特征工程
特征工程的主要工作就是基于数据分析与探索提取潜在有价值的特征。
特征工程对于模型的构建起着至关重要的作用。
特征工程的第一步是特征提取，第二步则是对提取出来的特征进行组合。
特征提取
- 基于数据分析与探索提取
- 箱型图分析
交易量与周几有关 ⇒ \Rightarrow ⇒ 得到 6 个 0-1 型特征（是否周一，是否周二，…，是否周末）
- 点线图分析
user_start_level（用户星级）、user_occupation_id（用户职业）均与 is_trade（是否
交易）有关 ⇒ \Rightarrow ⇒ 虽然 user_start_level 连续，但是可以考虑将其离散化
- 离散型特征的重要性：
- 可用于设计规则
- 易于模型拟合，xgboost、lightgbm、catboost 等都以决策树为基模型
- 便于理解
- 便于做特征组合
- 在推荐系统等领域很常见
特征选择
- 劣态：剔除几乎无关的特征，保留大量特征
- 优胜：挑选出良好特征，组成最优特征子集
- 特征重要性的分析方法：
- Mean Variance Test
- SHAP (SHapley Additive exPlanations)
- Python 的 shap 库
- 可以用来解释任意机器学习模型的输出
- SHAP Value 为正，表明变量对预测值有促进作用
- SHAP Value 为负，表明变量对预测值有抑制作用
- 绝对值越大 ⇒ \Rightarrow ⇒ 对预测影响越大
- Permutation Importance
- Python 的 eli5 库
- 一个特征被处理为随机数后，若模型效果下降明显，则认为该特征重要
特征组合
- 简单的特征组合方式：
- 直接进行进行加、减、乘、除、 l o g log log、 e x p exp exp 等运算
- 易于生成大量特征，但会易出现过拟合问题，且不易于解释
- 更好的特征提取与组合方式：
- 在充分理解问题背景的基础上进行数据分析与探索
- 以时间序列问题为例，常见的特征包括统计量（最大值、最小值、中位数、偏度、峰度等）、排序（各统计量在历史同期的排名）、分位数（各统计量在历史同期排名的分位数）等等 。周期因子也可以视为一些特征的组合
- 以推荐系统为例，常见的特征类别包括用户特征、商品特征、行为特征（按时段统计）等等。可对其中的离散型特征直接做组合
特征工程的实现
下面的分析过程只对赎回总额 (total_redeem_amt) 进行，申购总额特征工程的过程与之类似。
- 首先进行特征提取，并对赎回总额在不同特征上的分布情况进行可视化，根据可视化的结果可以直观地剔除一些无法有效分割数据集的特征以及方差较大的特征
- 对赎回总额与其他特征的相关性进行分析，剔除与赎回总额相关性较差的那些特征
- 利用 MVTest 从剔除的相关性较差的集合中挽回与赎回总额有依赖性的特征
- 利用 SHAP 和 Permutation Importance 选择优胜特征
- 汇总结果得到最终的特征集合
初始化
导入需要用到的包：
## 导入库函数
import pandas as pd
import numpy as np
import datetime
import shap
import eli5
import seaborn as sns
import matplotlib.pyplot as plt
from mvtest import *
from wordcloud import WordCloud
from scipy import stats
from eli5.sklearn import PermutationImportance
from sklearn import tree
from sklearn.preprocessing import OneHotEncoder
from sklearn.linear_model import LinearRegression
from typing import *
import warnings
warnings.filterwarnings('ignore')
%matplotlib inline
为了方面后续操作，设置全局index变量：
## 为了方面后续操作，设置全局index变量
labels = ['total_purchase_amt','total_redeem_amt']
date_indexs = ['week','year','month','weekday','day']
数据预处理
定义处理数据要用到的函数：
## 读取 user_balance_table (用户申购赎回数据表)
def load_data(path: str = 'user_balance_table.csv')->pd.DataFrame:
data_balance = pd.read_csv(path)
return data_balance.reset_index(drop=True)
## 添加时间戳
def add_timestamp(data: pd.DataFrame, time_index: str = 'report_date')->pd.DataFrame:
data_balance = data.copy()
data_balance['date'] = pd.to_datetime(data_balance[time_index], format= "%Y%m%d")
data_balance['day'] = data_balance['date'].dt.day
data_balance['month'] = data_balance['date'].dt.month
data_balance['year'] = data_balance['date'].dt.year
data_balance['week'] = data_balance['date'].dt.week
data_balance['weekday'] = data_balance['date'].dt.weekday
return data_balance.reset_index(drop=True)
## 计算每天的申购、赎回总额
def get_total_balance(data: pd.DataFrame, date: str = '2014-03-31')->pd.DataFrame:
df_tmp = data.copy()
df_tmp = df_tmp.groupby(['date'])['total_purchase_amt','total_redeem_amt'].sum()
df_tmp.reset_index(inplace=True)
return df_tmp[(df_tmp['date']>= date)].reset_index(drop=True)
## 生成测试数据
def generate_test_data(data: pd.DataFrame)->pd.DataFrame:
total_balance = data.copy()
start = datetime.datetime(2014,9,1)
testdata = []
while start != datetime.datetime(2014,10,15):
temp = [start, np.nan, np.nan] # 用 nan 初始化待预测的数据
testdata.append(temp)
start += datetime.timedelta(days = 1)
testdata = pd.DataFrame(testdata)
testdata.columns = total_balance.columns
# 把生成的测试数据和原始total_balance拼接在一起
total_balance = pd.concat([total_balance, testdata], axis = 0)
total_balance = total_balance.reset_index(drop=True)
return total_balance.reset_index(drop=True)
## 读取 user_profile_table （用户信息表）
def load_user_information(path: str = 'user_profile_table.csv')->pd.DataFrame:
return pd.read_csv(path)
读取数据并进行处理：
## 读取数据并进行处理
balance_data = load_data('Dataset/user_balance_table.csv')
