BayesCombine.py 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235
  1. import yaml
  2. import os
  3. import pandas as pd
  4. import numpy as np
  5. from commonFunc import update_dict, get_db_engine
  6. from functools import reduce
  7. from itertools import product
  8. from datetime import datetime
  9. from statsmodels.stats.proportion import proportions_ztest
  10. class BayesFeatures(object):
  11. """
  12. 计算素材单维度的特征值
  13. """
  14. def __init__(self, signature, target_type, dim, dim_config):
  15. self.signature = signature
  16. self.target_type = target_type
  17. self.dim = dim
  18. self.dim_config = dim_config
  19. self.table = self.dim_config['table']
  20. self.file_name = self.dim_config['fieldName']
  21. self.dim_lst = self.dim_config['Lst']
  22. self.window_size = self.dim_config['windowSize']
  23. self.window_combine = []
  24. self.bayes_feature = {}
  25. def get_window_combine(self):
  26. """
  27. 根据size得到指定维度的滑窗组合。
  28. 如 age: ['18-23岁', '24-30岁','31-40岁',....] 和 size = 2
  29. 得到 ['18-23岁', '24-30岁'],['24-30岁','31-40岁'],......
  30. """
  31. start = 0
  32. while (start < len(self.dim_lst)) and (start + self.window_size <= len(self.dim_lst)):
  33. self.window_combine.append(self.dim_lst[start: start + self.window_size])
  34. start += 1
  35. # 添加一个全部的组合,等同于'不限'
  36. self.window_combine.append(['不限'])
  37. def get_bayes_feature(self):
  38. # 1、获取指定素材在指定维度下的人群数据(近一个月内的数据表现)
  39. sql = '''
  40. select signature,
  41. %s,
  42. sum(aclick) aclick,
  43. sum(bclick) bclick,
  44. sum(activation) activation
  45. from %s
  46. where signature = '%s'
  47. -- and datediff(now(),stat_date)<=30
  48. group by signature, %s
  49. ''' % (self.file_name, self.table, self.signature, self.file_name)
  50. df = pd.read_sql(sql, engine)
  51. if self.target_type == 'action_ratio':
  52. # 计算行为率(bclick/aclick)的组合特征值
  53. df['unbclick'] = df['aclick'] - df['bclick']
  54. sig_pos_pct = df['bclick'].sum() / df['aclick'].sum() # 素材曝光-点击的概率
  55. sig_neg_pct = 1 - sig_pos_pct # 素材曝光-未点击的概率
  56. self.bayes_feature['sig_pos_pct'] = sig_pos_pct
  57. self.bayes_feature['sig_neg_pct'] = sig_neg_pct
  58. self.bayes_feature['sample_size'] = df['aclick'].sum() # 素材曝光量,后续计算显著差异用
  59. for sub_combine in self.window_combine:
  60. if sub_combine != ['不限']:
  61. key = self.dim + '_' + '|'.join(sub_combine)
  62. if self.dim != 'city':
  63. pos_pct = df[df[self.file_name].isin(sub_combine)].bclick.sum() / df.bclick.sum()
  64. neg_pct = df[df[self.file_name].isin(sub_combine)].unbclick.sum() / df.unbclick.sum()
  65. if self.dim == 'city':
  66. # 获取 city_level 对应的 city_name
  67. city_lst = []
  68. for city_level in sub_combine:
  69. get_city_sql = """
  70. select city_name from ctop_kuaishou_city_level where city_level = '%s'
  71. """ % city_level
  72. city_df = pd.read_sql(get_city_sql, engine)
  73. city_lst.extend(city_df['city_name'].values)
  74. pos_pct = df[df[self.file_name].isin(city_lst)].bclick.sum() / df.bclick.sum()
  75. neg_pct = df[df[self.file_name].isin(city_lst)].unbclick.sum() / df.unbclick.sum()
  76. self.bayes_feature[key] = {'pos_pct': pos_pct, 'neg_pct': neg_pct}
  77. else:
