一、特征分析（EDA，探索性数据分析）

1.1 seaborn特征分析

roc_cure
lineplot("X", "y", data=df))

一个特征不同值对生的影响，有限个数：
barplot("X", "y", data=df)

连续且个数比较多
sns.distplot(train['SibSp'][train['Survived'] == 1], bins=50)
sns.distplot(train['SibSp'][train['Survived'] == 0], bins=50)
等价于
sns.distplot(train.loc[ train['Survived'] == 1, 'SibSp'], bins=50)
sns.distplot(train.loc[ train['Survived'] == 0, 'SibSp'], bins=50)

一个值分别对生死的影响
countplot("Embarked", hue='Survived', data=df)

1.2 特征概述

data.head(10)
data.describe()
data.describe().T
data.info()
train['Survived'].value_counts() #查看生存比重

二、特征选择、处理

2.1 连续值分隔处理

使用pd.cut自动分割
train['Age'] = pd.cut(train['Age'], 5, labels=[0, 1, 2, 3, 4])
手动分割
def ProcessLabel(val):
if val < 3:
return 0
elif val < 7:
return 1
else:
return 2
train['FamliySize'] = train['Sisbp'] + train['Parch'] + 1
train['FamLable'] = train[FamilySize].apply(ProcessLabel)

2.2 字符串处理

train['Embarked'] = train['Embarked'].map({'S': 0, 'P':1, 'S': 2})

2.3 缺失值处理

字符串填充：

train['Embarked'] = train['Embarked'].fillna('S')

使用均值填充：

avg = train['Age'].mean()
std = train['Age'].std()
age_null_count  = train['Age'].isnull().sum()
age_list = np.random.randint(avg-std, avg+std, size = age_null_count)
train.loc[train['Age'].isnull(), 'Age'] = age_list

当缺失较多时，使用回归模型预测值：
from sklearn.ensemble import RandomForestRegressor
import lightgbm as lgbm

data = train[['Age', 'Pclass', 'Sex', 'Title']]
data = pd.get_dummies(data)
model = RandomForestRegressor(n_estimators=128, n_jobs=-1)
# model = lgbm.LGBMRegressor(n_estimators=128, n_jobs=-1)
tr= data[data['Age'].notnull()].values
te = data[data['Age'].isnull()].values
tr_X = tr[:, 1:]
tr_y = tr[:, 0]
te_X = te[:, 1:]
model.fit(tr_X, tr_y)
pred_age = model.predict(te_X)
train.loc[data['Age'].isnull(), 'Age'] = pre_age

2.4 one hot 编码

一定要对all_data进行，否则容易训练集，测试集不匹配：

all_data = pd.get_dummise(all_data)


Emb = pd.get_dummies(all_data)
all_data = pd.concat([all_data, Emb], axis = 1)

2.5 数据合并分开

all_data = pd.concat([train, test], ignore_index = True)

分开：

train=all_data.loc[all_data['Survived'].notnull()]
test=all_data.loc[all_data['Survived'].isnull()]

2.6 特征缩放，标准化

from sklearn.preprocessing import StandardScaler
sc =StandardScaler()
data_new[['Amount', 'Hour']] =sc.fit_transform(data_new[['Amount', 'Hour']])
data_new.head()

三、模型调参

lgbm:

objective=(regression,binary/multiclass)

3.1 GridSearchCV参数寻优

import lightgbm as lgb
from sklearn.model_selection import cross_val_score
from sklearn.model_selection improt GridSearchCV

params = {'num_leaves': [32, 64, 128, 256, 1024], 'max_depth': [10, 20, 30, 60], 'learning_rate': [0.01, 0.05, 0.1], 'n_estimators': [100, 200, 300]}
model = lgb.LGBMClassifier()
gridS = GridSearchCV(model, params, cv=5, n_jobs=-1)
gridS.fit(X, y)
gridS.best_estimator_

四、结果

4.1 画roc曲线

需要最好是概率，如果是0, 1值的话，只有一个点，所以要使用lgb.train()，而不是LGBMClassifier()的模型

from sklearn.metrics import roc_curve
from matplotlib import pyplot as plt
import seaborn as sns

sns.set()
fpr, tpr, thresh = roc_curve(y, pred)
plt.plot(fpr, tpr)
plt.show()

4.2 求交叉准确率

from sklearn.model_selection import cross_val_score

score =  cross_val_score(model, X, y, scoring='accuracy', cv=5)
print(np.mean(score))

4.3 保存csv

res = pd.DataFrame({'PassageID': passage_id, 'Survived': pred.as_type(np.int32)})
res.to_csv('pred.csv', index=False)

Others

模型训练时报Input contains NaN, infinity or a value too large for dtype('float64')：
因为特征里包含nan