python 进行各种回归

基本回归：线性、决策树、SVM、KNN
集成方法：随机森林、Adaboost、GradientBoosting、Bagging、ExtraTrees
##学会了数据分层抽样，以及各种回归的代码书写。可能还需要注意调参等。
继续学习网址：使用sklearn做各种回归

数据准备

from matplotlib import pyplot as plt
%matplotlib inline
plt.style.use('fivethirtyeight')  #设置matplotlib作图风格
import seaborn as sns
import pandas as pd
sns.set()   #设置sns作图默认风格
import warnings
warnings.filterwarnings('ignore')
import numpy as np

data = pd.read_csv("../论文/mianbanshu.csv")

data.head()

	city	year	rkmd	rjys	ys	rq	gd	dl	ps	ws	jz	gymj	lh	ldl	shlj	whh	qntz	wstz	Unnamed: 18
0	AH	2008	2043.0	174.61	95.11	87.60	7.50	14.15	1.75	78.86	58.64	9.29	35.97	32.03	94.75	53.95	2722949.0	25465	NaN
1	AH	2009	2114.0	160.96	95.25	88.62	9.66	14.91	8.23	83.24	62.40	10.23	37.16	33.63	95.51	60.91	3762160.0	27900	NaN
2	AH	2010	2469.0	160.83	96.06	90.52	9.88	16.01	8.81	88.46	71.58	10.95	37.50	33.67	95.59	64.56	4756917.0	30324	NaN
3	AH	2011	2265.0	168.99	96.55	93.35	10.66	18.00	10.25	91.09	79.04	11.88	39.47	34.55	97.06	86.99	6203403.0	32884	NaN
4	AH	2012	2401.0	165.45	98.02	94.61	11.13	18.47	11.72	94.53	86.39	11.92	38.80	34.72	95.28	91.14	6339795.0	39962	NaN

数据探索

f, ax = plt.subplots(figsize = (11, 7))
sns.factorplot(data = data, x = 'year', y = "wstz", 
               palette = 'plasma',
               hue = 'city', ax=ax) 

f, ax = plt.subplots(figsize = (11, 7))
sns.factorplot(data = data, x = 'year', y = "rkmd", 
               palette = 'plasma',
               hue = 'city', ax=ax) 

f, ax = plt.subplots(figsize = (11, 7))
sns.factorplot(data = data, x = 'year', y = "rjys", 
               palette = 'plasma',
               hue = 'city', ax=ax) 

data = data.drop(["Unnamed: 18"],axis=1)

#Correlation Matrix（另一种画法）

corr = kiva_loans_data.corr() plt.figure(figsize=(12,12))
sns.heatmap(corr,
xticklabels=corr.columns.values,
yticklabels=corr.columns.values, annot=True, cmap=‘cubehelix’, square=True) plt.title(‘Correlation between
different features’) corr

corr_all = data.drop(["city"],axis = 1).corr()

mask = np.zeros_like(corr_all, dtype = np.bool)
mask[np.triu_indices_from(mask)] = True
f, ax = plt.subplots(figsize = (11, 9))
sns.heatmap(corr_all, mask = mask,
            square = True, linewidths = .5, ax = ax, cmap = "BuPu")      
#plt.savefig('heatmap.png')

a = set(data["city"])
b = [9] * 31

实现分层抽样，抽到的即为训练样本

gbr=data.groupby('city')

# 分层抽样字典定义 组名：数据个数
typicalNDict = dict(zip(a,b))
 
# 函数定义
def typicalsamling(group, typicalNDict):
    name = group.name
    n = typicalNDict[name]
    return (group.sample(n=n))