balance_data = add_timestamp(balance_data, time_index='report_date')
total_balance = get_total_balance(balance_data)
total_balance = generate_test_data(total_balance)
total_balance = add_timestamp(total_balance, 'date')
user_information = load_user_information('Dataset/user_profile_table.csv')
读取结果如下：
balance_data
total_balance
user_information
特征提取
基于日期的特征
对是否月初月中月末、是否周末、是否节假日等日期特征进行提取：
## 获取节假日的日期集合
def get_holiday_set()->Set[datetime.date]:
holiday_set = set()
# 清明节
holiday_set = holiday_set | {datetime.date(2014,4,5), datetime.date(2014,4,6), datetime.date(2014,4,7)}
# 劳动节
holiday_set = holiday_set | {datetime.date(2014,5,1), datetime.date(2014,5,2), datetime.date(2014,5,3)}
# 端午节
holiday_set = holiday_set | {datetime.date(2014,5,31), datetime.date(2014,6,1), datetime.date(2014,6,2)}
# 中秋节
holiday_set = holiday_set | {datetime.date(2014,9,6), datetime.date(2014,9,7), datetime.date(2014,9,8)}
# 国庆节
holiday_set = holiday_set | {datetime.date(2014,10,1), datetime.date(2014,10,2), datetime.date(2014,10,3),\
datetime.date(2014,10,4), datetime.date(2014,10,5), datetime.date(2014,10,6),\
datetime.date(2014,10,7)}
# 中秋节
holiday_set = holiday_set | {datetime.date(2013,9,19), datetime.date(2013,9,20), datetime.date(2013,9,21)}
# 国庆节
holiday_set = holiday_set | {datetime.date(2013,10,1), datetime.date(2013,10,2), datetime.date(2013,10,3),\
datetime.date(2013,10,4), datetime.date(2013,10,5), datetime.date(2013,10,6),\
datetime.date(2013,10,7)}
return holiday_set
提取所有是/否类型的特征：
## 提取所有是/否类型的特征
def extract_is_feature(data: pd.DataFrame)->pd.DataFrame:
total_balance = data.copy().reset_index(drop=True)
# 是否是Weekend
total_balance['is_weekend'] = 0
total_balance.loc[total_balance['weekday'].isin((5,6)), 'is_weekend'] = 1
# 是否是假期
total_balance['is_holiday'] = 0
total_balance.loc[total_balance['date'].isin(get_holiday_set()), 'is_holiday'] = 1
# 是否是节假日的第一天
last_day_flag = 0 # 前一天是否为 holiday
total_balance['is_firstday_of_holiday'] = 0
# iterrows() 是在数据框中的行进行迭代的一个生成器，它返回每行的索引及一个包含行本身的对象
# 可以使用 iterrows()方法实现行数据的遍历
for index, row in total_balance.iterrows():
if last_day_flag == 0 and row['is_holiday'] == 1: # 如果当天为 holiday 但前一天不是 holiday
total_balance.loc[index, 'is_firstday_of_holiday'] = 1
last_day_flag = row['is_holiday']
# 是否是节假日的最后一天
total_balance['is_lastday_of_holiday'] = 0
for index, row in total_balance.iterrows():
if row['is_holiday'] == 1 and total_balance.loc[index+1, 'is_holiday'] == 0:
total_balance.loc[index, 'is_lastday_of_holiday'] = 1
# 是否是节假日后的上班第一天
total_balance['is_firstday_of_work'] = 0
last_day_flag = 0
for index, row in total_balance.iterrows():
if last_day_flag == 1 and row['is_holiday'] == 0:
total_balance.loc[index, 'is_firstday_of_work'] = 1
last_day_flag = row['is_lastday_of_holiday']
# 是否要上班 （除了节假日和周末之外都要上班）
total_balance['is_work'] = 1
total_balance.loc[(total_balance['is_holiday'] == 1) | (total_balance['is_weekend'] == 1), 'is_work'] = 0
special_work_day_set = {datetime.date(2014,5,4), datetime.date(2014,9,28)} # 调休上班
total_balance.loc[total_balance['date'].isin(special_work_day_set), 'is_work'] = 1
# 是否明天要上班
total_balance['is_gonna_work_tomorrow'] = 0
for index, row in total_balance.iterrows():
if index == len(total_balance)-1:
break
if row['is_work'] == 0 and total_balance.loc[index+1, 'is_work'] == 1: # 今天不上班，明天要上班
total_balance.loc[index, 'is_gonna_work_tomorrow'] = 1
# 昨天上班了吗
total_balance['is_worked_yestday'] = 0
for index, row in total_balance.iterrows():
if index <= 1:
continue
if total_balance.loc[index-1, 'is_work'] == 1:
total_balance.loc[index, 'is_worked_yestday'] = 1
# 是否是放假前一天
total_balance['is_lastday_of_workday'] = 0
for index, row in total_balance.iterrows():
if index == len(total_balance)-1:
break
