BayesCombine.py 14 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285
  1. import json
  2. import yaml
  3. import os
  4. import pandas as pd
  5. import numpy as np
  6. from utils.commonFunc import update_dict, get_db_engine, city_code_transform, age_code_transform, gender_code_transform
  7. from functools import reduce
  8. from itertools import product
  9. from datetime import datetime
  10. from statsmodels.stats.proportion import proportions_ztest
  11. import traceback
  12. import requests
  13. from utils.BaseClass import NpEncoder
  14. from config.url import estimate_people_number_url, headers
  15. import logging
  16. from concurrent_log import ConcurrentTimedRotatingFileHandler
  17. log_formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s', '%Y/%m/%d %I:%M:%S %p')
  18. log_handler = ConcurrentTimedRotatingFileHandler("/data/pythonProject/ai_target/logs/BayesCombine.log", when="midnight", backupCount=100)
  19. log_handler.setFormatter(log_formatter)
  20. logger = logging.getLogger('bayes_combine_logger')
  21. logger.addHandler(log_handler)
  22. logger.setLevel(logging.DEBUG)
  23. print('id of bayes_combine_logger %s' % id(logger))
  24. class BayesFeatures(object):
  25. """
  26. 计算素材单维度的特征值
  27. """
  28. def __init__(self, signature, target_type, dim, dim_config):
  29. self.signature = signature
  30. self.target_type = target_type
  31. self.dim = dim
  32. self.dim_config = dim_config
  33. self.table = self.dim_config['table']
  34. self.file_name = self.dim_config['fieldName']
  35. self.dim_lst = self.dim_config['Lst']
  36. self.window_size = self.dim_config['windowSize']
  37. self.window_combine = []
  38. self.bayes_feature = {}
  39. def get_window_combine(self):
  40. """
  41. 根据size得到指定维度的滑窗组合。
  42. 如 age: ['18-23岁', '24-30岁','31-40岁',....] 和 size = 2
  43. 得到 ['18-23岁', '24-30岁'],['24-30岁','31-40岁'],......
  44. """
  45. start = 0
  46. while (start < len(self.dim_lst)) and (start + self.window_size <= len(self.dim_lst)):
  47. self.window_combine.append(self.dim_lst[start: start + self.window_size])
  48. start += 1
  49. # 添加一个全部的组合,等同于'不限'
  50. self.window_combine.append(['不限'])
  51. def get_bayes_feature(self):
  52. # 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)
  53. sql = '''
  54. select signature,
  55. %s,
  56. sum(aclick) aclick,
  57. sum(bclick) bclick,
  58. sum(activation) activation
  59. from %s
  60. where signature = '%s'
  61. and datediff(now(),stat_date)<= %s
  62. group by signature, %s
  63. ''' % (self.file_name, self.table, self.signature, config['getBaysCombineMaterialFilterRule']['days'], self.file_name)
  64. df = pd.read_sql(sql, product_engine)
  65. if self.target_type == 'action_ratio':
  66. # 计算行为率(bclick/aclick)的组合特征值
  67. df['unbclick'] = df['aclick'] - df['bclick']
  68. for sub_combine in self.window_combine:
  69. if sub_combine != ['不限']:
  70. key = self.dim + '_' + '|'.join(sub_combine)
  71. if self.dim != 'city':
  72. pos_pct = df[df[self.file_name].isin(sub_combine)].bclick.sum() / df.bclick.sum()
  73. neg_pct = df[df[self.file_name].isin(sub_combine)].unbclick.sum() / df.unbclick.sum()
  74. if self.dim == 'city':
  75. # 获取 city_level 对应的 city_name
  76. city_lst = []
  77. for city_level in sub_combine:
  78. get_city_sql = """
