随机森林
集成学习一般可分为三大类：boosting，bagging，stacking，随机森林算法归属于bagging类，它的特点是使用多个没有依赖关系的弱学习器进行并行拟合，最后的决策也很简单，对于分类问题则使用简单的投票法，对于回归问题，则使用平均法。
在随机森林算法中建立每棵树的过程是：
- 1、随机在N个样本中选择一个样本，重复N次（样本是有可能重复的）
- 2、随机在M个特征中选择m个特征，不同于普通的决策树选择信息增益最大或者根据基尼系数等选择特征
注意使用随机森林算法的采样方式与与GBDT算法不同，GBDT算法采用的是无放回采样，而随机森林算法采用的是有放回采样。
采用有放回采样方式可以保证随机森林的每棵树之间都有交集，否则每棵决策树都有可能是“有偏的”。
随机森林算法的优势：
- 1、能够有效的运行在大数据集上，精度高
- 2、能够处理高维数据而不需要降维处理
- 3、采用随机采样，训练出的模型的方差小，泛化能力强
- 4、能够评估各个特征在分类问题上的重要性
- 5、对缺失值不敏感
承接上篇博客，对2万个数据进行算法验证，并比较CART、GBDT及随机森林算法。
导入数据
import pandas as pd
df = pd.read_csv("./train_modified.csv")
df.head()
| Disbursed | Existing_EMI | ID | Loan_Amount_Applied | Loan_Tenure_Applied | Monthly_Income | Var4 | Var5 | Age | EMI_Loan_Submitted_Missing | ... | Var2_2 | Var2_3 | Var2_4 | Var2_5 | Var2_6 | Mobile_Verified_0 | Mobile_Verified_1 | Source_0 | Source_1 | Source_2 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | 0 | 0.0 | ID000002C20 | 300000 | 5 | 20000 | 1 | 0 | 37 | 1 | ... | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 1 | 0 | 0 |
| 1 | 0 | 0.0 | ID000004E40 | 200000 | 2 | 35000 | 3 | 13 | 30 | 0 | ... | 0 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 0 | 0 |
| 2 | 0 | 0.0 | ID000007H20 | 600000 | 4 | 22500 | 1 | 0 | 34 | 1 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 1 |
| 3 | 0 | 0.0 | ID000008I30 | 1000000 | 5 | 35000 | 3 | 10 | 28 | 1 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 1 |
| 4 | 0 | 25000.0 | ID000009J40 | 500000 | 2 | 100000 | 3 | 17 | 31 | 1 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 1 |
5 rows × 51 columns
这里我们以 Disbursed 这一列作为分类结果，从结果上看是二分类问题，观察样本发现样本分布不均匀，这时候对分类的好坏评估就需要使用 AUC 评估参数了
df["Disbursed"].value_counts()
0 19680
1 320
Name: Disbursed, dtype: int64
数据可视化（随便选了两个特征，似乎不明显-_-!）
from sklearn.model_selection import train_test_split
from matplotlib import pyplot as plt
%matplotlib inline
x_columns = [x for x in df.columns if x not in ["Disbursed", "ID"]] # 挑选除了Disbursed、ID这两列的数据
X = df[x_columns]
y = df["Disbursed"]
# 划分数据集
x_train, x_test, y_train, y_test = train_test_split(X, y, test_size=0.25)
# 数据可视化
fig = plt.figure()
plt.scatter(x_train[y_train==0]["Loan_Tenure_Applied"], x_train[y_train==0]["Var4"])
plt.scatter(x_train[y_train==1]["Loan_Tenure_Applied"], x_train[y_train==1]["Var4"])
plt.legend([0, 1])
plt.show()
使用单个决策树CART模型，不调任何参数，观察结果， 发现尽管分类的精度很高，但是在样本分布不均匀的情况下，AUC得分接近0.5，说明这个分类器性能很差
from sklearn.tree import DecisionTreeClassifier
from sklearn.metrics import roc_auc_score, classification_report
dtc = DecisionTreeClassifier()
dtc.fit(x_train, y_train)
accuracy = dtc.score(x_test, y_test)
print("Accuracy (test): \n", accuracy)
y_pred = dtc.predict(x_test)
print("混淆矩阵：\n", classification_report(y_test, y_pred))
y_predprob = dtc.predict_proba(x_test)[:, 1]
print("AUC Score (test): %f" % roc_auc_score(y_test, y_predprob))
Accuracy (test):
0.9658
混淆矩阵：
precision recall f1-score support
0 0.99 0.98 0.98 4926
1 0.05 0.07 0.06 74
avg / total 0.97 0.97 0.97 5000
AUC Score (test): 0.523336
建立集成学习算法模型，首先是GBDT模型，不调参数，使用默认的参数设置，观察结果
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.metrics import roc_auc_score, classification_report
gbc = GradientBoostingClassifier()
gbc.fit(x_train, y_train)
accuracy = gbc.score(x_test, y_test)
print("Accuracy (test): \n", accuracy)
y_pred = gbc.predict(x_test)
print("混淆矩阵：\n", classification_report(y_test, y_pred))
y_predprob = gbc.predict_proba(x_test)[:, 1]
print("AUC Score (test): %f" % roc_auc_score(y_test, y_predprob))
Accuracy (test):
0.9848
混淆矩阵：
precision recall f1-score support
0 0.99 1.00 0.99 4926
1 0.00 0.00 0.00 74
avg / total 0.97 0.98 0.98 5000
AUC Score (test): 0.824064
建立集成学习算法模型，使用随机森林模型，不调参数，使用默认的参数设置，观察结果，发现袋外分数和精度都挺高，但是AUC得分还是接近0.5，所以分类模型并不是很好
