Adtributor python实现

import numpy as np
import pandas as pd
import scipy.stats

lists = [['联通', '北京', 20, 5],
         ['联通', '上海', 15, 15],
         ['联通', '广东', 10, 10],
         ['电信', '北京', 10, 7],
         ['电信', '上海', 25, 15],
         ['电信', '广东', 20, 20]]
df = pd.DataFrame(lists, columns=['isp', 'area', 'pv_pred', 'pv_actual'])


# 这里要分析当pv实际的数值跟预测的数值有差距的时候，到底是哪个维度的数据出了问题


def js_divergence(p, q):
    p = np.array(p)
    q = np.array(q)
    m = (p + q) / 2

    # 方法一：自定义函数
    js1 = 0.5 * np.sum(p * np.log(p / m)) + 0.5 * np.sum(q * np.log(q / m))
    # 方法二：调用scipy包
    js2 = 0.5 * scipy.stats.entropy(p, m) + 0.5 * scipy.stats.entropy(q, m)

    return round(float(js1), 6)

def root_cause_analysis(dframe):
    '''
    数据总共有两个维度，分别为isp和area，isp维度下共有两个值，分别为联通、电信，area维度下有三个值，分别为北京、上海、广东
    
    '''
    pv_sum = dframe.sum(numeric_only=True)

    # 先计算isp维度的
    group_isp = dframe.groupby('isp').sum()
    group_isp = group_isp.reset_index()
    group_isp['pred_sum'] = pv_sum['pv_pred']
    group_isp['actual_sum'] = pv_sum['pv_actual']

    group_isp['p'] = group_isp['pv_pred'] / group_isp['pred_sum']
    group_isp['q'] = group_isp['pv_actual'] / group_isp['actual_sum']

    # 再计算area维度下的
    group_area = dframe.groupby('area').sum()
    group_area = group_area.reset_index()
    group_area['pred_sum'] = pv_sum['pv_pred']
    group_area['actual_sum'] = pv_sum['pv_actual']

    group_area['p'] = group_area['pv_pred'] / group_area['pred_sum']
    group_area['q'] = group_area['pv_actual'] / group_area['actual_sum']

    # 第1步：分别计算两个维度下，不同维值的Surprise
    group_isp['surprise'] = group_isp[['p', 'q']].apply(lambda x: js_divergence(x['p'], x['q']), axis=1)
    group_area['surprise'] = group_area[['p', 'q']].apply(lambda x: js_divergence(x['p'], x['q']), axis=1)

    # 第2步：计算每个维度下每个维值的EP值
    group_isp['EP'] = group_isp[['pv_pred', 'pv_actual', 'pred_sum', 'actual_sum']].apply(
        lambda x: (x['pv_actual'] - x['pv_pred']) / (x['actual_sum'] - x['pred_sum']), axis=1)
    group_area['EP'] = group_area[['pv_pred', 'pv_actual', 'pred_sum', 'actual_sum']].apply(
        lambda x: (x['pv_actual'] - x['pv_pred']) / (x['actual_sum'] - x['pred_sum']), axis=1)

    # 第3步：根据Surprise值和EP值输出不同维度下维值的影响大小，从大到小
    isp_surprise = group_isp['surprise'].sum()
    area_surprise = group_area['surprise'].sum()

    return ('isp', group_isp[['isp', 'EP']].sort_values(by=['EP'], ascending=False)) \
        if isp_surprise > area_surprise \
        else ('area', group_area[['area', 'EP']].sort_values(by=['EP'], ascending=False))


dim, ep = root_cause_analysis(df)
print('root cause dim is %s' % dim)
print('dim value and ep:')
print(ep)

CK:
https://github.com/penchy-zju/code_gist/blob/master/adtributor.py