if row['is_holiday'] == 0 and total_balance.loc[index+1, 'is_holiday'] == 1:
total_balance.loc[index, 'is_lastday_of_workday'] = 1
# 是否周日要上班
total_balance['is_work_on_sunday'] = 0
for index, row in total_balance.iterrows():
if index == len(total_balance)-1:
break
if row['weekday'] == 6 and row['is_work'] == 1:
total_balance.loc[index, 'is_work_on_sunday'] = 1
# 是否是月初第一天
total_balance['is_firstday_of_month'] = 0
total_balance.loc[total_balance['day'] == 1, 'is_firstday_of_month'] = 1
# 是否是月初第二天
total_balance['is_secday_of_month'] = 0
total_balance.loc[total_balance['day'] == 2, 'is_secday_of_month'] = 1
# 是否是月初
total_balance['is_premonth'] = 0
total_balance.loc[total_balance['day'] <= 10, 'is_premonth'] = 1
# 是否是月中
total_balance['is_midmonth'] = 0
total_balance.loc[(10 < total_balance['day']) & (total_balance['day'] <= 20), 'is_midmonth'] = 1
# 是否是月末
total_balance['is_tailmonth'] = 0
total_balance.loc[20 < total_balance['day'], 'is_tailmonth'] = 1
# 是否是每个月第一个周
total_balance['is_first_week'] = 0
total_balance.loc[total_balance['week'] % 4 == 1, 'is_first_week'] = 1
# 是否是每个月第一个周
total_balance['is_second_week'] = 0
total_balance.loc[total_balance['week'] % 4 == 2, 'is_second_week'] = 1
# 是否是每个月第一个周
total_balance['is_third_week'] = 0
total_balance.loc[total_balance['week'] % 4 == 3, 'is_third_week'] = 1
# 是否是每个月第四个周
total_balance['is_fourth_week'] = 0
total_balance.loc[total_balance['week'] % 4 == 0, 'is_fourth_week'] = 1
return total_balance.reset_index(drop=True)
## 提取所有是/否类型的特征到数据集
total_balance = extract_is_feature(total_balance)
对 weekday 特征进行 OneHot 编码：
## 对weekday特征进行OneHot编码
def encode_data(data: pd.DataFrame, feature_name:str = 'weekday', encoder=OneHotEncoder())->pd.DataFrame():
total_balance = data.copy()
week_feature = encoder.fit_transform(np.array(total_balance[feature_name]).reshape(-1, 1)).toarray()
week_feature = pd.DataFrame(week_feature,columns= [feature_name + '_onehot_'+ str(x) for x in range(len(week_feature[0]))])
featureWeekday = pd.concat([total_balance, week_feature], axis = 1)
return featureWeekday
## 编码weekday特征到数据集
total_balance = encode_data(total_balance)
total_balance.head()
生成是/否类型的特征集合：
## 生成是/否类型的特征集合
feature = total_balance[[x for x in total_balance.columns if x not in date_indexs]] # date_indexs = ['week','year','month','weekday','day']
feature.head()
进行标签分布分析:
绘制箱型图：
## 绘制箱型图
def draw_boxplot(data: pd.DataFrame)->None:
f, axes = plt.subplots(7, 4, figsize=(18, 24))
global date_indexs, labels
count = 0
for i in [x for x in data.columns if x not in date_indexs + labels + ['date']]:
sns.boxenplot(x=i, y='total_redeem_amt', data=data, ax=axes[count // 4][count % 4])
count += 1
## 画出 total_redeem_amt 关于各个特征的箱线图
draw_boxplot(feature)
根据箱型图剔除部分看起来较差的特征：
## 剔除看起来较差的特征
redeem_feature_seems_useless = [
#样本量太少，建模时无效；但若确定这是一个有用规则，可以对结果做修正
'is_work_on_sunday'
#中位数差异不明显
'is_premonth'
]
相关性分析：
画出相关性热力图：
## 画相关性热力图
def draw_correlation_heatmap(data: pd.DataFrame, way:str = 'pearson')->None:
feature = data.copy()
plt.figure(figsize=(20,10))
plt.title('The ' + way +' coleration between total redeem and each feature')
sns.heatmap(feature[[x for x in feature.columns if x not in ['total_purchase_amt', 'date'] ]].corr(way),linecolor='white',
linewidths=0.1,
cmap="RdBu")
draw_correlation_heatmap(feature, 'spearman')
剔除与 total_redeem_amt 相关性较低的特征：
## 剔除相关性较低的特征
temp = np.abs(feature[[x for x in feature.columns
if x not in ['total_purchase_amt', 'date'] ]].corr('spearman')['total_redeem_amt'])
feature_low_correlation = list(set(temp[temp < 0.1].index))
print('其他特征与 total_redeem_amt 的相关系数')
print(temp)
print('与 total_redeem_amt 相关性较低的几个特征')
print(feature_low_correlation)
基于距离的特征
提取距离特征：
## 提取距离特征
def extract_distance_feature(data: pd.DataFrame)->pd.DataFrame:
total_balance = data.copy()
