kaggle競賽-保險轉化-homesite

• 時間格式的轉化
• 查看數據類型
• 查看DataFrame的詳細信息
• 填充缺失值
• category 數據類型轉化
• 模型參數設定
• 結論

• 該項目是針對kaggle中的homesite進行的算法預測，使用xgboost的sklearn接口，進行數據建模，購買預測。

import pandas as pd
import numpy as np
import xgboost as xgb
from sklearn.model_selection import StratifiedKFold
from sklearn.model_selection import GridSearchCV

train = pd.read_csv('train.csv')
test = pd.read_csv('test.csv')
train.head()

	QuoteNumber	Original_Quote_Date	QuoteConversion_Flag	Field6	Field7	Field8	Field9	Field10	Field11	Field12	...	GeographicField59A	GeographicField59B	GeographicField60A	GeographicField60B	GeographicField61A	GeographicField61B	GeographicField62A	GeographicField62B	GeographicField63	GeographicField64
0	1	2013-08-16	0	B	23	0.9403	0.0006	965	1.0200	N	...	9	9	-1	8	-1	18	-1	10	N	CA
1	2	2014-04-22	0	F	7	1.0006	0.0040	548	1.2433	N	...	10	10	-1	11	-1	17	-1	20	N	NJ
2	4	2014-08-25	0	F	7	1.0006	0.0040	548	1.2433	N	...	15	18	-1	21	-1	11	-1	8	N	NJ
3	6	2013-04-15	0	J	10	0.9769	0.0004	1,165	1.2665	N	...	6	5	-1	10	-1	9	-1	21	N	TX
4	8	2014-01-25	0	E	23	0.9472	0.0006	1,487	1.3045	N	...	18	22	-1	10	-1	11	-1	12	N	IL

5 rows × 299 columns

train=train.drop('QuoteNumber',axis=1)

test = test.drop('QuoteNumber', axis=1)

時間格式的轉化

train['Date']=pd.to_datetime(train['Original_Quote_Date'])
train= train.drop('Original_Quote_Date',axis=1)

test['Date']=pd.to_datetime(test['Original_Quote_Date'])
test= test.drop('Original_Quote_Date',axis=1)

train['year']=train['Date'].dt.year

train['month']=train['Date'].dt.month
train['weekday']=train['Date'].dt.weekday

train.head()

	QuoteConversion_Flag	Field6	Field7	Field8	Field9	Field10	Field11	Field12	CoverageField1A	CoverageField1B	...	GeographicField61A	GeographicField61B	GeographicField62A	GeographicField62B	GeographicField63	GeographicField64	Date	year	month	weekday
0	0	B	23	0.9403	0.0006	965	1.0200	N	17	23	...	-1	18	-1	10	N	CA	2013-08-16	2013	8	4
1	0	F	7	1.0006	0.0040	548	1.2433	N	6	8	...	-1	17	-1	20	N	NJ	2014-04-22	2014	4	1
2	0	F	7	1.0006	0.0040	548	1.2433	N	7	12	...	-1	11	-1	8	N	NJ	2014-08-25	2014	8	0
3	0	J	10	0.9769	0.0004	1,165	1.2665	N	3	2	...	-1	9	-1	21	N	TX	2013-04-15	2013	4	0
4	0	E	23	0.9472	0.0006	1,487	1.3045	N	8	13	...	-1	11	-1	12	N	IL	2014-01-25	2014	1	5

5 rows × 301 columns

test['year']=test['Date'].dt.year
test['month']=test['Date'].dt.month
test['weekday']=test['Date'].dt.weekday

train = train.drop('Date', axis=1)  
test = test.drop('Date', axis=1)

