随机森林是数据挖掘中非常常用的分类预测算法，以分类或回归的决策树为基分类器。算法的一些基本要点：
*对大小为m的数据集进行样本量同样为m的有放回抽样；
*对K个特征进行随机抽样，形成特征的子集，样本量的确定方法可以有平方根、自然对数等；
*每棵树完全生成，不进行剪枝；
*每个样本的预测结果由每棵树的预测投票生成（回归的时候，即各棵树的叶节点的平均）
著名的python机器学习包scikit learn的文档对此算法有比较详尽的介绍: http://scikit-learn.org/stable/modules/ensemble.html#random-forests
出于个人研究和测试的目的，基于经典的Kaggle 101 泰坦尼克号乘客的数据集，建立模型并进行评估。比赛页面及相关数据集的下载：https://www.kaggle.com/c/titanic
泰坦尼克号的沉没，是历史上非常著名的海难。突然感到，自己面对的不再是冷冰冰的数据，而是用数据挖掘的方法，去研究具体的历史问题，也是饶有兴趣。言归正传，模型的主要的目标，是希望根据每个乘客的一系列特征，如性别、年龄、舱位、上船地点等，对其是否能生还进行预测，是非常典型的二分类预测问题。数据集的字段名及实例如下：
| PassengerId | Survived | Pclass | Name | Sex | Age | SibSp | Parch | Ticket | Fare | Cabin | Embarked |
| 1 | 0 | 3 | Braund, Mr. Owen Harris | male | 22 | 1 | 0 | A/5 21171 | 7.25 | S | |
| 2 | 1 | 1 | Cumings, Mrs. John Bradley (Florence Briggs Thayer) | female | 38 | 1 | 0 | PC 17599 | 71.2833 | C85 | C |
| 3 | 1 | 3 | Heikkinen, Miss. Laina | female | 26 | 0 | 0 | STON/O2. 3101282 | 7.925 | S | |
| 4 | 1 | 1 | Futrelle, Mrs. Jacques Heath (Lily May Peel) | female | 35 | 1 | 0 | 113803 | 53.1 | C123 | S |
| 5 | 0 | 3 | Allen, Mr. William Henry | male | 35 | 0 | 0 | 373450 | 8.05 | S |
值得说明的是，SibSp是指sister brother spouse，即某个乘客随行的兄弟姐妹、丈夫、妻子的人数，Parch指parents,children
下面给出整个数据处理及建模过程，基于ubuntu+python 3.４ （ anaconda科学计算环境已经集成一系列常用包，pandas numpy sklearn等，这里强烈推荐）
懒得切换输入法，写的时候主要的注释都是英文，中文的注释是后来补充的 :-)
# -*- coding: utf-8 -*-
"""
@author: kim
"""
from model import * ＃载入基分类器的代码
#ETL:same procedure to training set and test set
training=pd.read_csv('train.csv',index_col=0)
test=pd.read_csv('test.csv',index_col=0)
SexCode=pd.DataFrame([1,0],index=['female','male'],columns=['Sexcode']) ＃将性别转化为０１
training=training.join(SexCode,how='left',on=training.Sex)
training=training.drop(['Name','Ticket','Embarked','Cabin','Sex'],axis=1) ＃删去几个不参与建模的变量，包括姓名、船票号，船舱号
test=test.join(SexCode,how='left',on=test.Sex)
test=test.drop(['Name','Ticket','Embarked','Cabin','Sex'],axis=1)
print('ETL IS DONE!')
