【特征工程构建】

类别特征目标编码

统计每一个类对于标签的均值，count等

eidmean = train_data.groupby('eid')['target'].mean()
train_data['eid_freq'] = train_data['eid'].map(train_data['eid'].value_counts())
train_data['eid_mean'] = train_data['eid'].map(eidmean)

test_data['eid_freq'] = test_data['eid'].map(test_data['eid'].value_counts())
test_data['eid_mean'] = test_data['eid'].map(eidmean)

查看对于目标的影响

train_data[['eid','target']].groupby(['eid'], as_index=False).mean().sort_values(by='target', ascending=False)

数值特征

采用交叉特征的方式，如行交叉与业务间相互交叉（加减乘）

num_features =['x7','x2','x1']

# 3. 特征交互与多项式特征
interaction_features = num_features

for i in range(len(interaction_features)):
    for j in range(i + 1, len(interaction_features)):
        interaction_name = interaction_features[i] + '_' + interaction_features[j]
        train_data[interaction_name] = train_data[interaction_features[i]] + train_data[interaction_features[j]]
        test_data[interaction_name] = test_data[interaction_features[i]] + test_data[interaction_features[j]]

分桶操作

bins = [0,10,12,14,17,19,32]  # 区间边界
# 使用cut函数进行分桶
train_data['day'] = pd.cut(train_data['day'], bins=bins,labels=False, include_lowest=True)
test_data['day'] = pd.cut(test_data['day'], bins=bins,labels=False, include_lowest=True)

时间转化

提取weekday等，hour minute

train_data['tm'] = pd.to_datetime(train_data['common_ts'], unit='ms')
test_data['tm'] = pd.to_datetime(test_data['common_ts'], unit='ms')
train_data['weekday'] = train_data['tm'].dt.dayofweek
test_data['weekday'] = test_data['tm'].dt.dayofweek