  78. self.bayes_feature[self.dim + '_' + '不限'] = {'pos_pct': 1, 'neg_pct': 1}
  79. if self.target_type == 'convert_ratio':
  80. # 计算转化率(activation/bclick)的组合特征值
  81. df['unactivation'] = df['bclick'] - df['activation']
  82. sig_pos_pct = df['activation'].sum() / df['bclick'].sum() # 素材点击-转化的概率
  83. sig_neg_pct = 1 - sig_pos_pct # 素材点击-未转化的概率
  84. self.bayes_feature['sig_pos_pct'] = sig_pos_pct
  85. self.bayes_feature['sig_neg_pct'] = sig_neg_pct
  86. self.bayes_feature['sample_size'] = df['bclick'].sum() # 素材行为量,后续计算显著差异用
  87. for sub_combine in self.window_combine:
  88. if sub_combine != '不限':
  89. key = self.dim + '_' + '|'.join(sub_combine)
  90. if self.dim != 'city':
  91. pos_pct = df[df[self.file_name].isin(sub_combine)].activation.sum() / df.activation.sum()
  92. neg_pct = df[df[self.file_name].isin(sub_combine)].unactivation.sum() / df.unactivation.sum()
  93. if self.dim == 'city':
  94. # 获取 city_level 对应的 city_name
  95. city_lst = []
  96. for city_level in sub_combine:
  97. get_city_sql = """
  98. select city_name from ctop_kuaishou_city_level where city_level = '%s'
  99. """ % city_level
  100. city_df = pd.read_sql(get_city_sql, engine)
  101. city_lst.extend(city_df['city_name'].values)
  102. pos_pct = df[df[self.file_name].isin(city_lst)].activation.sum() / df.activation.sum()
  103. neg_pct = df[df[self.file_name].isin(city_lst)].unactivation.sum() / df.unactivation.sum()
  104. self.bayes_feature[key] = {'pos_pct': pos_pct, 'neg_pct': neg_pct}
  105. else:
  106. self.bayes_feature[self.dim + '_' + '不限'] = {'pos_pct': 1, 'neg_pct': 1}
  107. class BayesCombine(object):
  108. """
  109. 依据特征值,计算多维度的组合预估值
  110. """
  111. def __init__(self, signature, project_id, target_type, dim_features):
  112. self.signature = signature
  113. self.project_id = project_id
  114. self.target_type = target_type
  115. self.dim_features = dim_features
  116. self.bayes_combine_df = pd.DataFrame()
  117. self.actual_prob = None
  118. self.sample_size = None
  119. def get_bayes_estimate(self):
  120. dim_combine_lst = [[key for key in fea.keys() if key not in ['sig_pos_pct', 'sig_neg_pct', 'sample_size']]
  121. for fea in self.dim_features]
  122. self.dim_features = reduce(update_dict, self.dim_features)
  123. self.actual_prob = self.dim_features['sig_pos_pct']
  124. self.sample_size = self.dim_features['sample_size']
  125. for ele in product(*dim_combine_lst):
  126. pos_pct_lst = [self.dim_features[key]['pos_pct'] for key in ele]
  127. prob_pos = reduce(lambda x, y: x * y, pos_pct_lst)
  128. neg_pct_lst = [self.dim_features[key]['neg_pct'] for key in ele]
  129. prob_neg = reduce(lambda x, y: x * y, neg_pct_lst)
  130. prob = (prob_pos * self.dim_features['sig_pos_pct']) / (prob_neg * self.dim_features['sig_neg_pct'])
  131. out_dict = dict(zip([e.split('_')[0] for e in ele], [e.split('_')[-1] for e in ele]))
  132. # 计算该组合的概率值 与 实际投放的概率值 是否存在显著性差异:显著好、 显著差
  133. count = np.array([self.sample_size * prob, self.sample_size * self.actual_prob])
  134. nobs = np.array([self.sample_size, self.sample_size])
  135. z_score, p_value = proportions_ztest(count=count, nobs=nobs, value=None, alternative='two-sided', prop_var=False)
