一起加入这次沉浸式学习吧,本次分享的方案基本上包好了结构化数据比赛的基本流程：数据分析、数据预处理，特征工程、模型训练以及模型融合，大家可以留在周末学习一波。

比赛名称:Sberbank Russian Housing Market 比赛链接：https://www.kaggle.com/c/sberbank-russian-housing-market

1

竞赛背景

住房成本需要消费者和开发商的大量投资。 在规划预算时（无论是个人预算还是公司预算），任何一方不到最后就是不确定其中哪一项是最大开支。 俄罗斯最早、最大的银行Sberbank通过预测房地产价格来帮助客户预测预算，因此租户，开发商和贷方在签订租约或购买建筑物时更加相互信任。

尽管俄罗斯的住房市场相对稳定，但该国动荡的经济形势使得根据公寓价格预测成为一项独特的挑战。 房屋数量（如卧室数量和位置）之间复杂的相互关系足以使价格预测变得复杂。 加上不稳定的经济因素，意味着Sberbank及其客户需要的不仅仅是其机器学习库中的简单回归模型。

在这场竞赛中，Sberbank向Kagglers提出挑战，要求他们开发使用多种特征来预测房地产价格的算法。 竞争对手将依靠丰富的数据集，其中包括住房数据和宏观经济模式。 准确的预测模型将使Sberbank在不确定的经济环境中为其客户提供更多的确定性。

2

赛题解析

这个竞赛目的是预测每一处房产的销售价格。目标变量在train.csv中称为price_doc。训练数据为2011年8月至2015年6月，测试集为2015年7月至2016年5月。该数据集还包括俄罗斯经济和金融部门的总体状况信息，因此您可以专注于为每个房产生成准确的价格预测，而无需猜测商业周期将如何变化。

3

竞赛数据

train.csv，test.csv：有关单个交易的信息。 这些行由“ id”字段索引，该字段引用单个事务（特定属性在单独的事务中可能出现多次）。 这些文件还包括有关每个属性的本地区域的补充信息。

macro.csv：有关俄罗斯宏观经济和金融部门的数据（可以根据“时间戳”与训练集和测试集合并）

data_dictionary.txt：其他数据文件中可用字段的说明

sample_submission.csv：格式正确的示例提交文件 其中字段比较多，我们可以通过data_dictionary文件可以发现至少有200+个字段，所以本次比赛的数据还是比较丰富，比较客观，同时也具有研究价值。

4

数据分析

来源：https://www.kaggle.com/sudalairajkumar/simple-exploration-notebook-sberbank

• 房产价格分布 我们将价格按照从小到大排序，画出如下每处房产价格分布：

plt.figure(figsize=(8,6))
plt.scatter(range(train_df.shape[0]), np.sort(train_df.price_doc.values))
plt.xlabel('index', fontsize=12)
plt.ylabel('price', fontsize=12)
plt.show()

• 房产价格随着时间变化趋势

train_df['yearmonth'] = train_df['timestamp'].apply(lambda x: x[:4]+x[5:7])
grouped_df = train_df.groupby('yearmonth')['price_doc'].aggregate(np.median).reset_index()plt.figure(figsize=(12,8))
sns.barplot(grouped_df.yearmonth.values, grouped_df.price_doc.values, alpha=0.8, color=color[2])
plt.ylabel('Median Price', fontsize=12)
plt.xlabel('Year Month', fontsize=12)
plt.xticks(rotation='vertical')
plt.show()

• 特征重要性较高的特征 因为有292个变量，让我们构建一个基本的xgboost模型，然后先研究重要的变量。

for f in train_df.columns:if train_df[f].dtype=='object':lbl = preprocessing.LabelEncoder()lbl.fit(list(train_df[f].values)) train_df[f] = lbl.transform(list(train_df[f].values))train_y = train_df.price_doc.values
train_X = train_df.drop(["id", "timestamp", "price_doc"], axis=1)xgb_params = {'eta': 0.05,'max_depth': 8,'subsample': 0.7,'colsample_bytree': 0.7,'objective': 'reg:linear','eval_metric': 'rmse','silent': 1
}
dtrain = xgb.DMatrix(train_X, train_y, feature_names=train_X.columns.values)
model = xgb.train(dict(xgb_params, silent=0), dtrain, num_boost_round=100)# plot the important features #
fig, ax = plt.subplots(figsize=(12,18))
xgb.plot_importance(model, max_num_features=50, height=0.8, ax=ax)
plt.show()