查看數據類型

train.dtypes

QuoteConversion_Flag      int64
Field6                   object
Field7                    int64
Field8                  float64
Field9                  float64
Field10                  object
Field11                 float64
Field12                  object
CoverageField1A           int64
CoverageField1B           int64
CoverageField2A           int64
CoverageField2B           int64
CoverageField3A           int64
CoverageField3B           int64
CoverageField4A           int64
CoverageField4B           int64
CoverageField5A           int64
CoverageField5B           int64
CoverageField6A           int64
CoverageField6B           int64
CoverageField8           object
CoverageField9           object
CoverageField11A          int64
CoverageField11B          int64
SalesField1A              int64
SalesField1B              int64
SalesField2A              int64
SalesField2B              int64
SalesField3               int64
SalesField4               int64
                         ...   
GeographicField50B        int64
GeographicField51A        int64
GeographicField51B        int64
GeographicField52A        int64
GeographicField52B        int64
GeographicField53A        int64
GeographicField53B        int64
GeographicField54A        int64
GeographicField54B        int64
GeographicField55A        int64
GeographicField55B        int64
GeographicField56A        int64
GeographicField56B        int64
GeographicField57A        int64
GeographicField57B        int64
GeographicField58A        int64
GeographicField58B        int64
GeographicField59A        int64
GeographicField59B        int64
GeographicField60A        int64
GeographicField60B        int64
GeographicField61A        int64
GeographicField61B        int64
GeographicField62A        int64
GeographicField62B        int64
GeographicField63        object
GeographicField64        object
year                      int64
month                     int64
weekday                   int64
Length: 300, dtype: object

查看DataFrame的詳細信息

train.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 260753 entries, 0 to 260752
Columns: 300 entries, QuoteConversion_Flag to weekday
dtypes: float64(6), int64(267), object(27)
memory usage: 596.8+ MB

填充缺失值

train = train.fillna(-999)
test = test.fillna(-999)

category 數據類型轉化

from sklearn import preprocessing
features = list(train.columns[1:])  
for i in features:
    if train[i].dtype=='object':
        le=preprocessing.LabelEncoder()
        le.fit(list(train[i].values)+list(test[i].values))
        train[i] = le.transform(list(train[i].values))
        test[i] = le.transform(list(test[i].values))

模型參數設定

#brute force scan for all parameters, here are the tricks
#usually max_depth is 6,7,8
#learning rate is around 0.05, but small changes may make big diff
#tuning min_child_weight subsample colsample_bytree can have 
#much fun of fighting against overfit 
#n_estimators is how many round of boosting
#finally, ensemble xgboost with multiple seeds may reduce variance

xgb_model = xgb.XGBClassifier()

parameters = {'nthread':[4], #when use hyperthread, xgboost may become slower
              'objective':['binary:logistic'],
              'learning_rate': [0.05,0.1], #so called `eta` value
              'max_depth': [6],
              'min_child_weight': [11],
              'silent': [1],
              'subsample': [0.8],
              'colsample_bytree': [0.7],
              'n_estimators': [5], #number of trees, change it to 1000 for better results
              'missing':[-999],
              'seed': [1337]}

sfolder = StratifiedKFold(n_splits=5,random_state=42,shuffle=True)
clf= GridSearchCV(xgb_model,parameters,n_jobs=4,cv=sfolder.split(train[features], train["QuoteConversion_Flag"]),scoring='roc_auc',
                   verbose=2, refit=True,return_train_score=True)
clf.fit(train[features], train["QuoteConversion_Flag"])

Fitting 5 folds for each of 2 candidates, totalling 10 fits


[Parallel(n_jobs=4)]: Done  10 out of  10 | elapsed:  2.4min finished





GridSearchCV(cv=<generator object _BaseKFold.split at 0x0000000018459888>,
       error_score='raise',
       estimator=XGBClassifier(base_score=0.5, booster='gbtree', colsample_bylevel=1,
       colsample_bytree=1, gamma=0, learning_rate=0.1, max_delta_step=0,
       max_depth=3, min_child_weight=1, missing=None, n_estimators=100,
       n_jobs=1, nthread=None, objective='binary:logistic', random_state=0,
       reg_alpha=0, reg_lambda=1, scale_pos_weight=1, seed=None,
       silent=True, subsample=1),
       fit_params=None, iid=True, n_jobs=4,
       param_grid={'nthread': [4], 'objective': ['binary:logistic'], 'learning_rate': [0.05, 0.1], 'max_depth': [6], 'min_child_weight': [11], 'silent': [1], 'subsample': [0.8], 'colsample_bytree': [0.7], 'n_estimators': [5], 'missing': [-999], 'seed': [1337]},
       pre_dispatch='2*n_jobs', refit=True, return_train_score=True,
       scoring='roc_auc', verbose=2)

clf.grid_scores_

c:\anaconda3\envs\nlp\lib\site-packages\sklearn\model_selection\_search.py:761: DeprecationWarning: The grid_scores_ attribute was deprecated in version 0.18 in favor of the more elaborate cv_results_ attribute. The grid_scores_ attribute will not be available from 0.20
  DeprecationWarning)