# 距离下一个假期还有多少天
total_balance['dis_to_nowork'] = 0
for index, row in total_balance.iterrows():
if row['is_work'] == 0: # 遍历到假期时，给假期前面的数据的 dis_to_nowork 赋值
step = 1
flag = 1
while flag:
if index - step >= 0 and total_balance.loc[index - step, 'is_work'] == 1:
# 如果 index - step > 0，并且那天工作则给 dis_to_nowork 赋值
total_balance.loc[index - step, 'dis_to_nowork'] = step
step += 1
else:
flag = 0
# 距上一个假期已经过去了多少天，即已经工作了多少天
total_balance['dis_from_nowork'] = 0
step = 0
for index, row in total_balance.iterrows():
step += 1
if row['is_work'] == 1:
total_balance.loc[index, 'dis_from_nowork'] = step
else:
step = 0
# 距离上班还有多少天
total_balance['dis_to_work'] = 0
for index, row in total_balance.iterrows():
if row['is_work'] == 1:
step = 1
flag = 1
while flag:
if index - step >= 0 and total_balance.loc[index - step, 'is_work'] == 0:
total_balance.loc[index - step, 'dis_to_work'] = step
step += 1
else:
flag = 0
# 已经休假了几天
total_balance['dis_from_work'] = 0
step = 0
for index, row in total_balance.iterrows():
step += 1
if row['is_work'] == 0:
total_balance.loc[index, 'dis_from_work'] = step
else:
step = 0
# 距离下一个节假日还有多少天
total_balance['dis_to_holiday'] = 0
for index, row in total_balance.iterrows():
if row['is_holiday'] == 1:
step = 1
flag = 1
while flag:
if index - step >= 0 and total_balance.loc[index - step, 'is_holiday'] == 0:
total_balance.loc[index - step, 'dis_to_holiday'] = step
step += 1
else:
flag = 0
# 距离上一个节假日已经过去了多少天
total_balance['dis_from_holiday'] = 0
step = 0
for index, row in total_balance.iterrows():
step += 1
if row['is_holiday'] == 0:
total_balance.loc[index, 'dis_from_holiday'] = step
else:
step = 0
# 距离下一个节假日的最后一天还有多少天
total_balance['dis_to_holiendday'] = 0
for index, row in total_balance.iterrows():
if row['is_lastday_of_holiday'] == 1:
step = 1
flag = 1
while flag:
if index - step >= 0 and total_balance.loc[index - step, 'is_lastday_of_holiday'] == 0:
total_balance.loc[index - step, 'dis_to_holiendday'] = step
step += 1
else:
flag = 0
# 距离上一个节假日的最后一天已经过去了多少天
total_balance['dis_from_holiendday'] = 0
step = 0
for index, row in total_balance.iterrows():
step += 1
if row['is_lastday_of_holiday'] == 0:
total_balance.loc[index, 'dis_from_holiendday'] = step
else:
step = 0
# 距离月初第几天
total_balance['dis_from_startofmonth'] = np.abs(total_balance['day'])
# 距离月的中心点有几天
total_balance['dis_from_middleofmonth'] = np.abs(total_balance['day'] - 15)
# 距离星期的中心有几天
total_balance['dis_from_middleofweek'] = np.abs(total_balance['weekday'] - 3)
# 距离星期日有几天
total_balance['dis_from_endofweek'] = np.abs(total_balance['weekday'] - 6)
return total_balance
## 拼接距离特征到原数据集
total_balance = extract_distance_feature(total_balance)
total_balance.head()
距离特征分析：
## 获取距离特征的列名
feature = total_balance[[x for x in total_balance.columns if x not in date_indexs]]
dis_feature_indexs = [x for x in feature.columns if (x not in date_indexs + labels + ['date']) & ('dis' in x)]
dis_feature_indexs
画出距离特征关于total_redeem_amt的点线图：
## 画出距离特征关于total_redeem_amt的点线图
def draw_point_feature(data: pd.DataFrame)->None:
feature = data.copy()
f, axes = plt.subplots(data.shape[1] // 3, 3, figsize=(30, data.shape[1] // 3 * 4))
count = 0
for i in [x for x in feature.columns if (x not in date_indexs + labels + ['date'])]:
# date_indexs + labels + ['date'] = ['week', 'year', 'month', 'weekday', 'day', 'total_purchase_amt', 'total_redeem_amt', 'date']
sns.pointplot(x=i, y="total_redeem_amt",
markers=["^", "o"], linestyles=["-", "--"],
kind="point", data=feature, ax=axes[count // 3][count % 3] if data.shape[1] > 3 else axes[count])
count += 1
draw_point_feature(feature[['total_redeem_amt'] + dis_feature_indexs])
处理距离过远的时间点：
## 处理距离过远的时间点
def dis_change(x):
if x > 5:
x = 10
return x