因此，数据特征中的重要性前5个变量及其描述为：

full_sq-以平方米为单位的总面积，包括凉廊，阳台和其他非住宅区
life_sq-居住面积（平方米），不包括凉廊，阳台和其他非居住区
floor-对于房屋，建筑物的当前层数
max_floor-建筑物中的总楼层数
build_year-建造年份

full_seq与房产价格的分布

ulimit = np.percentile(train_df.price_doc.values, 99.5)
llimit = np.percentile(train_df.price_doc.values, 0.5)
train_df['price_doc'].ix[train_df['price_doc']>ulimit] = ulimit
train_df['price_doc'].ix[train_df['price_doc']<llimit] = llimitcol = "full_sq"
ulimit = np.percentile(train_df[col].values, 99.5)
llimit = np.percentile(train_df[col].values, 0.5)
train_df[col].ix[train_df[col]>ulimit] = ulimit
train_df[col].ix[train_df[col]<llimit] = llimitplt.figure(figsize=(12,12))
sns.jointplot(x=np.log1p(train_df.full_sq.values), y=np.log1p(train_df.price_doc.values), size=10)
plt.ylabel('Log of Price', fontsize=12)
plt.xlabel('Log of Total area in square metre', fontsize=12)
plt.show()

life_sq与房产价格分布

col = "life_sq"
train_df[col].fillna(0, inplace=True)
ulimit = np.percentile(train_df[col].values, 95)
llimit = np.percentile(train_df[col].values, 5)
train_df[col].ix[train_df[col]>ulimit] = ulimit
train_df[col].ix[train_df[col]<llimit] = llimitplt.figure(figsize=(12,12))
sns.jointplot(x=np.log1p(train_df.life_sq.values), y=np.log1p(train_df.price_doc.values), kind='kde', size=10)
plt.ylabel('Log of Price', fontsize=12)
plt.xlabel('Log of living area in square metre', fontsize=12)
plt.show()

楼层与房产价格中位数分布

grouped_df = train_df.groupby('floor')['price_doc'].aggregate(np.median).reset_index()
plt.figure(figsize=(12,8))
sns.pointplot(grouped_df.floor.values, grouped_df.price_doc.values, alpha=0.8, color=color[2])
plt.ylabel('Median Price', fontsize=12)
plt.xlabel('Floor number', fontsize=12)
plt.xticks(rotation='vertical')
plt.show()

5

Top 1% 代码分享

代码链接：https://github.com/LenzDu/Kaggle-Competition-Sberbank

• Data.py: 数据清洗以及特征工程

• Exploration.py: 数据分析

• Model.py: XGBoost模型

• BaseModel.py: 基线模型：RandomForestRegressor、GradientBoostingRegressor、Lasso等

• lightGBM.py: lightGBM模型

• Stacking.py: model stacking (final model)：模型融合

因为代码比较清晰简洁，非常适合数据挖掘的新手解读学习，其中作者写的Stacking也是非常漂亮，我们可以感受下：

Stacking是通过一个元分类器或者元回归器整合多个模型的集成学习技术。基础模型利用整个训练集做训练，元模型利用基础模型做特征进行训练。一般Stacking多使用不同类型的基础模型

