【Python】决策树算法(DecisionTreeClassifier)——东北大学数据挖掘实训三

1.利用决策树算法对train_feature.csv进行训练对test_feature.csv进行预测（练习调参）,并计算预测正确的准确率。（由于训练数据的数据分布不平衡，所以将数据进行平衡处理，正样本保留，负样本从原负样本中取出正样本数量的n倍）说明：准确率=（测试集中预测为下载）&（测试集中实际为下载）/测试集实际为下载的数据量

import pandas as pd
from sklearn import datasets
from sklearn.tree import DecisionTreeClassifier
from sklearn.model_selection import train_test_split,GridSearchCV
import time,datetime

train_df=pd.read_csv("C:\\Users\\zzh\\Desktop\\dataMiningExperment\\数据挖掘实训课件\\数据挖掘第3次实训\\数据\\训练和预测用数据--做题用\\train_feature.csv")

train_df.head()

	ip	app	device	os	channel	is_attributed	day	hour	minute	ip_count	app_count	device_count	os_count	channel_count	hour_count	minute_count
0	83230	3	1	13	379	0	2017-11-06	14	32	938	774123	6527713	1541988	101195	48	110457
1	17357	3	1	19	379	0	2017-11-06	14	33	677	774123	6527713	1644220	101195	48	112948
2	35810	3	1	13	379	0	2017-11-06	14	34	351	774123	6527713	1541988	101195	48	112532
3	45745	14	1	13	478	0	2017-11-06	14	34	7786	316214	6527713	1541988	11355	48	112532
4	161007	3	1	13	379	0	2017-11-06	14	35	132	774123	6527713	1541988	101195	48	115570

test_df=pd.read_csv("C:\\Users\\zzh\\Desktop\\dataMiningExperment\\数据挖掘实训课件\\数据挖掘第3次实训\\数据\\训练和预测用数据--做题用\\test_feature.csv")

test_df.head()

	click_id	ip	app	device	os	channel	is_attributed	day	hour	minute	ip_count	app_count	device_count	os_count	channel_count	hour_count	minute_count
0	0	19870	2	1	13	435	0	2017-11-06	23	1	99	308059	2853433	657790	42678	2308568	68675
1	1	50314	15	1	17	265	0	2017-11-06	23	1	233	307505	2853433	153419	68057	2308568	68675
2	2	183513	15	1	13	153	0	2017-11-06	23	1	105	307505	2853433	657790	104935	2308568	68675
3	3	35731	12	1	19	178	0	2017-11-06	23	1	550	348786	2853433	765928	89744	2308568	68675
4	4	186444	12	1	3	265	0	2017-11-06	23	1	16	348786	2853433	45955	68057	2308568	68675

（由于训练数据的数据分布不平衡，所以将数据进行平衡处理，正样本保留，负样本从原负样本中取出正样本数量的n倍）

train_df["is_attributed"].value_counts()

0    6986725
1      13275
Name: is_attributed, dtype: int64

tmp_is1 = train_df[train_df['is_attributed']==1] #13275
tmp_is0 = train_df[train_df['is_attributed']==0] #6986725
tmp_is0 = tmp_is0.sample(n=tmp_is1.shape[0]*5)
train_df= tmp_is1.append(tmp_is0) #合并

删除‘day’列

print(train_df["day"].value_counts()) #都是同一天，没有用，删掉
print(test_df["day"].value_counts()) 

2017-11-06    79650
Name: day, dtype: int64
2017-11-06    2308568
2017-11-07     691432
Name: day, dtype: int64

train_df1=train_df.drop(['day'],axis=1)
test_df1=test_df.drop(['day'],axis=1)
test_df1=test_df1.drop(['click_id'],axis=1)

# 分割数据，取两个数据特征做为训练数据的特征,测试时发现如何将四个特征都做用起来，
# 准确率基本为 1，这样反而不方便调试了
y_train=train_df1[['is_attributed']].values
y_test=test_df1[['is_attributed']].values
x_train=train_df1.drop(['is_attributed'],axis=1)
x_test=test_df1.drop(['is_attributed'],axis=1)

# 初始化模型, max_depth 限制树的最大深度
# 可以使用"gini"或者"entropy"，前者代表基尼系数，后者代表信息增益。
# 一般说使用默认的基尼系数"gini"就可以了，即CART算法。除非你更喜欢类似ID3, C4.5的最优特征选择方法。
clf = DecisionTreeClassifier()

# 训练模型
clf.fit(x_train, y_train)

DecisionTreeClassifier(class_weight=None, criterion='gini', max_depth=None,
                       max_features=None, max_leaf_nodes=None,
                       min_impurity_decrease=0.0, min_impurity_split=None,
                       min_samples_leaf=1, min_samples_split=2,
                       min_weight_fraction_leaf=0.0, presort=False,
                       random_state=None, splitter='best')

print("训练数据的score：", (clf.score(x_train, y_train)))
print("测试数据的score：", (clf.score(x_test, y_test)))

训练数据的score： 0.9996359070935342
测试数据的score： 0.8894056666666666

predict=clf.predict(x_test) #预测

submission = pd.DataFrame ( {
    'click_id':test_df['click_id'],
    'is_attributed':predict
} )

# submission.to_csv('submission.csv',index=False)

说明：准确率=（测试集中预测为下载）&（测试集中实际为下载）/测试集实际为下载的数据量

print("准确率:",sum((predict == 1) & (test_df.is_attributed==1)) / sum(test_df.is_attributed==1))

准确率: 0.7458654906284454

# 笨的求准确率的办法
# denominator=test_df[test_df.is_attributed==1].shape[0]  #分母
# denominator
# test_df['is_attributed']=test_df['is_attributed'].replace(0,2) 
# submission['is_attributed']=submission['is_attributed'].replace(0,3) 
# molecule=(test_df['is_attributed']==submission['is_attributed']).sum() ##分子
# molecule
# Precision=molecule/denominator
# print(Precision)  #准确率