  79. select city_name from ctop_kuaishou_city_level where city_level = '%s'
  80. """ % city_level
  81. city_df = pd.read_sql(get_city_sql, product_engine)
  82. city_lst.extend(city_df['city_name'].values)
  83. pos_pct = df[df[self.file_name].isin(city_lst)].bclick.sum() / df.bclick.sum()
  84. neg_pct = df[df[self.file_name].isin(city_lst)].unbclick.sum() / df.unbclick.sum()
  85. self.bayes_feature[key] = {'pos_pct': pos_pct, 'neg_pct': neg_pct}
  86. else:
  87. self.bayes_feature[self.dim + '_' + '不限'] = {'pos_pct': 1, 'neg_pct': 1}
  88. if self.target_type == 'convert_ratio':
  89. # 计算转化率(activation/bclick)的组合特征值
  90. df['unactivation'] = df['bclick'] - df['activation']
  91. for sub_combine in self.window_combine:
  92. if sub_combine != '不限':
  93. key = self.dim + '_' + '|'.join(sub_combine)
  94. if self.dim != 'city':
  95. pos_pct = df[df[self.file_name].isin(sub_combine)].activation.sum() / df.activation.sum()
  96. neg_pct = df[df[self.file_name].isin(sub_combine)].unactivation.sum() / df.unactivation.sum()
  97. if self.dim == 'city':
  98. # 获取 city_level 对应的 city_name
  99. city_lst = []
  100. for city_level in sub_combine:
  101. get_city_sql = """
  102. select city_name from ctop_kuaishou_city_level where city_level = '%s'
  103. """ % city_level
  104. city_df = pd.read_sql(get_city_sql, product_engine)
  105. city_lst.extend(city_df['city_name'].values)
  106. pos_pct = df[df[self.file_name].isin(city_lst)].activation.sum() / df.activation.sum()
  107. neg_pct = df[df[self.file_name].isin(city_lst)].unactivation.sum() / df.unactivation.sum()
  108. self.bayes_feature[key] = {'pos_pct': pos_pct, 'neg_pct': neg_pct}
  109. else:
  110. self.bayes_feature[self.dim + '_' + '不限'] = {'pos_pct': 1, 'neg_pct': 1}
  111. class BayesCombine(object):
  112. """
  113. 依据特征值,计算多维度的组合预估值
  114. """
  115. def __init__(self, signature, project_id, target_type, dim_features):
  116. self.signature = signature
  117. self.project_id = project_id
  118. self.target_type = target_type
  119. self.dim_features = dim_features
  120. self.bayes_combine_df = pd.DataFrame()
  121. self.actual_prob = None
  122. self.sample_size = None
  123. self.sig_pos_pct = None
  124. self.sig_neg_pct = None
  125. def get_bayes_estimate(self):
  126. dim_combine_lst = [[key for key in fea.keys()] for fea in self.dim_features]
  127. self.dim_features = reduce(update_dict, self.dim_features)
  128. # 通过 年龄对应的表: ctop_kuaishou_audience_daily_report_by_signature_age 来获取样本量,和正负比例值(其他表的数据可能会丢失导致不全)
  129. get_sample_and_pct_sql = """
  130. select sum(aclick) aclick, sum(bclick) bclick, sum(activation) activation from %s where
  131. signature = '%s'
  132. and datediff(now(),stat_date)<= %s
  133. """ % (config['bayesDim']['age']['table'], self.signature, config['getBaysCombineMaterialFilterRule']['days'])
  134. sample_pct_df = pd.read_sql(get_sample_and_pct_sql, product_engine)
  135. if self.target_type == 'action_ratio':
  136. self.sig_pos_pct = sample_pct_df['bclick'].sum() / sample_pct_df['aclick'].sum()
  137. self.sig_neg_pct = 1 - self.sig_pos_pct
  138. self.actual_prob = self.sig_pos_pct
  139. self.sample_size = int(sample_pct_df.aclick.sum())
  140. elif self.target_type == 'convert_ratio':
  141. self.sig_pos_pct = sample_pct_df['activation'].sum() / sample_pct_df['bclick'].sum()
