python-sklearn中RandomForestClassifier函数以及ROC曲线绘制

先转自博主http://blog.itpub.net/12199764/viewspace-1572056/

介绍一下RandomForestClassifier函数的简单用法

# -*- coding: utf-8 -*-
from sklearn.tree import DecisionTreeClassifier
from matplotlib.pyplot import *
from sklearn.cross_validation import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.externals.joblib import Parallel, delayed
from sklearn.tree import export_graphviz

final = open('c:/test/final.dat' , 'r')
data = [line.strip().split('\t') for line in final]
feature = [[float(x) for x in row[3:]] for row in data]
target = [int(row[0]) for row in data]

#拆分训练集和测试集
feature_train, feature_test, target_train, target_test = train_test_split(feature, target, test_size=0.1, random_state=42)

#分类型决策树
clf = RandomForestClassifier(n_estimators = 8)

#训练模型
s = clf.fit(feature_train , target_train)
print s

#评估模型准确率
r = clf.score(feature_test , target_test)
print r

print '判定结果：%s' % clf.predict(feature_test[0])
#print clf.predict_proba(feature_test[0])

print '所有的树:%s' % clf.estimators_

print clf.classes_
print clf.n_classes_

print '各feature的重要性：%s' % clf.feature_importances_

print clf.n_outputs_

def _parallel_helper(obj, methodname, *args, **kwargs):
    return getattr(obj, methodname)(*args, **kwargs)

all_proba = Parallel(n_jobs=10, verbose=clf.verbose, backend="threading")(
            delayed(_parallel_helper)(e, 'predict_proba', feature_test[0]) for e in clf.estimators_)
print '所有树的判定结果：%s' % all_proba

proba = all_proba[0]
for j in range(1, len(all_proba)):
    proba += all_proba[j]
proba /= len(clf.estimators_)
print '数的棵树：%s ， 判不作弊的树比例：%s' % (clf.n_estimators , proba[0,0])
print '数的棵树：%s ， 判作弊的树比例：%s' % (clf.n_estimators , proba[0,1])

#当判作弊的树多余不判作弊的树时，最终结果是判作弊
print '判断结果：%s' % clf.classes_.take(np.argmax(proba, axis=1), axis=0)

#把所有的树都保存到word
for i in xrange(len(clf.estimators_)):
    export_graphviz(clf.estimators_[i] , '%d.dot'%i)

楼上博主将模型保存在dot文件中，还可以将模型保存在pkl文件中，这部分是转自 https://www.zhihu.com/question/31604690/answer/52708757
import pickle
model = RandomForestClassifier(n_estimators=100)
model.fit(x, y) # 建立随机森林，其中x,y分别为训练集和训练集的标签

‘’‘
然后就可以把训练好的随机森林存进pickle中了，省去了每次训练，其实并不是所有的类都能存进pickle， 但随机森林是可以的，这里不局限于存一个model，你可以把训练集、训练集的标签、随机森林放到一个list一起扔进pickle
’‘’
with open("data.pkl", "wb") as f:
pickle.dump(model, f)
# 如果下次要用就再读取
with open("data.pkl", "rb") as f:
model = pickle.load(f)

下面是自己写的代码片段

import matplotlib.pyplot as plt
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import roc_curve, auc
import pickle
df_train = pd.read_csv(utils.HEATMAP_FEATURE_CSV_TRAIN)
df_validation = pd.read_csv(utils.HEATMAP_FEATURE_CSV_VALIDATION)
n_columns = len(df_train.columns)
feature_column_names = df_train.columns[:n_columns - 1]#返回的是每个列表的第一行，对应的是特征的名字
label_column_name = df_train.columns[n_columns - 1]
train_x = df_train[feature_column_names]
train_y = df_train[label_column_name]
validation_x = df_validation[feature_column_names]
validation_y = df_validation[label_column_name]
clf = RandomForestClassifier(n_estimators=50, n_jobs=2)  #分类型决策树
s = clf.fit(train_x, train_y) # 训练模型
r = clf.score(validation_x,validation_y) #评估模型准确率
print r
predict_y_validation = clf.predict(validation_x)#直接给出预测结果，每个点在所有label的概率和为1，内部还是调用predict——proba()
# print(predict_y_validation)
prob_predict_y_validation = clf.predict_proba(validation_x)#给出带有概率值的结果，每个点所有label的概率和为1
predictions_validation = prob_predict_y_validation[:, 1]
fpr, tpr, _ = roc_curve(validation_y, predictions_validation)
    #
roc_auc = auc(fpr, tpr)
plt.title('ROC Validation')
plt.plot(fpr, tpr, 'b', label='AUC = %0.2f' % roc_auc)
plt.legend(loc='lower right')
plt.plot([0, 1], [0, 1], 'r--')
plt.xlim([0, 1])
plt.ylim([0, 1])
plt.ylabel('True Positive Rate')
plt.xlabel('False Positive Rate')
plt.show()

#
#

#

#