用Python研究了三千套房子,告訴你究竟是什么抬高了房價(jià)?
關(guān)于房價(jià),一直都是全民熱議的話題,畢竟不少人終其一生都在為之奮斗。
房地產(chǎn)的泡沫究竟有多大不得而知?今天我們拋開泡沫,回歸房屋最本質(zhì)的內(nèi)容,來分析一下房價(jià)的影響因素究竟是什么?
1、導(dǎo)入數(shù)據(jù)
- import numpy as np
- import pandas as pd
- import matplotlib.pyplot as plt
- import seaborn as sn
- import missingno as msno
- %matplotlib inline
- train = pd.read_csv('train.csv',index_col=0)
- #導(dǎo)入訓(xùn)練集
- test = pd.read_csv('test.csv',index_col=0)
- #導(dǎo)入測試集
- train.head(3)
- print('train訓(xùn)練集缺失數(shù)據(jù)分布圖')
- msno.matrix(train)
- print('test測試集缺失數(shù)據(jù)分布圖')
- msno.matrix(test)
從上面的數(shù)據(jù)缺失可視化圖中可以看出,部分特征的數(shù)據(jù)缺失十分嚴(yán)重,下面我們來對特征的缺失數(shù)量進(jìn)行統(tǒng)計(jì)。
2、目標(biāo)Y值分析
- ##分割Y和X數(shù)據(jù)
- y=train['SalePrice']
- #看一下y的值分布
- prices = pd.DataFrame({'price':y,'log(price+1)':np.log1p(y)})
- prices.hist()
觀察目標(biāo)變量y的分布和取對數(shù)后的分布看,取完對數(shù)后更傾向于符合正太分布,故我們對y進(jìn)行對數(shù)轉(zhuǎn)化。
- y = np.log1p(y) #+1的目的是防止對數(shù)轉(zhuǎn)化后的值無意義
3、合并數(shù)據(jù) 缺失處理
- #合并訓(xùn)練特征和測試集
- all_df = pd.concat((X,test),axis=0)
- print('all_df缺失數(shù)據(jù)圖')
- msno.matrix(all_df)
- #定義缺失統(tǒng)計(jì)函數(shù)
- def show_missing(feature):
- missing = feature.columns[feature.isnull().any()].tolist()
- return missing
- print('缺失特征的數(shù)據(jù)缺失量統(tǒng)計(jì):')
- all_df[show_missing(all_df)].isnull().sum()
- #先處理numeric數(shù)值型數(shù)據(jù)
- #挨個(gè)兒看一下分布
- fig,axs = plt.subplots(3,2,figsize=(16,9))
- all_df['BsmtFinSF1'].hist(ax = axs[0,0])#眾數(shù)填充
- all_df['BsmtFinSF2'].hist(ax = axs[0,1])#眾數(shù)
- all_df['BsmtUnfSF'].hist(ax = axs[1,0])#中位數(shù)
- all_df['TotalBsmtSF'].hist(ax = axs[1,1])#均值填充
- all_df['BsmtFullBath'].hist(ax = axs[2,0])#眾數(shù)
- all_df['BsmtHalfBath'].hist(ax = axs[2,1])#眾數(shù)
- #lotfrontage用均值填充
- mean_lotfrontage = all_df.LotFrontage.mean()
- all_df.LotFrontage.hist()
- print('用均值填充:')
- cat_input(all_df,'LotFrontage',mean_lotfrontage)
- cat_input(all_df,'BsmtFinSF1',0.0)
- cat_input(all_df,'BsmtFinSF2',0.0)
- cat_input(all_df,'BsmtFullBath',0.0)
- cat_input(all_df,'BsmtHalfBath',0.0)
- cat_input(all_df,'BsmtUnfSF',467.00)
- cat_input(all_df,'TotalBsmtSF',1051.78)
- #在處理字符型,同樣,挨個(gè)看下分布
- fig,axs = plt.subplots(4,2,figsize=(16,9))
- all_df['MSZoning'].hist(ax = axs[0,0])#眾數(shù)填充
- all_df['Utilities'].hist(ax = axs[0,1])#眾數(shù)
- all_df['Exterior1st'].hist(ax = axs[1,0])#眾數(shù)
- all_df['Exterior2nd'].hist(ax = axs[1,1])#眾數(shù)填充
- all_df['KitchenQual'].hist(ax = axs[2,0])#眾數(shù)
- all_df['Functional'].hist(ax = axs[2,1])#眾數(shù)
- all_df['SaleType'].hist(ax = axs[3,0])#眾數(shù)
- cat_input(all_df,'MSZoning','RL')
- cat_input(all_df,'Utilities','AllPub')
- cat_input(all_df,'Exterior1st','VinylSd')
- cat_input(all_df,'Exterior2nd','VinylSd')
- cat_input(all_df,'KitchenQual','TA')
- cat_input(all_df,'Functional','Typ')
- cat_input(all_df,'SaleType','WD')
- #再看一下缺失分布
- msno.matrix(all_df)
binggo,數(shù)據(jù)干凈啦!下面開始處理特征,經(jīng)過上述略微復(fù)雜的處理,數(shù)據(jù)集中所有的缺失數(shù)據(jù)都已處理完畢,可以開始接下來的工作啦!
