psi计算

基础概念：https://zhuanlan.zhihu.com/p/344754828

import sys
import pandas as pd
import numpy as np
import math


# all_list = []
# df = pd.DataFrame(columns = ['date', 'data'])
# counter = 0
# for line in sys.stdin:
#     line = line.strip('\r\n')
#     df.loc[counter] = line.split(',')
#     counter += 1
# df.to_excel('./output2.xlsx', sheet_name='Sheet1')

def calc_psi(dataframe):
    date_col = dataframe['date']
    data_col = dataframe['data']
    # print(dataframe['data'].max(), dataframe['data'].min())
    
    # 分组
    result = pd.qcut(dataframe['data'], 10)
    # print(result)
    dataframe["group"] = result
    group = result.drop_duplicates(inplace=False).sort_values(ascending=True)
    group.index = range(1, len(group) + 1)
    print(group)
    print('group 1:', group[1])

    # 日期
    from datetime import datetime
    months = date_col.apply(lambda x: str(datetime.strptime(str(x), '%Y%m%d').date().year)+str(datetime.strptime(str(x), '%Y%m%d').date().month))
    months_drop = months.drop_duplicates(inplace=False)
    print(months_drop.values)

    dataframe['new_date'] = months
    print(dataframe)

    # 先取月份
    # print((months == months_drop.values[1]) and (dataframe['group'] == group[1]))
    # 取分组
    # print(dataframe['group'])
    # print('group 11:', group[1])
    ################ 9月数据在第一组的数量/9月总数 ###########
    # print(dataframe['group'] == (group[1]))
    # print((months == months_drop.values[1]))
    # print('ggg:', (months == months_drop.values[1]) & (dataframe['group'] == group[1]))
    actual_i = []
    psi_indexs = []
    for month_idx in range(0, 7):
        group_i = []
        for group_index in range(1, 11):
            dec_group_i = dataframe[(months == months_drop.values[month_idx]) & (dataframe['group'] == group[group_index])]
            # print(dec_group_i.size, dataframe[months == months_drop.values[month_idx]].size)
            group_i.append(dec_group_i.size / dataframe[months == months_drop.values[month_idx]].size)
        print('group_i:', group_i)
        # 计算index = (实际占比 - 预期占比)* ln(实际占比 / 预期占比) 
        actual_i.append(group_i)
        if month_idx > 0:
            actual_ = np.array(actual_i[month_idx])
            except_ = np.array(actual_i[month_idx - 1])
            psi_index = (actual_ - except_) * np.log(actual_ / except_)
            psi_indexs.append(psi_index)
            print('psi_index:', psi_index)

    print(actual_i)
    # print(psi_indexs) 
    final_index = [np.sum(psi_index) for psi_index in psi_indexs]
    print('final_index: ', final_index)   
    
    # data_202109 = dataframe[(months == months_drop.values[1]) & (dataframe['group'] == group[1])]
    # print(data_202109.size, dataframe[(months == months_drop.values[1])].size)
    # group2 = dataframe[(months == months_drop.values[1]) & (dataframe['group'] == group[2])]
    # print(group2.size)


# date_drop = data.drop_duplicates(subset=['date', 'date'], keep='first', inplace=False)
data = pd.read_csv('./lyx_data.txt', sep=',', names=['date', 'data'])
# print(data.head())
calc_psi(data)


# 生成数据
# import numpy as np
# date = ['20210812', '20210922', '20211009', '20211102', '20211202', '20220112', '20220202']
# new_date = []
# for i in range(100):
#     new_date.extend(date)
# new_data = []
# for i in range(700):
#     new_data.append(np.random.random(1))

# data = pd.DataFrame(columns=['date', 'data'])
# data['date'] = np.array(new_date)
# data['data'] = np.array(new_data)
# data.to_csv('./lyx_data.txt', sep=',', index=None)