[mean: 0.94416, std: 0.00118, params: {'colsample_bytree': 0.7, 'learning_rate': 0.05, 'max_depth': 6, 'min_child_weight': 11, 'missing': -999, 'n_estimators': 5, 'nthread': 4, 'objective': 'binary:logistic', 'seed': 1337, 'silent': 1, 'subsample': 0.8},
 mean: 0.94589, std: 0.00120, params: {'colsample_bytree': 0.7, 'learning_rate': 0.1, 'max_depth': 6, 'min_child_weight': 11, 'missing': -999, 'n_estimators': 5, 'nthread': 4, 'objective': 'binary:logistic', 'seed': 1337, 'silent': 1, 'subsample': 0.8}]

pd.DataFrame(clf.cv_results_['params'])

	colsample_bytree	learning_rate	max_depth	min_child_weight	missing	n_estimators	nthread	objective	seed	silent	subsample
0	0.7	0.05	6	11	-999	5	4	binary:logistic	1337	1	0.8
1	0.7	0.10	6	11	-999	5	4	binary:logistic	1337	1	0.8

best_parameters, score, _ = max(clf.grid_scores_, key=lambda x: x[1])
print('Raw AUC score:', score)
for param_name in sorted(best_parameters.keys()):
    print("%s: %r" % (param_name, best_parameters[param_name]))

Raw AUC score: 0.9458947562485674
colsample_bytree: 0.7
learning_rate: 0.1
max_depth: 6
min_child_weight: 11
missing: -999
n_estimators: 5
nthread: 4
objective: 'binary:logistic'
seed: 1337
silent: 1
subsample: 0.8


c:\anaconda3\envs\nlp\lib\site-packages\sklearn\model_selection\_search.py:761: DeprecationWarning: The grid_scores_ attribute was deprecated in version 0.18 in favor of the more elaborate cv_results_ attribute. The grid_scores_ attribute will not be available from 0.20
  DeprecationWarning)

test_probs = clf.predict_proba(test[features])[:,1]

sample = pd.read_csv('sample_submission.csv')
sample.QuoteConversion_Flag = test_probs
sample.to_csv("xgboost_best_parameter_submission.csv", index=False)

clf.best_estimator_.predict_proba(test[features])

array([[0.6988076 , 0.3011924 ],
       [0.6787684 , 0.3212316 ],
       [0.6797658 , 0.32023418],
       ...,
       [0.5018287 , 0.4981713 ],
       [0.6988076 , 0.3011924 ],
       [0.62464744, 0.37535256]], dtype=float32)

下面的截斷值0.5能夠本身根據實際的項目設定截斷值

kears_result=pd.read_csv('keras_nn_test.csv')
result1=[1 if i>0.5 else 0 for i in kears_result['QuoteConversion_Flag']]
xgb_result=pd.read_csv('xgboost_best_parameter_submission.csv')
result2=[1 if i>0.5 else 0 for i in xgb_result['QuoteConversion_Flag']]
from sklearn import metrics
metrics.accuracy_score(result1,result2)

0.8566004740099864

metrics.confusion_matrix(result1,result2)

array([[148836,  24862],
       [    66,     72]], dtype=int64)

結論

• 對數據的時間進行了預處理
• 對數據中的category類型進行了label化，我以爲有必要對這個進行從新考慮，我的以爲應該使用one-hot進行category的處理，而不是LabelEncoder處理（疑慮）
• Label encoding在某些狀況下頗有用，可是場景限制不少。再舉一例：好比有[dog,cat,dog,mouse,cat]，咱們把其轉換爲[1,2,1,3,2]。這裏就產生了一個奇怪的現象：dog和mouse的平均值是cat。因此目前尚未發現標籤編碼的普遍使用。
• 獲得的模型對測試集進行處理，Raw AUC 0.94，而對應的準確率只有85%，實際上並無實際的分類效果，對於其實是0的，預測成1的太多了，也就是假陽性過高了，實際中的轉換率也不會很高。
• 其實模型還有不少能夠調整的參數都沒有調整，若是對調參有興趣的能夠查看美團的文本分類項目中的例子。