概述
在做的过程中,浏览了好多出色的报告,受益匪浅,浏览的文章主要包括:
import pandas as pd
import numpy as np
import seaborn as sns
from scipy import stats
from scipy.stats import skew
from scipy.stats import norm
import matplotlib.pyplot as plt
from sklearn.preprocessing import StandardScaler
from sklearn.manifold import TSNE
from sklearn.cluster import KMeans
from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler
# import warnings
# warnings.filterwarnings('ignore')
%config InlineBackend.figure_format = 'retina' #set 'png' here when working on notebook
%matplotlib inline
train_df = pd.read_csv("../input/train.csv")
test_df = pd.read_csv("../input/test.csv")
查看数据
我们拿到数据后,先对数据要有个大致的了解,我们有1460的训练数据和1460的测试数据,数据的特征列有81个,其中35个是数值类型的,44个类别类型。
我们通过阅读数据的描述说明,会发现列MSSubClass,OverallQual,OverallCond 这些数据可以将其转换为类别类型.
但是去具体看OverallQual,OverallCond 的时候,其没有缺失列,可以当做int来处理
all_df = pd.concat((train_df.loc[:,'MSSubClass':'SaleCondition'], test_df.loc[:,'MSSubClass':'SaleCondition']), axis=0,ignore_index=True)
all_df['MSSubClass'] = all_df['MSSubClass'].astype(str)
quantitative = [f for f in all_df.columns if all_df.dtypes[f] != 'object']
qualitative = [f for f in all_df.columns if all_df.dtypes[f] == 'object']
print("quantitative: {}, qualitative: {}" .format (len(quantitative),len(qualitative)))
quantitative: 35, qualitative: 44
处理缺失数据
对于缺失值的处理
缺失的行特别对,弃用该列
缺失的值比较少,取均值
缺失的值中间,对于类别信息的列可以将缺失作为新的类别做 one-hot
missing = all_df.isnull().sum()
missing.sort_values(inplace=True,ascending=False)
missing = missing[missing > 0]
types = all_df[missing.index].dtypes
percent = (all_df[missing.index].isnull().sum()/all_df[missing.index].isnull().count()).sort_values(ascending=False)
missing_data = pd.concat([missing, percent,types], axis=1, keys=['Total', 'Percent','Types'])
missing_data.sort_values('Total',ascending=False,inplace=True)
missing_data
image.png
missing.plot.bar()
output_14_1.png
上述缺失的列中有6列大于了15%的缺失率,其余主要是 BsmtX 和 GarageX 两大类,我们在具体决定这些列的处理之前,我们来看下我们要预测的价格的一些特征
数据统计分析
单变量分析
先看下我们要预测的价格的一些统计信息
train_df.describe()['SalePrice']
count 1460.000000
mean 180921.195890
std 79442.502883
min 34900.000000
25% 129975.000000
50% 163000.000000
75% 214000.000000
max 755000.000000
Name: SalePrice, dtype: float64
#skewness and kurtosis
print("Skewness: %f" % train_df['SalePrice'].skew())
print("Kurtosis: %f" % train_df['SalePrice'].kurt())
# 在统计学中,峰度(Kurtosis)衡量实数随机变量概率分布的峰态。峰度高就意味着方差增大是由低频度的大于或小于平均值的极端差值引起的。
Skewness: 1.882876
Kurtosis: 6.536282
相关性
我们先通过计算变量相关性,大致看下最相关的列都有什么
corrmat = train_df.corr()
#saleprice correlation matrix
k = 10 #number of variables for heatmap
cols = corrmat.nlargest(k, 'SalePrice')['SalePrice'].index
cm = np.corrcoef(train_df[cols].values.T)
sns.set(font_scale=1.25)
hm = sns.heatmap(cm, cbar=True, annot=True, square=True, fmt='.2f', annot_kws={'size': 10}, yticklabels=cols.values, xticklabels=cols.values)
plt.show()
output_21_0.png
## 同时是相关性列,也是缺失数据的
missing_data.index.intersection(cols)
Index(['GarageCars', 'GarageArea', 'TotalBsmtSF'], dtype='object')
missing_data.loc[missing_data.index.intersection(cols)]
image.png
从上面最相关的图中,我们可以首先将缺失的数据都给删除的
#dealing with missing data
all_df = all_df.drop((missing_data[missing_data['Total'] > 1]).index,1)
# df_train = df_train.drop(df_train.loc[df_train['Electrical'].isnull()].index)
all_df.isnull().sum().max() #just checking that there's no missing data missing...