#返回值：抽样后的数据框
train = data.groupby('city').apply(typicalsamling, typicalNDict)
train.to_csv('TrainData.csv', index=False)

train.head()

		city	year	rkmd	rjys	ys	rq	gd	dl	ps	ws	jz	gymj	lh	ldl	shlj	whh	qntz	wstz
city
AH	4	AH	2012	2401.0	165.45	98.02	94.61	11.13	18.47	11.72	94.53	86.39	11.92	38.80	34.72	95.28	91.14	6339795.0	39962
	2	AH	2010	2469.0	160.83	96.06	90.52	9.88	16.01	8.81	88.46	71.58	10.95	37.50	33.67	95.59	64.56	4756917.0	30324
	8	AH	2016	2487.0	180.19	99.20	98.05	12.76	21.82	13.18	97.36	92.03	14.02	41.71	37.67	99.94	99.94	7417525.0	67256
	0	AH	2008	2043.0	174.61	95.11	87.60	7.50	14.15	1.75	78.86	58.64	9.29	35.97	32.03	94.75	53.95	2722949.0	25465
	3	AH	2011	2265.0	168.99	96.55	93.35	10.66	18.00	10.25	91.09	79.04	11.88	39.47	34.55	97.06	86.99	6203403.0	32884

train = pd.read_csv("../论文/TrainData.csv")

train.head()

	city	year	rkmd	rjys	ys	rq	gd	dl	ps	ws	jz	gymj	lh	ldl	shlj	whh	qntz	wstz
0	AH	2012	2401.0	165.45	98.02	94.61	11.13	18.47	11.72	94.53	86.39	11.92	38.80	34.72	95.28	91.14	6339795.0	39962
1	AH	2010	2469.0	160.83	96.06	90.52	9.88	16.01	8.81	88.46	71.58	10.95	37.50	33.67	95.59	64.56	4756917.0	30324
2	AH	2016	2487.0	180.19	99.20	98.05	12.76	21.82	13.18	97.36	92.03	14.02	41.71	37.67	99.94	99.94	7417525.0	67256
3	AH	2008	2043.0	174.61	95.11	87.60	7.50	14.15	1.75	78.86	58.64	9.29	35.97	32.03	94.75	53.95	2722949.0	25465
4	AH	2011	2265.0	168.99	96.55	93.35	10.66	18.00	10.25	91.09	79.04	11.88	39.47	34.55	97.06	86.99	6203403.0	32884

train.shape

(279, 18)

测试样本

test = data[~data["wstz"].isin(list(train["wstz"]))]

test.shape

(31, 18)

test.head()

	city	year	rkmd	rjys	ys	rq	gd	dl	ps	ws	jz	gymj	lh	ldl	shlj	whh	qntz	wstz
7	AH	2015	2458.0	168.90	98.79	97.55	12.38	20.82	12.67	96.68	91.80	13.37	41.16	37.16	99.55	99.55	6324122.0	106486
10	BJ	2008	1181.0	187.22	100.00	100.00	6.77	6.21	0.73	78.92	74.52	8.56	37.15	35.85	97.71	97.71	5189156.0	98295
29	FJ	2017	2854.0	203.69	99.56	97.45	13.85	17.41	9.42	92.21	89.07	14.13	43.69	40.28	99.38	99.38	7128613.0	260721
32	GS	2010	3793.0	155.12	91.57	74.29	6.89	12.20	4.89	62.59	58.58	8.12	27.12	23.14	97.84	37.95	944242.0	6289
47	GD	2015	3060.0	248.95	98.46	97.60	17.74	13.60	9.51	93.65	93.25	17.40	41.43	37.32	97.63	91.56	8496440.0	644310

从上面可以看出每个city之间的指标随着年份可能平稳，或趋于增加，但是各折线相交都较少，说明这些指标已足以将样本进行区分。

另外从热力图可以看出，year这个变量与其他的变量有明显相关关系的情况比较少，大部分指标在不同年份相差不大，故考虑也将year这个变量删除。

去除城市和年份这一列

train = train.drop(["city","year"],axis=1)
X_train = train[train.columns[:-1]]
y_train = train[train.columns[-1]]

test = test.drop(["city","year"],axis=1)
X_test = test[test.columns[:-1]]
y_test = test[test.columns[-1]]