缺失處理總結(jié):在本篇文章所使用的數(shù)據(jù)集中存在比較多的缺失,缺失數(shù)據(jù)包括數(shù)值型和字符型,處理原則主要有兩個(gè):
一、根據(jù)繪制數(shù)據(jù)分布直方圖,觀察數(shù)據(jù)分布的狀態(tài),采取合適的方式填充缺失數(shù)據(jù);
二、非常重要的特征描述,認(rèn)真閱讀,按照特征描述填充可以解決大部分問題。
4、特征處理
讓我們在重新仔細(xì)審視一下數(shù)據(jù)有沒有問題?仔細(xì)觀察發(fā)現(xiàn)MSSubClass特征實(shí)際上是分類特征,但是數(shù)據(jù)顯示是int類型,這個(gè)需要改成str。
- #觀察特征屬性發(fā)現(xiàn),MSSubClass是分類特征,但是數(shù)據(jù)給的是數(shù)值型,需要對其做轉(zhuǎn)換
- all_df['MSSubClass']=all_df['MSSubClass'].astype(str)
- #將分類變量轉(zhuǎn)變成數(shù)值變量
- all_df = pd.get_dummies(all_df)
- print('分類變量轉(zhuǎn)換完成后有{}行{}列'.format(*all_df.shape))
分類變量轉(zhuǎn)換完成后有2919行316列
- #標(biāo)準(zhǔn)化處理
- numeric_cols = all_df.columns[all_df.dtypes !='uint8']
- #x-mean(x)/std(x)
- numeric_mean = all_df.loc[:,numeric_cols].mean()
- numeric_std = all_df.loc[:,numeric_cols].std()
- all_df.loc[:,numeric_cols] = (all_df.loc[:,numeric_cols]-numeric_mean)/numeric_std
再把數(shù)據(jù)拆分到訓(xùn)練集和測試集
- train_df = all_df.ix[0:1460]#訓(xùn)練集
- test_df = all_df.ix[1461:]#測試集
5、構(gòu)建基準(zhǔn)模型
- from sklearn import cross_validation
- from sklearn import linear_model
- from sklearn.learning_curve import learning_curve
- from sklearn.metrics import explained_variance_score
- from sklearn.grid_search import GridSearchCV
- from sklearn.model_selection import cross_val_score
- from sklearn.ensemble import RandomForestRegressor
- y = y.values #轉(zhuǎn)換成array數(shù)組
- X = train_df.values #轉(zhuǎn)換成array數(shù)組
- cv = cross_validation.ShuffleSplit(len(X),n_iter=3,test_size=0.2)
- print('嶺回歸交叉驗(yàn)證結(jié)果:')
- for train_index,test_index in cv:
- ridge = linear_model.Ridge(alpha=1).fit(X,y)
- print('train_score:{0:.3f},test_score:{1:.3f}\n'.format(ridge.score(X[train_index],y[train_index]), ridge.score(X[test_index],y[test_index])))
- print('隨機(jī)森林交叉驗(yàn)證結(jié)果:')
- for train_index,test_index in cv:
- rf = RandomForestRegressor().fit(X,y)
- print('train_score:{0:.3f},test_score:{1:.3f}\n'.format(rf.score(X[train_index],y[train_index]), rf.score(X[test_index],y[test_index])))
哇!好意外啊,這兩個(gè)模型的結(jié)果表現(xiàn)都不錯(cuò),但是隨機(jī)森林的結(jié)果似乎更好,下面來看看學(xué)習(xí)曲線情況。
我們采用的是默認(rèn)的參數(shù),沒有調(diào)優(yōu)處理,得到的兩個(gè)基準(zhǔn)模型都存在過擬合現(xiàn)象。下面,我們開始著手參數(shù)的調(diào)整,希望能夠改善模型的過擬合現(xiàn)象。
6、參數(shù)調(diào)優(yōu)
嶺回歸正則項(xiàng)縮放系數(shù)alpha調(diào)整
- alphas =[0.01,0.1,1,10,20,50,100,300]
- test_scores = []
- for alp in alphas:
- clf = linear_model.Ridge(alp)