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import roc_auc_score, classification_report
rfc = RandomForestClassifier(oob_score=True)
rfc.fit(x_train, y_train)
accuracy = rfc.score(x_test, y_test)
print("Accuracy (test): \n", accuracy)
y_pred = rfc.predict(x_test)
print("混淆矩阵(test)：\n", classification_report(y_test, y_pred))
y_predprob = rfc.predict_proba(x_test)[:, 1]
print("AUC Score (test): %f" % roc_auc_score(y_test, y_predprob))
print("袋外分数:\n", rfc.oob_score_)
Accuracy (test):
0.9852
混淆矩阵(test)：
precision recall f1-score support
0 0.99 1.00 0.99 4926
1 0.50 0.01 0.03 74
avg / total 0.98 0.99 0.98 5000
AUC Score (test): 0.603095
袋外分数:
0.9803333333333333
D:\anaconda\setup\lib\site-packages\sklearn\ensemble\forest.py:453: UserWarning: Some inputs do not have OOB scores. This probably means too few trees were used to compute any reliable oob estimates.
warn("Some inputs do not have OOB scores. "
D:\anaconda\setup\lib\site-packages\sklearn\ensemble\forest.py:458: RuntimeWarning: invalid value encountered in true_divide
predictions[k].sum(axis=1)[:, np.newaxis])
对随机森林算参数进行调参，发现精度并没有提高，但是AUC得分提高了很多
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import roc_auc_score, classification_report
from sklearn.model_selection import GridSearchCV
rfc = RandomForestClassifier(oob_score=True, max_features="sqrt")
params = {"max_depth": list(range(3,15, 2)),
"n_estimators": list(range(50, 201, 20)),
'min_samples_split': list(range(80,150,20)),
'min_samples_leaf': list(range(10,60,10))}
gs = GridSearchCV(estimator=rfc, param_grid=params, cv=5)
gs.fit(x_train, y_train)
accuracy = gs.score(x_test, y_test)
print("Accuracy (test): \n", accuracy)
y_pred = gs.predict(x_test)
print("混淆矩阵(test)：\n", classification_report(y_test, y_pred))
y_predprob = gs.predict_proba(x_test)[:, 1]
print("AUC Score (test): %f" % roc_auc_score(y_test, y_predprob))
Accuracy (test):
0.9852
混淆矩阵(test)：
precision recall f1-score support
0 0.99 1.00 0.99 4926
1 0.00 0.00 0.00 74
avg / total 0.97 0.99 0.98 5000
AUC Score (test): 0.803139
D:\anaconda\setup\lib\site-packages\sklearn\metrics\classification.py:1135: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples.
'precision', 'predicted', average, warn_for)
gs.best_estimator_, gs.best_score_, gs.best_params_
(RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
max_depth=3, max_features='sqrt', max_leaf_nodes=None,
min_impurity_decrease=0.0, min_impurity_split=None,
min_samples_leaf=10, min_samples_split=80,
min_weight_fraction_leaf=0.0, n_estimators=50, n_jobs=1,
oob_score=True, random_state=None, verbose=0, warm_start=False),
0.9836,
{'max_depth': 3,
'min_samples_leaf': 10,
'min_samples_split': 80,
'n_estimators': 50})
最后使用网格搜索后的最优超参数进行训练，结果如下
rfc2 = RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
max_depth=3, max_features='sqrt', max_leaf_nodes=None,
min_impurity_decrease=0.0, min_impurity_split=None,
min_samples_leaf=10, min_samples_split=80,
min_weight_fraction_leaf=0.0, n_estimators=50, n_jobs=1,
oob_score=True, random_state=None, verbose=0, warm_start=False)
rfc2.fit(x_train, y_train)
accuracy = rfc2.score(x_test, y_test)
print("Accuracy (test): \n", accuracy)
y_pred = rfc2.predict(x_test)
print("混淆矩阵(test)：\n", classification_report(y_test, y_pred))
y_predprob = rfc2.predict_proba(x_test)[:, 1]
print("AUC Score (test): %f" % roc_auc_score(y_test, y_predprob))
print("袋外分数:\n", rfc2.oob_score_)
Accuracy (test):
0.9836
混淆矩阵(test)：
precision recall f1-score support
0 0.98 1.00 0.99 4918
1 0.00 0.00 0.00 82
avg / total 0.97 0.98 0.98 5000
AUC Score (test): 0.808931
袋外分数:
0.9841333333333333
D:\anaconda\setup\lib\site-packages\sklearn\metrics\classification.py:1135: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples.
'precision', 'predicted', average, warn_for)