## 处理特殊距离
dis_holiday_feature = [x for x in total_balance.columns if 'dis' in x and 'holi' in x]
# ['dis_to_holiday', 'dis_from_holiday', 'dis_to_holiendday', 'dis_from_holiendday']
dis_month_feature = [x for x in total_balance.columns if 'dis' in x and 'month' in x]
# ['dis_from_startofmonth', 'dis_from_middleofmonth']
total_balance[dis_holiday_feature] = total_balance[dis_holiday_feature].applymap(dis_change)
total_balance[dis_month_feature] = total_balance[dis_month_feature].applymap(dis_change)
画出处理后的点线图：
## 处理后的点线图
feature = total_balance[[x for x in total_balance.columns if x not in date_indexs]]
draw_point_feature(feature[['total_redeem_amt'] + dis_feature_indexs])
根据点线图剔除方差较大的特征：
## 剔除看起来用处不大的特征
redeem_feature_seems_useless += [
#即使做了处理，但方差太大，不可信，规律不明显
'dis_to_holiday',
#方差太大，不可信
'dis_from_startofmonth',
#方差太大，不可信
'dis_from_middleofmonth'
]
相关性分析：
画出相关性热力图 ：
## 画出相关性热力图
draw_correlation_heatmap(feature[['total_redeem_amt'] + dis_feature_indexs])
剔除与 total_redeem_amt 相关性较差的特征
# 剔除相关性较差的特征
temp = np.abs(feature[[x for x in feature.columns
if ('dis' in x) | (x in ['total_redeem_amt']) ]].corr()['total_redeem_amt'])
feature_low_correlation += list(set(temp[temp < 0.1].index) )
feature_low_correlation
波峰波谷特征
提取波峰特征：
画出时序图观察波峰波谷的分布：
## 观察波峰特点
fig = plt.figure(figsize=(15,15))
for i in range(4,9):
plt.subplot(5,1,i - 3)
total_balance_2 = total_balance[(total_balance['date'] >= '2014-'+str(i)+'-01') & (total_balance['date'] < '2014-'+str(i+1)+'-01')]
sns.pointplot(x=total_balance_2['day'],y=total_balance_2['total_redeem_amt'])
plt.legend().set_title('Month:' + str(i))
通过观察图像可以看出来，波峰大都在周一，波谷大都在周六。
## 设定波峰波谷日期
def extract_peak_feature(data: pd.DataFrame)->pd.DataFrame:
total_balance = data.copy()
# 距离purchase波峰（即周一）有几天
total_balance['dis_from_redeem_peak'] = np.abs(total_balance['weekday'] - 1)
# 距离purchase波谷（即周六）有几天，与dis_from_endofweek相同
total_balance['dis_from_redeem_valley'] = np.abs(total_balance['weekday'] - 6)
return total_balance
## 提取波峰特征
total_balance = extract_peak_feature(total_balance)
feature = total_balance[[x for x in total_balance.columns if x not in date_indexs]]
feature.head()
画出波峰波谷特征的点线图：
draw_point_feature(feature[['total_redeem_amt'] + ['dis_from_redeem_peak','dis_from_redeem_valley']])
相关性分析：
temp = np.abs(feature[[x for x in feature.columns if ('peak' in x) or ('valley' in x) or (x in ['total_redeem_amt']) ]].corr()['total_redeem_amt'])
print('total_redeem_amt 与波峰波谷的相关性')
print(temp)
draw_correlation_heatmap(feature[['total_redeem_amt'] + ['dis_from_redeem_peak','dis_from_redeem_valley']])
可以看出来 total_redeem_amt 与波峰波谷特征都有一定的相关性。
周期因子特征
提取周期因子：
## 生成周期因子
def generate_rate(df, month_index):
total_balance = df.copy()
pure_balance = total_balance[['date','total_purchase_amt','total_redeem_amt']]
pure_balance = pure_balance[(pure_balance['date'] >= '2014-03-01') & (pure_balance['date'] < '2014-'+str(month_index)+'-01')]
pure_balance['weekday'] = pure_balance['date'].dt.weekday
pure_balance['day'] = pure_balance['date'].dt.day
pure_balance['week'] = pure_balance['date'].dt.week
pure_balance['month'] = pure_balance['date'].dt.month
# 计算星期因子
weekday_rate = pure_balance[['weekday']+labels].groupby('weekday',as_index=False).mean() # 申购总额根据周几进行聚合，并取平均
for name in labels:
weekday_rate = weekday_rate.rename(columns={name: name+'_weekdaymean'})
weekday_rate['total_purchase_amt_weekdaymean'] /= np.mean(pure_balance['total_purchase_amt']) # 根据周几分组后的均值/整体均值
weekday_rate['total_redeem_amt_weekdaymean'] /= np.mean(pure_balance['total_redeem_amt'])
pure_balance = pd.merge(pure_balance, weekday_rate, on='weekday', how='left')
# 计算日期因子
# 依据频次对周期因子total_purchase进行加权，获得日期因子
# 日期因子 = 周期因子*（周一到周日在(1~31)号出现的次数/共有几个月）
weekday_count = pure_balance[['day','weekday','date']].groupby(['day','weekday'],as_index=False).count() # 根据(1~31)号进行聚合，计算频次