1.线性回归

from sklearn import linear_model
model1_linear = linear_model.LinearRegression()
model1_linear = model1_linear.fit(X_train,y_train)
y_pred1 = model1_linear.predict(X_test)

from sklearn.metrics import mean_absolute_error 
mean_absolute_error(y_pred1,y_test)

84180.196193290452

from sklearn.metrics import r2_score
r2_score(y_pred1,y_test)

0.48478789713121917

plt.figure(figsize=(12, 6))
plt.plot(list(y_pred1),label="forcast")
plt.plot(list(y_test),label="test")
plt.ylabel('wstz',fontsize=14,horizontalalignment='center')
plt.legend()
plt.show()

2.决策树回归

from sklearn import tree
model2_tree = tree.DecisionTreeRegressor()
model2_tree = model2_tree.fit(X_train,y_train)
y_pred2 = model2_tree.predict(X_test)

mean_absolute_error(y_pred2,y_test)

52285.806451612902

r2_score(y_pred2,y_test)

0.55980167732053721

plt.figure(figsize=(12, 6))
plt.plot(list(y_pred2),label="forcast")
plt.plot(list(y_test),label="test")
plt.ylabel('wstz',fontsize=14,horizontalalignment='center')
plt.legend()
plt.show()

3.SVM回归

####3.3SVM回归####
from sklearn import svm
model3_SVR = svm.SVR()
model3_SVR = model3_SVR.fit(X_train,y_train)
y_pred3 = model3_SVR.predict(X_test)

mean_absolute_error(y_pred3,y_test)

90906.8702612044

r2_score(y_pred3,y_test)

-16123841.245722869

plt.figure(figsize=(12, 6))
plt.plot(list(y_pred3),label="forcast")
plt.plot(list(y_test),label="test")
plt.ylabel('wstz',fontsize=14,horizontalalignment='center')
plt.legend()
plt.show()

4.KNN回归

from sklearn.neighbors import KNeighborsRegressor
model4_KN = KNeighborsRegressor(n_neighbors=3)
model4_KN.fit(X_train,y_train)
y_pred4 = model4_KN.predict(X_test)

mean_absolute_error(y_pred4,y_test)

117266.68817204301

r2_score(y_pred4,y_test)

-1.5288016385577752

plt.figure(figsize=(12, 6))
plt.plot(list(y_pred4),label="forcast")
plt.plot(list(y_test),label="test")
plt.ylabel('wstz',fontsize=14,horizontalalignment='center')
plt.legend()
plt.show()

5.随机森林回归

from sklearn.ensemble import RandomForestRegressor

model5_RF = RandomForestRegressor()
model5_RF = model5_RF.fit(X_train,y_train)
y_pred5 = model5_RF.predict(X_test)
mean_absolute_error(y_pred5,y_test)

34605.676129032247

r2_score(y_pred5,y_test)

0.74665386911940801

plt.figure(figsize=(12, 6))
plt.plot(list(y_pred5),label="forcast")
plt.plot(list(y_test),label="test")
plt.ylabel('wstz',fontsize=14,horizontalalignment='center')
plt.legend()
plt.show()

6.AdaBoost回归

from sklearn.ensemble import AdaBoostRegressor
model6_AdaBoost = AdaBoostRegressor(n_estimators=100)
model6_AdaBoost = model6_AdaBoost.fit(X_train,y_train)
y_pred6 = model6_AdaBoost.predict(X_test)
mean_absolute_error(y_pred6,y_test)

71458.328974579621

r2_score(y_pred6,y_test)

0.30236366136673865

plt.figure(figsize=(12, 6))
plt.plot(list(y_pred6),label="forcast")
plt.plot(list(y_test),label="test")
plt.ylabel('wstz',fontsize=14,horizontalalignment='center')
plt.legend()
plt.show()

