Feature Engineering 特征工程 1. Baseline Model

learn from https://www.kaggle.com/learn/feature-engineering

下一篇：Feature Engineering 特征工程 2. Categorical Encodings

---

1. 读取数据

预测任务：用户是否会下载APP，当其点击广告以后
数据集：ks-projects-201801.csv

• 读取数据，指定两个特征'deadline','launched'，parse_dates解析为时间

ks = pd.read_csv('ks-projects-201801.csv',parse_dates=['deadline','launched'])

预测Kickstarter项目是否会成功。state作为结果label
可以使用类别category，货币currency，资金目标funding goal，国家country以及启动时间launched等特征

2. 处理label

• 准备标签列，看看有哪些值，转换成可用的数字格式

pd.unique(ks.state)

有6种数值

array(['failed', 'canceled', 'successful', 'live', 'undefined',
       'suspended'], dtype=object)

每种多少个？按state分组，每组中ID行数有多少

ks.groupby('state')['ID'].count()

state
canceled       38779
failed        197719
live            2799
successful    133956
suspended       1846
undefined       3562
Name: ID, dtype: int64

• 简单处理下标签列，正在进行的项目live丢弃，successful的标记为1，其余的为0

ks = ks.query('state != "live"') # live行不要
ks = ks.assign(outcome=(ks['state']=='successful').astype(int))
# label 转成1,0，int型

3. 添加特征

• 把launched时间拆分成，年月日小时，作为新的特征

ks = ks.assign(hour=ks.launched.dt.hour,
               day=ks.launched.dt.day,
               month=ks.launched.dt.month,
               year=ks.launched.dt.year)
ks.head()

• 转换文字特征category, currency, country为数字

from sklearn.preprocessing import LabelEncoder

cat_features = ['category','currency','country']
encoder = LabelEncoder()

encoded = ks[cat_features].apply(encoder.fit_transform)
encoded.head(10)

• 将选择使用的特征合并在一个数据里

X = ks[['goal', 'hour', 'day', 'month', 'year', 'outcome']].join(encoded)
X.head()

4. 数据集切片

• 数据切片，按比例分成训练集、验证集、测试集（0.8，0.1，0.1）
• 更高级的简单做法sklearn.model_selection.StratifiedShuffleSplit

valid_ratio = 0.1
valid_size = int(len(X)*valid_ratio)
train = X[ : -2*valid_size]
valid = X[-2*valid_size : -valid_size]
test = X[-valid_size : ]

需要关注下，label 在每个数据集中的占比是否接近

for each in [train, valid, test]:
    print("Outcome fraction = {:.4f}".format(each.outcome.mean()))

Outcome fraction = 0.3570
Outcome fraction = 0.3539
Outcome fraction = 0.3542

5. 训练

• 使用LightGBM模型进行训练

机器学习算法之LightGBM

feature_cols = train.columns.drop('outcome')

dtrain = lgb.Dataset(train[feature_cols], label=train['outcome'])
dvalid = lgb.Dataset(valid[feature_cols], label=valid['outcome'])

param = {'num_leaves': 64, 'objective': 'binary'}
param['metric'] = 'auc'
num_round = 1000
bst = lgb.train(param, dtrain, num_round, valid_sets=[dvalid],
				early_stopping_rounds=10, verbose_eval=False)

6. 预测

• 对测试集进行预测

from sklearn import metrics
ypred = bst.predict(test[feature_cols])
score = metrics.roc_auc_score(test['outcome'], ypred)

print(f"Test AUC score: {score}")

---

下一篇：Feature Engineering 特征工程 2. Categorical Encodings