weekday_count = pd.merge(weekday_count, weekday_rate, on = 'weekday')
weekday_count['total_purchase_amt_weekdaymean'] *= weekday_count['date'] / (len(set(pure_balance['month'])) - 1)
weekday_count['total_redeem_amt_weekdaymean'] *= weekday_count['date'] / (len(set(pure_balance['month'])) - 1)
day_rate = weekday_count.drop(['weekday','date'],axis=1).groupby('day',as_index=False).sum()
weekday_rate.columns = ['weekday','purchase_weekdayrate','redeem_weekdayrate']
day_rate.columns = ['day','purchase_dayrate','redeem_dayrate']
day_rate['date'] = datetime.datetime(2014, month_index, 1)
for index, row in day_rate.iterrows():
if month_index in (2,4,6,9) and row['day'] == 31:
continue
day_rate.loc[index, 'date'] = datetime.datetime(2014, month_index, int(row['day']))
day_rate['weekday'] = day_rate['date'].dt.weekday
day_rate = pd.merge(day_rate, weekday_rate, on='weekday')
day_rate['purchase_dayrate'] = day_rate['purchase_weekdayrate'] / day_rate['purchase_dayrate']
day_rate['redeem_dayrate'] = day_rate['redeem_weekdayrate'] / day_rate['redeem_dayrate']
weekday_rate['month'] = month_index
day_rate['month'] = month_index
return weekday_rate, day_rate[['day','purchase_dayrate','redeem_dayrate','month']].sort_values('day')
## 生成周期因子并合并到数据集
weekday_rate_list = []
day_rate_list = []
for i in range(3, 10):
weekday_rate, day_rate = generate_rate(total_balance, i)
weekday_rate_list.append(weekday_rate.reset_index(drop=True))
day_rate_list.append(day_rate.reset_index(drop=True))
weekday_rate_list = pd.concat(weekday_rate_list).reset_index(drop=True)
day_rate_list = pd.concat(day_rate_list).reset_index(drop=True)
total_balance = pd.merge(total_balance, weekday_rate_list, on=['weekday','month'], how='left')
total_balance = pd.merge(total_balance, day_rate_list, on=['day','month'], how='left')
total_balance.head()
## 对周期因子进行处理
for i in [x for x in total_balance.columns
if 'rate' in x and x not in labels + date_indexs]:
total_balance[i] = total_balance[i].fillna(np.nanmedian(total_balance[i])) # 填充空值
total_balance.head()
相关性分析：
画出相关性热力图：
## 画出相关性热力图
draw_correlation_heatmap(total_balance[['total_redeem_amt']
+ [x for x in total_balance.columns
if 'rate' in x and x not in labels + date_indexs]])
剔除与 total_redeem_amt 相关性低的特征：
## 剔除相关性低的特征
feature = total_balance.drop(date_indexs, axis=1)
temp = np.abs(feature[[x for x in feature.columns ]].corr()['total_redeem_amt'])
print('total_redeem_amt 和周期因子的相关性')
temp
动态时序特征
提取动态时序特征：
以星期为周期，统计申购总额的均值、中位数、最大值、最小值、偏度等：
## 提取动态特征
# 以星期为周期，统计申购总额的均值、中位数、最大值、最小值、偏度等
def get_amtfeature_with_time(data: pd.DataFrame)->pd.DataFrame:
df_tmp_ = data[labels + date_indexs + ['date']].copy()
total_balance = data.copy()
# 添加时间戳
df_tmp_ = df_tmp_[(df_tmp_['date']>= '2014-03-03')]
df_tmp_['weekday'] = df_tmp_['date'].dt.weekday + 1
df_tmp_['week'] = df_tmp_['date'].dt.week - min(df_tmp_['date'].dt.week) + 1 # 距开始日期过去了几周
df_tmp_['day'] = df_tmp_['date'].dt.day
df_tmp_['month'] = df_tmp_['date'].dt.month
df_tmp_.reset_index(inplace=True)
del df_tmp_['index']
df_redeem = pd.DataFrame(columns = ['weekday1','weekday2','weekday3','weekday4',
'weekday5','weekday6','weekday7']) # 初始化一个空数据框
count = 0
for i in range(len(df_tmp_)): # 把周一到周日变成列，第几周变成行
df_redeem.loc[count,'weekday'+str(df_tmp_.loc[i,'weekday'])] = df_tmp_.loc[i,'total_redeem_amt']
if df_tmp_.loc[i,'weekday'] == 7:
count = count + 1
df_tmp_['redeem_weekday_median'] = np.nan
df_tmp_['redeem_weekday_mean'] = np.nan
df_tmp_['redeem_weekday_min'] = np.nan
df_tmp_['redeem_weekday_max'] = np.nan
df_tmp_['redeem_weekday_std'] = np.nan
df_tmp_['redeem_weekday_skew'] = np.nan
for i in range(len(df_tmp_)):
# 从2014年3月31日开始统计
# 因为df_tmp_是从2014-03-03开始的，3+4*7-1就是日期大于30号的数据
if i > 4*7-1:
