过关斩将打进Kaggle竞赛Top 0.3%,我是这样做的

作者 | Lavanya Shukla

译者 | Monanfei

责编 | 夕颜

出品 | AI科技大本营(id:rgznai100)

导读:刚开始接触数据竞赛时,我们可能会被一些高大上的技术吓到。各界大佬云集,各种技术令人眼花缭乱,新手们就像蜉蝣一般渺小无助。今天本文就分享一下在 kaggle 的竞赛中,参赛者取得 top0.3% 的经验和技巧。让我们开始吧!

Top 0.3% 模型概览

赛题和目标

  • 数据集中的每一行都描述了某一房屋的特征
  • 在已知这些特征的条件下,预测每间房的销售价格
  • 预测价格对数和真实价格对数的RMSE(均方根误差)作为模型的评估指标。将RMSE转化为对数尺度,能够保证廉价房屋和高价房屋的预测误差,对模型分数的影响较为一致。

模型训练过程中的重要细节

  • 交叉验证:使用12-折交叉验证
  • 模型:在每次交叉验证中,同时训练七个模型(ridge, svr, gradient boosting, random forest, xgboost, lightgbm regressors)
  • Stacking 方法:使用 xgboot 训练了元 StackingCVRegressor 学习器
  • 模型融合:所有训练的模型都会在不同程度上过拟合,因此,为了做出最终的预测,将这些模型进行了融合,得到了鲁棒性更强的预测结果

模型性能

从下图可以看出,融合后的模型性能最好,RMSE 仅为 0.075,该融合模型用于最终预测。

In[1]:

from IPython.display import Image
Image("../input/kernel-files/model_training_advanced_regression.png")

Output[1]:

现在让我们正式开始吧!

In[2]:

# Essentials
import numpy as np
import pandas as pd
import datetime
import random
# Plots
import seaborn as sns
import matplotlib.pyplot as plt


# Models
from sklearn.ensemble import RandomForestRegressor, GradientBoostingRegressor, AdaBoostRegressor, BaggingRegressor
from sklearn.kernel_ridge import KernelRidge
from sklearn.linear_model import Ridge, RidgeCV
from sklearn.linear_model import ElasticNet, ElasticNetCV
from sklearn.svm import SVR
from mlxtend.regressor import StackingCVRegressor
import lightgbm as lgb
from lightgbm import LGBMRegressor
from xgboost import XGBRegressor


# Stats
from scipy.stats import skew, norm
from scipy.special import boxcox1p
from scipy.stats import boxcox_normmax


# Misc
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import KFold, cross_val_score
from sklearn.metrics import mean_squared_error
from sklearn.preprocessing import OneHotEncoder
from sklearn.preprocessing import LabelEncoder
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import scale
from sklearn.preprocessing import StandardScaler
from sklearn.preprocessing import RobustScaler
from sklearn.decomposition import PCA

pd.set_option('display.max_columns', None)

# Ignore useless warnings
import warnings
warnings.filterwarnings(action="ignore")
pd.options.display.max_seq_items = 8000
pd.options.display.max_rows = 8000

import os
print(os.listdir("../input/kernel-fi

Output[2]:

['model_training_advanced_regression.png']

In[3]:

Output[3]:

((1460, 81), (1459, 80))

EDA

目标

  • 数据集中的每一行都描述了某一间房的特征
  • 在已知这些特征的条件下,预测每间房的销售价格

对原始数据进行可视化

In[4]:

Output[5]:

SalePrice:目标值的特性探究

In[5]:

In[6]:

Skewness: 1.882876 Kurtosis: 6.536282 可用的特征:深入探索

数据可视化

In[7]:

探索这些特征以及 SalePrice 的相关性

In[8]:

Output[8]:

<matplotlib.axes._subplots.AxesSubplot at 0x7ff0e416e4e0>

选取部分特征,可视化它们和 SalePrice 的相关性

Input[9]:

Input[10]:

Input[11]:

.3, ylim=(0,800000));

Input[12]:

lim=(0,800000));

Input[13]:

ylim=(0,800000));

Input[14]:

Output[14]:

((1460, 80), (1459, 79)) 可视化 salePrice 的分布

Input[15]:

从上图中可以看出,SalePrice 有点向右边倾斜,由于大多数机器学习模型对非正态分布的数据的效果不佳,因此,我们对数据进行变换,修正这种倾斜:log(1+x)

Input[16]:

对 SalePrice 重新进行可视化

Input[17]:

sns.set_style("white")
sns.set_color_codes(palette='deep')
f, ax = plt.subplots(figsize=(8, 7))
#Check the new distribution
sns.distplot(train['SalePrice'] , fit=norm, color="b");


# Get the fitted parameters used by the function
(mu, sigma) = norm.fit(train['SalePrice'])
print( '\n mu = {:.2f} and sigma = {:.2f}\n'.format(mu, sigma))