import numpy as np
import pandas as pd
from sklearn.model_selection import ShuffleSplit, cross_val_score
from sklearn.cross_validation import KFold
from sklearn.ensemble import AdaBoostRegressor, RandomForestRegressor, ExtraTreesRegressor, GradientBoostingRegressor
from sklearn.neural_network import MLPRegressor
from sklearn.preprocessing import Imputer
from sklearn.tree import DecisionTreeRegressor
from sklearn.linear_model import LinearRegression
import xgboost as xgb
import lightgbm as lgb
from sklearn.preprocessing import StandardScaler# 封装一下lightgbm让其可以在stacking里面被调用
class LGBregressor(object):def __init__(self,params):self.params = paramsdef fit(self, X, y, w):y /= 10000000# self.scaler = StandardScaler().fit(y)# y = self.scaler.transform(y)split = int(X.shape[0] * 0.8)indices = np.random.permutation(X.shape[0])train_id, test_id = indices[:split], indices[split:]x_train, y_train, w_train, x_valid, y_valid,  w_valid = X[train_id], y[train_id], w[train_id], X[test_id], y[test_id], w[test_id],d_train = lgb.Dataset(x_train, y_train, weight=w_train)d_valid = lgb.Dataset(x_valid, y_valid, weight=w_valid)partial_bst = lgb.train(self.params, d_train, 10000, valid_sets=d_valid, early_stopping_rounds=50)num_round = partial_bst.best_iterationd_all = lgb.Dataset(X, label = y, weight=w)self.bst = lgb.train(self.params, d_all, num_round)def predict(self, X):return self.bst.predict(X) * 10000000# return self.scaler.inverse_transform(self.bst.predict(X))# 封装一下xgboost让其可以在stacking里面被调用
class XGBregressor(object):def __init__(self, params):self.params = paramsdef fit(self, X, y, w=None):if w==None:w = np.ones(X.shape[0])split = int(X.shape[0] * 0.8)indices = np.random.permutation(X.shape[0])train_id, test_id = indices[:split], indices[split:]x_train, y_train, w_train, x_valid, y_valid,  w_valid = X[train_id], y[train_id], w[train_id], X[test_id], y[test_id], w[test_id],d_train = xgb.DMatrix(x_train, label=y_train, weight=w_train)d_valid = xgb.DMatrix(x_valid, label=y_valid, weight=w_valid)watchlist = [(d_train, 'train'), (d_valid, 'valid')]partial_bst = xgb.train(self.params, d_train, 10000, early_stopping_rounds=50, evals = watchlist, verbose_eval=100)num_round = partial_bst.best_iterationd_all = xgb.DMatrix(X, label = y, weight=w)self.bst = xgb.train(self.params, d_all, num_round)def predict(self, X):test = xgb.DMatrix(X)return self.bst.predict(test)# This object modified from Wille on https://dnc1994.com/2016/05/rank-10-percent-in-first-kaggle-competition-en/
class Ensemble(object):def __init__(self, n_folds, stacker, base_models):self.n_folds = n_foldsself.stacker = stackerself.base_models = base_modelsdef fit_predict(self, trainDf, testDf):X = trainDf.drop(['price_doc', 'w'], 1).valuesy = trainDf['price_doc'].valuesw = trainDf['w'].valuesT = testDf.valuesX_fillna = trainDf.drop(['price_doc', 'w'], 1).fillna(-999).valuesT_fillna = testDf.fillna(-999).valuesfolds = list(KFold(len(y), n_folds=self.n_folds, shuffle=True))S_train = np.zeros((X.shape[0], len(self.base_models)))S_test = np.zeros((T.shape[0], len(self.base_models)))for i, clf in enumerate(self.base_models):print('Training base model ' + str(i+1) + '...')S_test_i = np.zeros((T.shape[0], len(folds)))for j, (train_idx, test_idx) in enumerate(folds):print('Training round ' + str(j+1) + '...')if clf not in [xgb1,lgb1]: # sklearn models cannot handle missing values.X = X_fillnaT = T_fillnaX_train = X[train_idx]y_train = y[train_idx]w_train = w[train_idx]X_holdout = X[test_idx]# w_holdout = w[test_idx]# y_holdout = y[test_idx]clf.fit(X_train, y_train, w_train)y_pred = clf.predict(X_holdout)S_train[test_idx, i] = y_predS_test_i[:, j] = clf.predict(T)S_test[:, i] = S_test_i.mean(1)self.S_train, self.S_test, self.y = S_train, S_test, y  # for diagnosis purposeself.corr = pd.concat([pd.DataFrame(S_train),trainDf['price_doc']],1).corr() # correlation of predictions by different models.# cv_stack = ShuffleSplit(n_splits=6, test_size=0.2)# score_stacking = cross_val_score(self.stacker, S_train, y, cv=cv_stack, n_jobs=1, scoring='neg_mean_squared_error')# print(np.sqrt(-score_stacking.mean())) # CV result of stackingself.stacker.fit(S_train, y)y_pred = self.stacker.predict(S_test)return y_predif __name__ == "__main__":trainDf = pd.read_csv('train_featured.csv')testDf = pd.read_csv('test_featured.csv')params1 = {'eta':0.05, 'max_depth':5, 'subsample':0.8, 'colsample_bytree':0.8, 'min_child_weight':1,'gamma':0, 'silent':1, 'objective':'reg:linear', 'eval_metric':'rmse'}xgb1 = XGBregressor(params1)params2 = {'booster':'gblinear', 'alpha':0,# for gblinear, delete this line if change back to gbtree'eta':0.1, 'max_depth':2, 'subsample':1, 'colsample_bytree':1, 'min_child_weight':1,'gamma':0, 'silent':1, 'objective':'reg:linear', 'eval_metric':'rmse'}xgb2 = XGBregressor(params2)RF = RandomForestRegressor(n_estimators=500, max_features=0.2)ETR = ExtraTreesRegressor(n_estimators=500, max_features=0.3, max_depth=None)Ada = AdaBoostRegressor(DecisionTreeRegressor(max_depth=15),n_estimators=200)GBR = GradientBoostingRegressor(n_estimators=200,max_depth=5,max_features=0.5)LR =LinearRegression()params_lgb = {'objective':'regression','metric':'rmse','learning_rate':0.05,'max_depth':-1,'sub_feature':0.7,'sub_row':1,'num_leaves':15,'min_data':30,'max_bin':20,'bagging_fraction':0.9,'bagging_freq':40,'verbosity':0}lgb1 = LGBregressor(params_lgb)E = Ensemble(5, xgb2, [xgb1,lgb1,RF,ETR,Ada,GBR])prediction = E.fit_predict(trainDf, testDf)output = pd.read_csv('test.csv')output = output[['id']]output['price_doc'] = predictionoutput.to_csv(r'Ensemble\Submission_Stack.csv',index=False)

6

我们还可以学习到什么

一般每个比赛的discussion部分，我们可以看到前排方案的讨论交流，感觉读了他们分享的总结以及简介比代码获得收益更大。

——ChallengeHub

链接为：https://www.kaggle.com/c/sberbank-russian-housing-market/discussion/35684

从第一名分享的方案中，对我收益比较大的是：

1. 没有对目标变量直接预测，而是对单位平方米的价格进行预测，之后转化

2. 尝试很多的独立模型，这里指的是因为他们发现有两个变量放在一块导致模型差异很大（Investment 和OwnerOccupier），然后将两个变量置于两组不同的特征输入给模型

3. 去除异常值，单独训练模型

END