# coding=utf8
%matplotlib inline
import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestClassifier as rf
from sklearn.cross_validation import cross_val_score
from sklearn.linear_model import LogisticRegression as lr
import matplotlib.pyplot as plt

def show_importance_of_sub_model(total_size,factor_size,special_percent):
    each_dt_size = total_size/factor_size
    dt = pd.DataFrame()
    y = pd.DataFrame()
    special_size = int(factor_size*special_percent)
    special = range(special_size)

    for i in range(factor_size):
        dt_current = pd.DataFrame(np.random.random((each_dt_size,5)))
        dt_current[‘type‘] = ‘type‘+str(100+i)
        y_current = pd.DataFrame(np.ones((each_dt_size,)))

        if i in special:
            y_current[(dt_current[0]+dt_current[1]+dt_current[2])>0.6] = -1
        else:
            y_current[(dt_current[0]+dt_current[1]+dt_current[2])>1.5] = -1

        y = pd.concat((y,y_current))
        dt = pd.concat((dt,dt_current))

    type = pd.get_dummies(dt[‘type‘])

    X_ori =  pd.concat((dt,type),1)
    del X_ori[‘type‘]
    y = np.array(y[0])
    
    clf = lr(penalty=‘l1‘)
    clf.fit(type,y)
    coef = clf.coef_[0]

    dt[‘coef‘] = 0
    for i in range(len(type.columns)):
        dt.loc[dt[‘type‘]==type.columns[i],‘coef‘] = coef[i]
    X_sub = dt
    del X_sub[‘type‘]
    
    clf = rf(n_estimators=1000)
    return [cross_val_score(clf,X_ori,y,cv=10).mean(),cross_val_score(clf,X_sub,y,cv=10).mean()]

def performance_ev(factor_size=100,special_percent=0.2,total_size=10000):
    x = None
    outcome = list()
    if type(factor_size) == list:
        x = factor_size
        factor_sizes = factor_size
        for factor_size in factor_sizes:
            outcome.append(show_importance_of_sub_model(total_size,factor_size,special_percent))
    else:
        x = special_percent
        special_percents = special_percent
        for special_percent in special_percents:
            outcome.append(show_importance_of_sub_model(total_size,factor_size,special_percent))
    
    performance = pd.DataFrame(outcome,columns=[‘dummy‘,‘sub_model‘])
    plt.plot(x,performance)

%matplotlib inline
performance_ev(factor_size = [5,10,20,40,80])

%matplotlib inline
performance_ev(special_percent = [0.02,0.04,0.08,0.16,0.32])

对离散特征，使用子模型建模能提升效果

离散特征越重要(重要的值越多，对应special_percent越高)，使用子模型带来的提升越大

离散特征的取值越多，RandomForest的预测效果越差