# 取当前周数-2条数据进行统计分析
df_tmp_.loc[i,'redeem_weekday_median'] = df_redeem.loc[:df_tmp_.loc[i,'week']-2,
'weekday'+str(df_tmp_.loc[i,'weekday'])].median()
df_tmp_.loc[i,'redeem_weekday_mean'] = df_redeem.loc[:df_tmp_.loc[i,'week']-2,
'weekday'+str(df_tmp_.loc[i,'weekday'])].mean()
df_tmp_.loc[i,'redeem_weekday_min'] = df_redeem.loc[:df_tmp_.loc[i,'week']-2,
'weekday'+str(df_tmp_.loc[i,'weekday'])].min()
df_tmp_.loc[i,'redeem_weekday_max'] = df_redeem.loc[:df_tmp_.loc[i,'week']-2,
'weekday'+str(df_tmp_.loc[i,'weekday'])].max()
df_tmp_.loc[i,'redeem_weekday_std'] = df_redeem.loc[:df_tmp_.loc[i,'week']-2,
'weekday'+str(df_tmp_.loc[i,'weekday'])].std()
df_tmp_.loc[i,'redeem_weekday_skew'] = df_redeem.loc[:df_tmp_.loc[i,'week']-2,
'weekday'+str(df_tmp_.loc[i,'weekday'])].skew() # 偏度
colList = ['redeem_weekday_median','redeem_weekday_mean','redeem_weekday_min',
'redeem_weekday_max','redeem_weekday_std','redeem_weekday_skew']
total_balance = pd.merge(total_balance, df_tmp_[colList+['day','month']], on=['day','month'], how='left')
return total_balance
## 合并特征到数据集
total_balance = get_amtfeature_with_time(total_balance)
total_balance.head()
相关性分析：
画出相关性热力图：
## 绘制动态特征的相关性图
draw_correlation_heatmap(total_balance[['total_redeem_amt'] +
['redeem_weekday_median',
'redeem_weekday_mean','redeem_weekday_min',
'redeem_weekday_max','redeem_weekday_std',
'redeem_weekday_skew'
]])
把提取出来的特征储存为 csv：
feature[labels + ['dis_to_nowork', 'dis_to_work', 'dis_from_work', 'purchase_weekdayrate',
'redeem_dayrate', 'weekday_onehot_5', 'weekday_onehot_6',
'dis_from_nowork', 'is_holiday', 'weekday_onehot_1', 'weekday_onehot_2',
'weekday_onehot_0', 'dis_from_middleofweek', 'dis_from_holiendday',
'weekday_onehot_3', 'is_lastday_of_holiday', 'is_firstday_of_holiday',
'weekday_onehot_4', 'is_worked_yestday', 'is_second_week',
'is_third_week', 'dis_from_startofmonth', 'dis_from_holiday', 'total_purchase_amt',
'total_redeem_amt', 'date']].to_csv('Feature/0615_residual_redeem_origined.csv', index=False)
特征剔除（劣汰）
剔除无法有效分割数据集的特征
画出各个特征分割数据集的分布估计图：
## 画出各个特征分割数据集的分布估计图
plt.figure(figsize=(4 * 6, 6 * len(feature.columns) / 6))
count = 0
for i in [x for x in feature.columns
if (x not in labels + date_indexs + ['date'])
& ('amt' not in x) & ('dis' not in x) & ('rate' not in x)]:
count += 1
if feature[feature[i] == 0].empty:
continue
plt.subplot(len(feature.columns) / 4, 4, count)
ax = sns.kdeplot(feature[feature[i] == 0]['total_redeem_amt'], label= str(i) + ' == 0, redeem')
ax = sns.kdeplot(feature[feature[i] == 1]['total_redeem_amt'], label= str(i) + ' == 1, redeem')
剔除对数据集划分不明显的特征 (分布图两条曲线几乎重合的特征)：
## 剔除对数据集划分不明显的特征
redeem_feature_seems_useless += ['is_third_week','is_fourth_week']
redeem_feature_seems_useless
使用MVTest挽回一些有依赖性但是不相关的特征
## MVtest Ref: https://github.com/ChuanyuXue/MVTest
l = mvtest()
name_list = []
Tn_list = []
p_list = []
for i in [i for i in feature_low_correlation if 'is' in i or 'discret' in i]:
pair = l.test(feature['total_redeem_amt'], feature[i])
name_list.append(str(i))
Tn_list.append(pair['Tn'])
p_list.append(pair['p-value'][0])
temp = pd.DataFrame([name_list,Tn_list]).T.sort_values(1)
temp[1] = np.abs(temp[1])
feature_saved_from_mv_redeem = list(temp.sort_values(1, ascending=False)[temp[1] > 0.5984][0])
剔除复共线特征
遍历相关性矩阵，如果特征 i 与特征 k 的相关性大于阈值，则考虑剔除特征 k，若特征 i 与特征 i * 特征 k 的相关性小于阈值，则保留 i*k 作为新特征（相当于交叉项）。
feature = feature[[x for x in feature.columns
if (x not in feature_low_correlation + redeem_feature_seems_useless) or\
(x in feature_saved_from_mv_redeem )]]
redeem_cors = feature.corr()
redeem_cors['total_redeem_amt'] = np.abs(redeem_cors['total_redeem_amt'])
feature_lists = list(redeem_cors.sort_values(by='total_redeem_amt',ascending=False).index)[2:] # 降序排列
print(feature_lists)
feature_temp = feature.dropna()