7.梯度上升回归

from sklearn.ensemble import GradientBoostingRegressor
model7_GBDT = GradientBoostingRegressor()
model7_GBDT = model7_GBDT.fit(X_train,y_train)
y_pred7 = model7_GBDT.predict(X_test)
mean_absolute_error(y_pred7,y_test)

30247.23832854132

r2_score(y_pred7,y_test)

0.79286645995236993

plt.figure(figsize=(12, 6))
plt.plot(list(y_pred7),label="forcast")
plt.plot(list(y_test),label="test")
plt.ylabel('wstz',fontsize=14,horizontalalignment='center')
plt.legend()
plt.show()

8.Bagging回归

from sklearn.ensemble import BaggingRegressor
model8_Bagging = BaggingRegressor()
model8_Bagging = model8_Bagging.fit(X_train,y_train)
y_pred8 = model8_Bagging.predict(X_test)
mean_absolute_error(y_pred8,y_test)

36788.058064516132

r2_score(y_pred8,y_test)

0.68475165003444116

plt.figure(figsize=(12, 6))
plt.plot(list(y_pred8),label="forcast")
plt.plot(list(y_test),label="test")
plt.ylabel('wstz',fontsize=14,horizontalalignment='center')
plt.legend()
plt.show()

9.ExtraTree极端随机树回归

####3.9ExtraTree极端随机树回归####
from sklearn.tree import ExtraTreeRegressor
model9_Etra = ExtraTreeRegressor()
model9_Etra = model9_Etra.fit(X_train,y_train)
y_pred9 = model9_Etra.predict(X_test)
mean_absolute_error(y_pred9,y_test)

45674.580645161288

r2_score(y_pred9,y_test)

0.31221911911576294

plt.figure(figsize=(12, 6))
plt.plot(list(y_pred9),label="forcast")
plt.plot(list(y_test),label="test")
plt.ylabel('wstz',fontsize=14,horizontalalignment='center')
plt.legend()
plt.show()

模型比较

modelnames = ['linear_model',
              'DecisionTreeRegressor',
              'RandomForestRegressor', 
              'AdaBoostRegressor', 
              'GradientBoostingRegressor',
              'BaggingRegressor',
              'ExtraTreeRegressor',
              ]
R_square = [r2_score(y_pred1,y_test),r2_score(y_pred2,y_test),
            r2_score(y_pred5,y_test),r2_score(y_pred6,y_test),
            r2_score(y_pred7,y_test),r2_score(y_pred8,y_test),
            r2_score(y_pred9,y_test)]
plt.figure(figsize=(10,9))

fig = plt.figure(figsize=(10,6))
ax = fig.add_subplot(1, 1, 1)
 
ticks = ax.set_xticks(range(0,7))#设定x轴有7个标签
ax.plot(R_square,'ko--')
labels = ax.set_xticklabels(modelnames,fontsize='14',rotation=90)  
plt.title("model comparsion",fontsize='14')
plt.grid(True)
plt.show()

误差衡量指标平均绝对误差和决定系数具有一致性，当决定系数系数越大，平均绝对误差就越小，上述试验数据也是这样。
相对较好的两个模型如下：
随机森林回归:平均绝对误差：34605 决定系数：0.747
梯度上升回归:平均绝对误差：30247 决定系数：0.793

---

前馈型神经网络，也叫多层感知器(Multi-Layer Perceptron,MLP)模型或全连接神经网络。它每一层的神经元从上一层神经元接收数据，经过计算之后产生输出数据，送入下一层神经元继续处理，最后一层神经元的输出是神经网络最终的输出值。

import numpy as np
from sklearn.neural_network import MLPRegressor


MLP = MLPRegressor()
MLP.fit(X_train,y_train)
y_pred10 = MLP.predict(X_test)
score = r2_score(y_test, y_pred10)  

plt.figure(figsize=(12, 6))
plt.plot(np.arange(len(y_pred10)),y_test,'go-',label = 'true value')
plt.plot(np.arange(len(y_pred10)),y_pred10,'ro-',label = 'predict value')
plt.legend()
plt.show()