- test_score = -cross_val_score(clf,X,y,cv=10,scoring='neg_mean_squared_error')
- test_scores.append(np.mean(test_score))
- import matplotlib.pyplot as plt
- %matplotlib inline
- plt.plot(alphas,test_scores)
- plt.title('alpha vs test_score')
alpha在10-20附近均方誤差最小
隨機(jī)森林參數(shù)調(diào)優(yōu)
隨機(jī)森林算法,本篇中主要調(diào)整三個(gè)參數(shù):maxfeatures,maxdepth,n_estimators
- #隨機(jī)森林的深度參數(shù)
- max_depth=[2,4,6,8,10]
- test_scores_depth = []
- for depth in max_depth:
- clf = RandomForestRegressor(max_depth=depth)
- test_score_depth = -cross_val_score(clf,X,y,cv=10,scoring='neg_mean_squared_error')
- test_scores_depth.append(np.mean(test_score_depth))
- #隨機(jī)森林的特征個(gè)數(shù)參數(shù)
- max_features =[.1, .3, .5, .7, .9, .99]
- test_scores_feature = []
- for feature in max_features:
- clf = RandomForestRegressor(max_features=feature)
- test_score_feature = -cross_val_score(clf,X,y,cv=10,scoring='neg_mean_squared_error')
- test_scores_feature.append(np.mean(test_score_feature))
- #隨機(jī)森林的估計(jì)器個(gè)位數(shù)參數(shù)
- n_estimators =[10,50,100,200,500]
- test_scores_n = []
- for n in n_estimators:
- clf = RandomForestRegressor(n_estimators=n)
- test_score_n = -cross_val_score(clf,X,y,cv=10,scoring='neg_mean_squared_error')
- test_scores_n.append(np.mean(test_score_n))
隨機(jī)森林的各項(xiàng)參數(shù)來看,深度位于8,選擇特征個(gè)數(shù)比例為0.5,估計(jì)器個(gè)數(shù)為500時(shí),效果***。下面分別利用上述得到的***參數(shù)分別重新訓(xùn)練,看一下學(xué)習(xí)曲線,過擬合現(xiàn)象是否得到緩解?
再回想一下,我們最初的基線模型學(xué)習(xí)曲線的形狀,是不是得到了一定程度的緩解?OK,下面我們采用模型融合技術(shù),對數(shù)據(jù)進(jìn)行預(yù)測。
- #預(yù)測
- ridge = linear_model.Ridge(alpha=10).fit(X,y)
- rf = RandomForestRegressor(n_estimators=500,max_depth=8,max_features=.5).fit(X,y)
- y_ridge = np.expm1(ridge.predict(test_df.values))
- y_rf = np.expm1(rf.predict(test_df.values))
- y_final = (y_ridge + y_rf)/2
本篇房價(jià)預(yù)測的模型搭建已經(jīng)完成。同樣,再梳理一邊思路:
一、本篇用到的房價(jià)數(shù)據(jù)集存在比較多的數(shù)據(jù)缺失,且分類變量十分多。在預(yù)處理階段需要將訓(xùn)練集和測試集合并,進(jìn)行缺失填充和one-hot獨(dú)熱變量處理,保證數(shù)據(jù)處理過程的一致性,在數(shù)據(jù)缺失填充過程中,需要綜合考慮特征的實(shí)際描述和數(shù)據(jù)的分布,選擇合適的填充方式填充;
二、為防止數(shù)據(jù)變量不統(tǒng)一帶來的模型準(zhǔn)確率下降,將數(shù)值型特征進(jìn)行標(biāo)準(zhǔn)化處理,數(shù)據(jù)處理完成后,按照數(shù)據(jù)合并的方式,再還原到訓(xùn)練集和測試集;
三、先構(gòu)建嶺回歸和隨機(jī)森林基準(zhǔn)模型,進(jìn)行三折交叉驗(yàn)證,繪制學(xué)習(xí)曲線,存在明顯的過擬合現(xiàn)象;
四、接下來分別對兩個(gè)基準(zhǔn)模型進(jìn)行參數(shù)調(diào)優(yōu),獲得使得均方誤差最小的參數(shù),返回到訓(xùn)練集進(jìn)行訓(xùn)練;
五、采用并行模型融合的方式,計(jì)算兩個(gè)模型預(yù)測結(jié)果的均值作為測試集的預(yù)測結(jié)果。