调参")
    f1_list.append(metrics_["f1-score"])
zhexiantu(range(1,3),f1_list,"f1 - PCA")

TSNE 降维，也没能达到指标

f1_list = []
for i in range(1,4):
    clf = tree.DecisionTreeClassifier(random_state=42)
    tsne = TSNE(n_components=i,random_state=42)
    Xtrain_ = tsne.fit_transform(Xtrain,Ytrain)
    clf.fit(Xtrain_, Ytrain)
    Xtest_ = tsne.fit_transform(Xtest)
    y_pre = clf.predict(Xtest_)
    metrics_ = res_metrics(Ytest,y_pre,"调参")
    f1_list.append(metrics_["f1-score"])

zhexiantu(range(1,4),f1_list,"tsne - F1")

结果为0.51 还需要继续调参

clf = tree.DecisionTreeClassifier(random_state=42)
tsne = TSNE(n_components=1,random_state=42)
Xtrain_ = tsne.fit_transform(Xtrain,Ytrain)
clf.fit(Xtrain_, Ytrain)
Xtest_ = tsne.fit_transform(Xtest)
y_pre = clf.predict(Xtest_)
metrics_ = res_metrics(Ytest,y_pre,"TSNE-f1")

#####################TSNE-f1######################
+-------------------+--------------------+--------------------+
|     precision     |       recall       |         f1         |
+-------------------+--------------------+--------------------+
| 0.802205521848379 | 0.3793103448275862 | 0.5150753564926158 |
+-------------------+--------------------+--------------------+

调参与剪枝

criterion

用于不纯度的计算，不纯度越低，对训练集的拟合越好。

entropy信息熵：对不纯度的乘法更强，决策树的生长更加精确但容易过拟合，出现欠拟合现象时使用。

gini基尼系数：计算速度较快，通常使用基尼系数。

max_depth

树的最大深度

min_samples_split

每个节点中至少包含的节点数，少于这个数值将不再继续分枝

min_samples_leaf

叶子节点至少包含的样本数量

# 搜索最佳参数
f1_list = []
best_f1 = -1.1
best_max_d = -1
best_min_sl = -1
best_min_ss = -1
for max_d in np.linspace(1,30,30,dtype=int):
    for min_sl in np.linspace(1,20,10,dtype=int):
        for min_ss in np.linspace(2,20,10,dtype=int):
            clf = DecisionTreeClassifier(max_depth=max_d,min_samples_leaf=min_sl,min_samples_split=min_ss)
            clf.fit(Xtrain_, Ytrain)
            y_pre = clf.predict(Xtest_)
            metrics_ = res_metrics(Ytest,y_pre,"调参")
            if best_f1 < metrics_["f1-score"]:
                best_max_d = max_d
                best_min_sl = min_sl
                best_min_ss = min_ss
                best_f1 = metrics_["f1-score"]
            f1_list.append(metrics_["f1-score"])
zhexiantu(np.linspace(1,3000,3000),f1_list,"params - F1")

出现上面这种趋势的原因：树的最大深度达到一定限制之后

再加大最大深度没有意义，变成f1随着其他两个属性呈周期性变化。

print("best_f1:{}\nmax_depth:{}\nmin_samples_leaf:{}\nmin_samples_split:{}".format(best_f1,best_max_d,best_min_sl,best_min_ss))

best_f1:0.6529176934256228
max_depth:10
min_samples_leaf:20
min_samples_split:2

结果提升了很多

f1-score 为0.65

clf = DecisionTreeClassifier(max_depth=10,min_samples_leaf=20,min_samples_split=2)
clf.fit(Xtrain_, Ytrain)
y_pre = clf.predict(Xtest_)
metrics_ = res_metrics(Ytest,y_pre,"DT-final")

#####################DT-final#####################
+-------------------+--------------------+--------------------+
|     precision     |       recall       |         f1         |
+-------------------+--------------------+--------------------+
| 0.799569364630173 | 0.5517241379310345 | 0.6529176934256228 |
+-------------------+--------------------+--------------------+

可视化

import graphviz
dot_data = tree.export_graphviz(clf,
                                filled=True,#填充颜色
                                rounded=True,#框的形状
                                out_file="./tree.dot",#用于生成决策树图片的文件
                                fontname="Microsoft YaHei"#设置字体，否则会乱码
)
graph = graphviz.Source(dot_data)

最终的决策树