## 这里要注意,保留的时候按照相关性降序排序，剔除按照相关性升序排序的顺序
thershold = 0.8
for i in range(len(feature_lists)): # 遍历相关性矩阵行
for k in range(len(feature_lists)-1, -1, -1): # 倒序遍历相关性矩阵的列
if i >= len(feature_lists) or k >= len(feature_lists) or i == k:
break
if np.abs(np.corrcoef(feature_temp[feature_lists[i]], feature_temp[feature_lists[k]])[0][1]) > thershold:
# 如果特征i与特征k的相关性大于阈值
higher_feature_temp = feature_temp[feature_lists[i]] * feature_temp[feature_lists[k]]
if np.abs(np.corrcoef(feature_temp[feature_lists[i]], higher_feature_temp)[0][1]) <= thershold:
# 如果特征i与特征i*特征k的相关性小于阈值，则保留i*k作为新特征（相当于交叉项）
name = str(feature_lists[i]) + '%%%%' + str(feature_lists[k])
feature_temp[name] = higher_feature_temp
feature[name] = feature[feature_lists[i]] * feature[feature_lists[k]]
feature_lists.append(name)
feature_temp = feature_temp.drop(feature_lists[k], axis=1)
# 删除特征k
feature_lists.remove(feature_lists[k])
feature = feature[[x for x in feature_lists if x not in labels] + labels + ['date']]
feature.to_csv('Feature/redeem_feature_droped_0614.csv',index=False)
feature.head()
特征选择 （优胜）
分割数据集，取一部分数据作为训练数据，另一部分作为测试数据：
## 分割数据集
def split_data_underline(data): # 分成4月到8月以及8月到9月
trainset = data[('2014-04-01' <= data['date']) & (data['date'] < '2014-08-01')]
testset = data[('2014-08-01' <= data['date']) & (data['date'] < '2014-09-01')]
return trainset, testset
使用 SHAP 获取优胜特征
shap.initjs() # 用来显示的模块
from sklearn import tree
model = tree.DecisionTreeRegressor()
train, test = split_data_underline(feature.dropna())
features = [x for x in train.columns if x not in date_indexs]
model.fit(train[features].drop(labels+['date'], axis=1), train['total_redeem_amt'])
explainer = shap.TreeExplainer(model)
shap_testues = explainer.shap_values(test[features].drop(labels+['date'], axis=1))
shap.summary_plot(shap_testues, test[features].drop(labels+['date'], axis=1), plot_type='bar')
shap.summary_plot(shap_testues, test[features].drop(labels+['date'], axis=1))
tree_important_redeem = pd.DataFrame(np.mean(np.abs(shap_testues), axis=0),[x for x in features if x not in labels + date_indexs + ['date']]).reset_index()
选择前20个重要性较大的特征：
tree_important_redeem = tree_important_redeem.sort_values(0, ascending=False).reset_index(drop=True)
tree_important_redeem = list(tree_important_redeem[:20]['index']) # 选择前20个重要性较大的特征
使用词云图显示输出特征：
## 输出选择的特征
def draw_cloud(feature_index: List[str])->None:
plt.figure(figsize=(20,10))
plt.subplot(1,2,1)
ciyun = WordCloud(background_color='white', max_font_size=40)
ciyun.generate(text=''.join([x+' ' for x in feature_index if x != 'total_redeem_amt']))
plt.imshow(ciyun, interpolation='bilinear')
plt.axis("off")
draw_cloud(tree_important_redeem)
使用 Permutation importance 获取优胜特征
model = LinearRegression()
train, test = split_data_underline(feature.dropna())
model.fit(train[features].drop(labels+['date'], axis=1), train['total_redeem_amt'])
perm = PermutationImportance(model, random_state=42).fit(test[features].drop(labels+['date'], axis=1), test['total_redeem_amt'])
liner_important_redeem = pd.DataFrame(np.abs(perm.feature_importances_), [x for x in features if x not in labels + date_indexs + ['date']]).reset_index()
eli5.show_weights(perm, feature_names=list(str(x) for x in features if x not in labels + ['date']))
选择前20个重要性较大的特征：
liner_important_redeem = liner_important_redeem.sort_values(0, ascending=False).reset_index(drop=True)
liner_important_redeem = list(liner_important_redeem[:20]['index'])
liner_important_redeem
使用词云图显示输出特征：
draw_cloud(liner_important_redeem)
特征集合取交集选出最终优胜特征
winer_features_redeem = list(set(tree_important_redeem)\
& set(liner_important_redeem))
winer_features_redeem
画出词云图：
draw_cloud(winer_features_redeem)
272
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