# 对于missing 1 的我们到时候已平均数填充
#histogram and normal probability plot
sns.distplot(train_df['SalePrice'], fit=norm);
fig = plt.figure()
res = stats.probplot(train_df['SalePrice'], plot=plt)
output_27_0.png
output_27_1.png
一个好的处理方法就是进行log
train_df['SalePrice'] = np.log(train_df['SalePrice'])
#histogram and normal probability plot
sns.distplot(train_df['SalePrice'], fit=norm);
fig = plt.figure()
res = stats.probplot(train_df['SalePrice'], plot=plt)
output_30_0.png
output_30_1.png
看下每个定量变量的分布图
quantitative = [f for f in all_df.columns if all_df.dtypes[f] != 'object']
qualitative = [f for f in all_df.columns if all_df.dtypes[f] == 'object']
print("quantitative: {}, qualitative: {}" .format (len(quantitative),len(qualitative)))
quantitative: 30, qualitative: 26
f = pd.melt(all_df, value_vars=quantitative)
g = sns.FacetGrid(f, col="variable", col_wrap=2, sharex=False, sharey=False)
g = g.map(sns.distplot, "value")
output_33_0.png
上面有些数据是类似于正态分布的,我们可以对其进行log操作了提升质量的,有些则不适合,合适的预选对象有LotArea,BsmtUnfSF,1stFlrSF,TotalBsmtSF,KitchenAbvGr
我们计算下我们定量数据的偏度
all_df[quantitative].apply(lambda x: skew(x.dropna())).sort_values(ascending=False)
MiscVal 21.947195
PoolArea 16.898328
LotArea 12.822431
LowQualFinSF 12.088761
3SsnPorch 11.376065
KitchenAbvGr 4.302254
BsmtFinSF2 4.145323
EnclosedPorch 4.003891
ScreenPorch 3.946694
OpenPorchSF 2.535114
WoodDeckSF 1.842433
1stFlrSF 1.469604
BsmtFinSF1 1.424989
GrLivArea 1.269358
TotalBsmtSF 1.162285
BsmtUnfSF 0.919351
2ndFlrSF 0.861675
TotRmsAbvGrd 0.758367
Fireplaces 0.733495
HalfBath 0.694566
OverallCond 0.570312
BedroomAbvGr 0.326324
GarageArea 0.241176
OverallQual 0.197110
MoSold 0.195884
FullBath 0.167606
YrSold 0.132399
GarageCars -0.218260
YearRemodAdd -0.451020
YearBuilt -0.599806
dtype: float64
定量特征分析
方差分析或变方分析(Analysis of variance,简称 ANOVA)为数据分析中常见的统计模型
train = all_df.loc[train_df.index]
train['SalePrice'] = train_df.SalePrice
def anova(frame):
anv = pd.DataFrame()
anv['feature'] = qualitative
pvals = []
for c in qualitative:
samples = []
for cls in frame[c].unique():
s = frame[frame[c] == cls]['SalePrice'].values
samples.append(s)
pval = stats.f_oneway(*samples)[1]
pvals.append(pval)
anv['pval'] = pvals
return anv.sort_values('pval')
a = anova(train)
a['disparity'] = np.log(1./a['pval'].values)
sns.barplot(data=a, x='feature', y='disparity')
x=plt.xticks(rotation=90)
/Users/zhuanxu/anaconda/envs/linear_regression_demo/lib/python3.6/site-packages/scipy/stats/stats.py:2958: RuntimeWarning: invalid value encountered in double_scalars
ssbn += _square_of_sums(a - offset) / float(len(a))
output_38_1.png
此处 stats.f_oneway 的作用是计算这种定性变量对于SalePrice的作用,如果GarageType的每个类别SalePrice的价格方差差不多,意味着该变量对于SalePrice就没什么作用,stats.f_oneway 返回的 pval > 0.05,基本就意味着量集合的相似,具体可以看
下面对这些定性变量进行下处理,对齐进行数值编码,让他转换为定性的列
def encode(frame, feature):
ordering = pd.DataFrame()
ordering['val'] = frame[feature].unique()
ordering.index = ordering.val
ordering['spmean'] = frame[[feature, 'SalePrice']].groupby(feature).mean()['SalePrice']
ordering = ordering.sort_values('spmean')
ordering['ordering'] = range(1, ordering.shape[0]+1)
ordering = ordering['ordering'].to_dict()
for cat, o in ordering.items():
frame.loc[frame[feature] == cat, feature+'_E'] = o
qual_encoded = []
for q in qualitative:
encode(train, q)
qual_encoded.append(q+'_E')
print(qual_encoded)
['MSSubClass_E', 'Street_E', 'LotShape_E', 'LandContour_E', 'LotConfig_E', 'LandSlope_E', 'Neighborhood_E', 'Condition1_E', 'Condition2_E', 'BldgType_E', 'HouseStyle_E', 'RoofStyle_E', 'RoofMatl_E', 'Exterior1st_E', 'Exterior2nd_E', 'ExterQual_E', 'ExterCond_E', 'Foundation_E', 'Heating_E', 'HeatingQC_E', 'CentralAir_E', 'Electrical_E', 'KitchenQual_E', 'PavedDrive_E', 'SaleType_E', 'SaleCondition_E']