#Now plot the distribution
plt.legend(['Normal dist. ($\mu=$ {:.2f} and $\sigma=$ {:.2f} )'.format(mu, sigma)],
            loc='best')
ax.xaxis.grid(False)
ax.set(ylabel="Frequency")
ax.set(xlabel="SalePrice")
ax.set(title="SalePrice distribution")
sns.despine(trim=True, left=True)

plt.show

mu = 12.02 and sigma = 0.40

从图中可以看到,当前的 SalePrice 已经变成了正态分布

Input[18]:

Input[19]:

# Split features and labels
train_labels = train['SalePrice'].reset_index(drop=True)
train_features = train.drop(['SalePrice'], axis=1)
test_features = test
# Combine train and test features in order to apply the feature transformation pipeline to the entire dataset
all_features = pd.concat([train_features, test_features]).reset_index(drop=True)
all_features.shape

Input[19]:

(2917, 79)

填充缺失值

Input[20]:

# determine the threshold for missing values
def percent_missing(df):
    data = pd.DataFrame(df)
    df_cols = list(pd.DataFrame(data))
    dict_x = {}
    for i in range(0, len(df_cols)):
        dict_x.update({df_cols[i]: round(data[df_cols[i]].isnull().mean()*100,2)})
    
    return dict_x

missing = percent_missing(all_features)
df_miss = sorted(missing.items(), key=lambda x: x[1], reverse=True)
print('Percent of missing data')
df_miss[0:10]

Percent of missing data

Output[20]:

[('PoolQC', 99.69),

('MiscFeature', 96.4),

('Alley', 93.21),

('Fence', 80.43),

('FireplaceQu', 48.68),

('LotFrontage', 16.66),

('GarageYrBlt', 5.45),

('GarageFinish', 5.45),

('GarageQual', 5.45),

('GarageCond', 5.45)]

Input[21]:

接下来,我们将分别对每一列填充缺失值

Input[22]:

Input[23]:

Input[24]:

Output[14]:

Percent of missing data

[('MSSubClass', 0.0),

('MSZoning', 0.0),

('LotFrontage', 0.0),

('LotArea', 0.0),

('Street', 0.0),

('Alley', 0.0),

('LotShape', 0.0),

('LandContour', 0.0),

('Utilities', 0.0),

('LotConfig', 0.0)]

从上面的结果可以看到,所有缺失值已经填充完毕

调整分布倾斜的特征

Input[25]:

Input[26]:

# Create box plots for all numeric features
sns.set_style("white")
f, ax = plt.subplots(figsize=(8, 7))
ax.set_xscale("log")
ax = sns.boxplot(data=all_features[numeric] , orient="h", palette="Set1")
ax.xaxis.grid(False)
ax.set(ylabel="Feature names")
ax.set(xlabel="Numeric values")
ax.set(title="Numeric Distribution of Features")
sns.despine(trim=True, left=True)

Input[27]:

Output[27]:

There are 25 numerical features with Skew > 0.5 :

MiscVal 21.939672

PoolArea 17.688664

LotArea 13.109495

LowQualFinSF 12.084539

3SsnPorch 11.372080

KitchenAbvGr 4.300550

BsmtFinSF2 4.144503

EnclosedPorch 4.002344

ScreenPorch 3.945101

BsmtHalfBath 3.929996

dtype: float64

使用 scipy 的函数 boxcox1来进行 Box-Cox 转换,将数据正态化

Input[28]:
# Normalize skewed features
for i in skew_index:
    all_features[i] = boxcox1p(all_features[i], 
    boxcox_normmax(all_features[i] + 1))

Input[29]:

从上图可以看到,所有特征都看上去呈正态分布了。

创建一些有用的特征

机器学习模型对复杂模型的认知较差,因此我们需要用我们的直觉来构建有效的特征,从而帮助模型更加有效的学习。

特征转换

通过对特征取对数或者平方,可以创造更多的特征,这些操作有利于发掘潜在的有用特征。

对集合特征进行编码

对集合特征进行数值编码,使得机器学习模型能够处理这些特征。

all_features = pd.get_dummies(all_features).reset_index(drop=True)
all_features.shape

(2917, 379)

all_features.head()
all_features.shape
(2917, 379)

重新创建训练集和测试集

X = all_features.iloc[:len(train_labels), :]
X_test = all_features.iloc[len(train_labels):, :]
X.shape, train_labels.shape, X_test.shape

((1458, 378), (1458,), (1459, 378))

对训练集中的部分特征进行可视化

# Finding numeric features
numeric_dtypes = ['int16', 'int32', 'int64', 'float16', 'float32', 'float64']
numeric = []
for i in X.columns:
    if X[i].dtype in numeric_dtypes:
        if i in ['TotalSF', 'Total_Bathrooms','Total_porch_sf','haspool','hasgarage','hasbsmt','hasfireplace']:
            pass
        else:
            numeric.append(i)
# visualising some more outliers in the data values
fig, axs = plt.subplots(ncols=2, nrows=0, figsize=(12, 150))
plt.subplots_adjust(right=2)
plt.subplots_adjust(top=2)
sns.color_palette("husl", 8)
for i, feature in enumerate(list(X[numeric]), 1):
    if(feature=='MiscVal'):
        break
    plt.subplot(len(list(numeric)), 3, i)
    sns.scatterplot(x=feature, y='SalePrice', hue='SalePrice', palette='Blues', data=train)
        
    plt.xlabel('{}'.format(feature), size=15,labelpad=12.5)
    plt.ylabel('SalePrice', size=15, labelpad=12.5)
    
    for j in range(2):
        plt.tick_params(axis='x', labelsize=12)
        plt.tick_params(axis='y', labelsize=12)
    
    plt.legend(loc='best', prop={'size': 10})
        
plt.show()