  136. out_dict['z_score'] = z_score
  137. out_dict['p_value'] = p_value
  138. out_dict['sample_size'] = self.sample_size
  139. # TODO 调用人群预估覆盖接口,得到该组合的人群覆盖数
  140. out_dict['crowd_coverage_cnt'] = 1000
  141. out_dict['combine_estimate_prob'] = prob
  142. out_dict['actual_prob'] = self.actual_prob
  143. out_dict['signature'] = self.signature
  144. out_dict['project_id'] = self.project_id
  145. out_dict['target_type'] = self.target_type
  146. out_dict['stat_date'] = str(datetime.now().date())
  147. combine = pd.DataFrame([out_dict])
  148. self.bayes_combine_df = self.bayes_combine_df.append(combine, ignore_index=True)
  149. def write_to_db(self):
  150. self.bayes_combine_df.to_sql(name="ctop_ai_kuaishou_signature_recommended_target_combine",
  151. con=engine,
  152. if_exists='append',
  153. index=False)
  154. if __name__ == '__main__':
  155. # 1、读取配置文件
  156. with open('config/config.yaml', mode='r', encoding='utf-8') as f:
  157. config = yaml.load(f.read(), Loader=yaml.FullLoader)
  158. # 数据库连接引擎,依据开发环境/生产环境 进行切换
  159. if os.getenv('environment', 'unknown') == 'dev':
  160. engine = get_db_engine(config['devDB'])
  161. else:
  162. engine = get_db_engine(config['productDB'])
  163. # 2、 参与定向组合的维度
  164. target_dim = [key for key in config['bayesDim'].keys() if config['bayesDim'][key]['isOn']]
  165. # 4、计算贝叶斯组合入库
  166. for project_id in config['projectId']:
  167. # 4.1 获取指定项目下当前活跃的素材信息, 如近3天内累计激活个数达到50个(开发环境或者生产环境,这部分都读取生产数据库)
  168. product_engine = get_db_engine(config['productDB'])
  169. sql = '''
  170. select signature from ctop_kuaishou_report_daily_material
  171. where account_id in (select account_id from ctop_user_allocation where project_id = %s)
  172. and datediff(now(),stat_date) <= %s
  173. group by signature
  174. having sum(activation) >= %s
  175. ''' % (project_id,
  176. config['activeMaterialFilterRule']['days'],
  177. config['activeMaterialFilterRule']['activation'])
  178. df = pd.read_sql(sql, product_engine)
  179. active_signatures = df[~df.signature.isnull()].signature.values
  180. # 4.2 在素材人群报表筛选里面近一个月累计100个激活 TODO 等生产的素材人群表存在之后,修改表名,添加 days
  181. sql = """
  182. select signature from ctop_kuaishou_audience_daily_report_by_signature_age
  183. where signature in %s
  184. group by signature
  185. having sum(activation) >= %s
  186. """ % (tuple(active_signatures),
  187. config['getBaysCombineMaterialFilterRule']['activation'])
  188. df = pd.read_sql(sql, engine)
  189. signature_lst = df['signature'].values
  190. # TODO 这行代码用于测试
  191. signature_lst = ['0070efb7557b2a04cf3d4a6f243c3cd8', '03b93728f0d82ea7865c3c7cf632bc1b']
  192. # 4.2 计算指定素材的贝叶斯特征
  193. for sig in signature_lst:
  194. for t_type in config['targetType']:
  195. bayes_feature_lst = []
  196. for dimension in target_dim:
  197. cls = BayesFeatures(sig, t_type, dimension, config['bayesDim'][dimension])
  198. # 计算滑窗组合
  199. cls.get_window_combine()
  200. cls.get_bayes_feature()
  201. bayes_feature_lst.append(cls.bayes_feature)
  202. # 依据特征值,计算多维度的组合预估值
  203. cls = BayesCombine(sig, project_id, t_type, bayes_feature_lst)
  204. cls.get_bayes_estimate()
  205. cls.write_to_db()