# 选出了包含缺失数据的行,处理一下
missing_data = all_df.isnull().sum()
missing_data = missing_data[missing_data>0]
ids = all_df[missing_data.index].isnull()
# index (0), columns (1)
all_df.loc[ids[ids.any(axis=1)].index][missing_data.index]
image.png
# 处理完后对于nan的数据,其值还是nan
train.loc[1379,'Electrical_E']
nan
相关性计算
def spearman(frame, features):
spr = pd.DataFrame()
spr['feature'] = features
#Signature: a.corr(other, method='pearson', min_periods=None)
#Docstring:
#Compute correlation with `other` Series, excluding missing values
# 计算特征和 SalePrice的 斯皮尔曼 相关系数
spr['spearman'] = [frame[f].corr(frame['SalePrice'], 'spearman') for f in features]
spr = spr.sort_values('spearman')
plt.figure(figsize=(6, 0.25*len(features))) # width, height
sns.barplot(data=spr, y='feature', x='spearman', orient='h')
features = quantitative + qual_encoded
spearman(train, features)
output_45_0.png
从上图我们可以看到特征 OverallQual Neighborhood GrLiveArea 对价格影响都比较大
下面我们分析下特征列之间的相关性,如果两特征相关,在做回归的时候会导致共线性问题
plt.figure(1)
corr = train[quantitative+['SalePrice']].corr()
sns.heatmap(corr)
plt.figure(2)
corr = train[qual_encoded+['SalePrice']].corr()
sns.heatmap(corr)
plt.figure(3)
# [31,27]
corr = pd.DataFrame(np.zeros([len(quantitative)+1, len(qual_encoded)+1]), index=quantitative+['SalePrice'], columns=qual_encoded+['SalePrice'])
for q1 in quantitative+['SalePrice']:
for q2 in qual_encoded+['SalePrice']:
corr.loc[q1, q2] = train[q1].corr(train[q2])
sns.heatmap(corr)
output_47_1.png
output_47_2.png
output_47_3.png
Pairplots
def pairplot(x, y, **kwargs):
ax = plt.gca()
ts = pd.DataFrame({'time': x, 'val': y})
ts = ts.groupby('time').mean()
ts.plot(ax=ax)
plt.xticks(rotation=90)
f = pd.melt(train, id_vars=['SalePrice'], value_vars=quantitative+qual_encoded)
g = sns.FacetGrid(f, col="variable", col_wrap=2, sharex=False, sharey=False, size=5)
g = g.map(pairplot, "value", "SalePrice")
IOPub data rate exceeded.
The notebook server will temporarily stop sending output
to the client in order to avoid crashing it.
To change this limit, set the config variable
`--NotebookApp.iopub_data_rate_limit`.
从上面的数据我们能清晰的看到哪些变量是线性关系比较好的,哪些是非线性关系,还有一些能看到如果加二次项可能会表现出比较的线性相关性出来
价格分段
我们对于价格简单的做一个二分,然后看下特征的不同,我们先看下SalePrice的图
a = train['SalePrice']
a.plot.hist()
output_51_1.png
features = quantitative
standard = train[train['SalePrice'] < np.log(200000)]
pricey = train[train['SalePrice'] >= np.log(200000)]
diff = pd.DataFrame()
diff['feature'] = features
diff['difference'] = [(pricey[f].fillna(0.).mean() - standard[f].fillna(0.).mean())/(standard[f].fillna(0.).mean())
for f in features]
sns.barplot(data=diff, x='feature', y='difference')
x=plt.xticks(rotation=90)
![Uploading output_52_0_342062.png . . .]
上图可以看到贵的房子,泳池会影响比较大
分类
我们先对数据做一个简单的分类
features = quantitative + qual_encoded
model = TSNE(n_components=2, random_state=0, perplexity=50)
X = train[features].fillna(0.).values
tsne = model.fit_transform(X)
std = StandardScaler()
s = std.fit_transform(X)
pca = PCA(n_components=30)
pca.fit(s)
pc = pca.transform(s)
kmeans = KMeans(n_clusters=5)
kmeans.fit(pc)
fr = pd.DataFrame({'tsne1': tsne[:,0], 'tsne2': tsne[:, 1], 'cluster': kmeans.labels_})
sns.lmplot(data=fr, x='tsne1', y='tsne2', hue='cluster', fit_reg=False)
print(np.sum(pca.explained_variance_ratio_))
0.838557886152
output_55_1.png
30个成分能覆盖83%的方差,整体看来,这种聚类方法不太好
总结
本文对数据进行了一些分析,下一篇会基于这个分析做模型处理