#MODEL FITTING
#===============PARAMETER AJUSTMENT============
min_leaf=1
min_dec_gini=0.0001
n_trees=5
n_fea=int(math.sqrt(len(training.columns)-1))
#==============================================
'''
BEST SCORE:0.83
min_leaf=30
min_dec_gini=0.001
n_trees=20
'''
#ESSEMBLE BY RANDOM FOREST
FOREST={}
tmp=list(training.columns)
tmp.pop(tmp.index('Survived'))
feaList=pd.Series(tmp)
for t in range(n_trees):
# fea=[]
feasample=feaList.sample(n=n_fea,replace=False)#select feature
fea=feasample.tolist()
fea.append('Survived')
# feaNew=fea.append(target)
subset=training.sample(n=len(training),replace=True) #generate the dataset with replacement
subset=subset[fea]
# print(str(t)+' Classifier built on feature:')
# print(list(fea))
FOREST[t]=tree_grow(subset,'Survived',min_leaf,min_dec_gini) #save the tree
#MODEL PREDICTION
#======================
currentdata=training
output='submission_rf_20151116_30_0.001_20'
#======================
prediction={}
for r in currentdata.index:#a row
prediction_vote={1:0,0:0}
row=currentdata.get(currentdata.index==r)
for n in range(n_trees):
tree_dict=FOREST[n] #a tree
p=model_prediction(tree_dict,row)
prediction_vote[p]+=1
vote=pd.Series(prediction_vote)
prediction[r]=list(vote.order(ascending=False).index)[0]#the vote result
result=pd.Series(prediction,name='Survived_p')
#del prediction_vote
#del prediction
#result.to_csv(output)
t=training.join(result,how='left')
accuracy=round(len(t[t['Survived']==t['Survived_p']])/len(t),5)
print(accuracy)
上述是随机森林的代码，如上所述，随机森林是一系列决策树的组合，决策树每次分裂，用Gini系数衡量当前节点的“不纯净度”，如果按照某个特征的某个分裂点对数据集划分后，能够让数据集的Gini下降最多（显著地减少了数据集输出变量的不纯度），则选为当前最佳的分割特征及分割点。代码如下：
# -*- coding: utf-8 -*-
"""
@author: kim
"""
import pandas as pd
import numpy as np
#import sklearn as sk
import math
def tree_grow(dataframe,target,min_leaf,min_dec_gini):
tree={} #renew a tree
is_not_leaf=(len(dataframe)>min_leaf)
if is_not_leaf:
fea,sp,gd=best_split_col(dataframe,target)
if gd>min_dec_gini:
tree['fea']=fea
tree['val']=sp
# dataframe.drop(fea,axis=1) #1116 modified
l,r=dataSplit(dataframe,fea,sp)
l.drop(fea,axis=1)
r.drop(fea,axis=1)
tree['left']=tree_grow(l,target,min_leaf,min_dec_gini)
tree['right']=tree_grow(r,target,min_leaf,min_dec_gini)
else:#return a leaf
return leaf(dataframe[target])
else:
return leaf(dataframe[target])
return tree
def leaf(class_lable):
tmp={}
for i in class_lable:
if i in tmp:
tmp[i]+=1
else:
tmp[i]=1
s=pd.Series(tmp)
s.sort(ascending=False)
return s.index[0]
def gini_cal(class_lable):
p_1=sum(class_lable)/len(class_lable)
p_0=1-p_1
gini=1-(pow(p_0,2)+pow(p_1,2))
return gini
def dataSplit(dataframe,split_fea,split_val):
left_node=dataframe[dataframe[split_fea]<=split_val]
right_node=dataframe[dataframe[split_fea]>split_val]
return left_node,right_node
def best_split_col(dataframe,target_name):
best_fea=''#modified 1116
best_split_point=0
col_list=list(dataframe.columns)
col_list.remove(target_name)
gini_0=gini_cal(dataframe[target_name])
n=len(dataframe)
gini_dec=-99999999
for col in col_list:
node=dataframe[[col,target_name]]
unique=node.groupby(col).count().index
for split_point in unique: #unique value
left_node,right_node=dataSplit(node,col,split_point)
if len(left_node)>0 and len(right_node)>0:
gini_col=gini_cal(left_node[target_name])*(len(left_node)/n)+gini_cal(right_node[target_name])*(len(right_node)/n)
if (gini_0-gini_col)>gini_dec:
gini_dec=gini_0-gini_col#decrease of impurity
best_fea=col
best_split_point=split_point
#print(col,split_point,gini_0-gini_col)
return best_fea,best_split_point,gini_dec
def model_prediction(model,row): #row is a df
fea=model['fea']
val=model['val']
left=model['left']
right=model['right']
if row[fea].tolist()[0]<=val:#get the value
branch=left
else:
branch=right
if ('dict' in str( type(branch) )):
prediction=model_prediction(branch,row)
else:
prediction=branch
return prediction
实际上，上面的代码还有很大的效率提升的空间，数据集不是很大的情况下，如果选择一个较大的输入参数，例如生成100棵树，就会显著地变慢；同时，将预测结果提交至kaggle进行评测，发现在测试集上的正确率不是很高，比使用sklearn里面相应的包进行预测的正确率（0.77512）要稍低一点 :-( 如果要提升准确率，两个大方向： 构造新的特征；调整现有模型的参数。
这里是抛砖引玉，欢迎大家对我的建模思路和算法的实现方法提出修改意见。