模型训练

模型训练过程中的重要细节

  • 交叉验证:使用12-折交叉验证
  • 模型:在每次交叉验证中,同时训练七个模型(ridge, svr, gradient boosting, random forest, xgboost, lightgbm regressors)
  • Stacking 方法:使用xgboot训练了元 StackingCVRegressor 学习器
  • 模型融合:所有训练的模型都会在不同程度上过拟合,因此,为了做出最终的预测,将这些模型进行了融合,得到了鲁棒性更强的预测结果

初始化交叉验证,定义误差评估指标

# Setup cross validation folds
kf = KFold(n_splits=12, random_state=42, shuffle=True)

建立模型

训练模型

计算每个模型的交叉验证的得分

scores = {}

score = cv_rmse(lightgbm)
print("lightgbm: {:.4f} ({:.4f})".format(score.mean(), score.std()))
scores['lgb'] = (score.mean(), score.std())

lightgbm: 0.1159 (0.0167)

score = cv_rmse(xgboost)
print("xgboost: {:.4f} ({:.4f})".format(score.mean(), score.std()))
scores['xgb'] = (score.mean(), score.std())

xgboost: 0.1364 (0.0175)

score = cv_rmse(svr)
print("SVR: {:.4f} ({:.4f})".format(score.mean(), score.std()))
scores['svr'] = (score.mean(), score.std())

SVR: 0.1094 (0.0200)

score = cv_rmse(ridge)
print("ridge: {:.4f} ({:.4f})".format(score.mean(), score.std()))
scores['ridge'] = (score.mean(), score.std())

ridge: 0.1101 (0.0161)

score = cv_rmse(rf)
print("rf: {:.4f} ({:.4f})".format(score.mean(), score.std()))
scores['rf'] = (score.mean(), score.std())

rf: 0.1366 (0.0188

score = cv_rmse(gbr)
print("gbr: {:.4f} ({:.4f})".format(score.mean(), score.std()))
scores['gbr'] = (score.mean(), score.std())
gbr: 0.1121 (0.0164)

拟合模型

print('stack_gen')
stack_gen_model = stack_gen.fit(np.array(X), np.array(train_labels))
stack_gen
print('lightgbm')
lgb_model_full_data = lightgbm.fit(X, train_labels)

lightgbm

print('xgboost')
xgb_model_full_data = xgboost.fit(X, train_labels)

xgboost

print('Svr')
svr_model_full_data = svr.fit(X, train_labels)

Svr

print('Ridge')
ridge_model_full_data = ridge.fit(X, train_labels)

Ridge

print('RandomForest')
rf_model_full_data = rf.fit(X, train_labels)

RandomForest

print('GradientBoosting')
gbr_model_full_data = gbr.fit(X, train_labels)

GradientBoosting

融合各个模型,并进行最终预测

# Blend models in order to make the final predictions more robust to overfitting
def blended_predictions(X):
    return ((0.1 * ridge_model_full_data.predict(X)) + \
            (0.2 * svr_model_full_data.predict(X)) + \
            (0.1 * gbr_model_full_data.predict(X)) + \
            (0.1 * xgb_model_full_data.predict(X)) + \
            (0.1 * lgb_model_full_data.predict(X)) + \
            (0.05 * rf_model_full_data.predict(X)) + \
            (0.35 * stack_gen_model.predict(np.array(X))))
# Get final precitions from the blended model
blended_score = rmsle(train_labels, blended_predictions(X))
scores['blended'] = (blended_score, 0)
print('RMSLE score on train data:')
print(blended_score)

RMSLE score on train data: 0.07537440195302639

各模型性能比较

从上图可以看出,融合后的模型性能最好,RMSE 仅为 0.075,该融合模型用于最终预测。

提交预测结果

(1459, 2)

# Append predictions from blended models
submission.iloc[:,1] = np.floor(np.expm1(blended_predictions(X_test)))
 
# Fix outleir predictions
q1 = submission['SalePrice'].quantile(0.0045)
q2 = submission['SalePrice'].quantile(0.99)
submission['SalePrice'] = submission['SalePrice'].apply(lambda x: x if x > q1 else x*0.77)
submission['SalePrice'] = submission['SalePrice'].apply(lambda x: x if x < q2 else x*1.1)
submission.to_csv("submission_regression1.csv", index=False)

(*本文为 AI科技大本营翻译文章,转载请联系 1092722531

原文发布于微信公众号 - AI科技大本营(rgznai100)

原文发表时间:2019-07-11

本文参与腾讯云自媒体分享计划,欢迎正在阅读的你也加入,一起分享。

发表于

我来说两句

0 条评论
登录 后参与评论

扫码关注云+社区

领取腾讯云代金券