  142. self.sig_neg_pct = 1 - self.sig_pos_pct
  143. self.actual_prob = self.sig_pos_pct
  144. self.sample_size = int(sample_pct_df.bclick.sum())
  145. for ele in product(*dim_combine_lst):
  146. pos_pct_lst = [self.dim_features[key]['pos_pct'] for key in ele]
  147. prob_pos = reduce(lambda x, y: x * y, pos_pct_lst)
  148. neg_pct_lst = [self.dim_features[key]['neg_pct'] for key in ele]
  149. prob_neg = reduce(lambda x, y: x * y, neg_pct_lst)
  150. prob = (prob_pos * self.sig_pos_pct) / (prob_neg * self.sig_neg_pct)
  151. out_dict = dict(zip([e.split('_')[0] for e in ele], [e.split('_')[-1] for e in ele]))
  152. # 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
  153. count = np.array([self.sample_size * prob, self.sample_size * self.actual_prob])
  154. nobs = np.array([self.sample_size, self.sample_size])
  155. z_score, p_value = proportions_ztest(count=count, nobs=nobs, value=None, alternative='two-sided', prop_var=False)
  156. out_dict['z_score'] = z_score
  157. out_dict['p_value'] = p_value
  158. out_dict['sample_size'] = self.sample_size
  159. try:
  160. # 获取该项目下在投的账号
  161. get_acc_sql = """select account_id from ctop_user_allocation where project_id = %s and account_status=0 limit 1""" % \
  162. self.project_id
  163. account_df = pd.read_sql(get_acc_sql, product_engine)
  164. account_id = account_df['account_id'].values[0]
  165. request_data = {'region': city_code_transform(out_dict.get('city')),
  166. 'ages_range': age_code_transform(out_dict.get('age')),
  167. 'gender': gender_code_transform(out_dict.get('gender')),
  168. 'advertiser_id': account_id}
  169. request = requests.post(url=estimate_people_number_url,
  170. headers=headers,
  171. data=json.dumps(request_data, cls=NpEncoder))
  172. response_data = json.loads(request.text)
  173. if response_data['code'] == 0:
  174. out_dict['crowd_coverage_cnt'] = response_data['data'].get('audience_prediction_num')
  175. else:
  176. logger.error("人群预估覆盖接口调用报错,请求数据为%s,返回数据为%s" % (str(request_data), str(response_data)))
  177. except:
  178. logger.error("人群预估覆盖接口调用报错,请求数据为%s,请求地址为%s, 返回数据为%s, 异常信息为%s" %
  179. (str(request_data), estimate_people_number_url, str(response_data), traceback.format_exc()))
  180. out_dict['combine_estimate_prob'] = prob
  181. out_dict['actual_prob'] = self.actual_prob
  182. out_dict['signature'] = self.signature
  183. out_dict['project_id'] = self.project_id
  184. out_dict['target_type'] = self.target_type
  185. out_dict['stat_date'] = str(datetime.now().date())
  186. combine = pd.DataFrame([out_dict])
  187. self.bayes_combine_df = self.bayes_combine_df.append(combine, ignore_index=True)
  188. def write_to_db(self):
  189. self.bayes_combine_df.to_sql(name="ctop_ai_kuaishou_signature_recommended_target_combine",
  190. con=engine,
  191. if_exists='append',
  192. index=False)
  193. if __name__ == '__main__':
  194. # 1、读取配置文件
  195. with open('/data/pythonProject/ai_target/config/config.yaml', mode='r', encoding='utf-8') as f:
  196. config = yaml.load(f.read(), Loader=yaml.FullLoader)
  197. # 1-1 数据库连接引擎,依据开发环境/生产环境 进行切换
  198. if os.getenv('LYY_DEV', 'unknown') == 'dev':
  199. engine = get_db_engine(config['devDB'])
  200. else:
  201. engine = get_db_engine(config['productDB'])
  202. product_engine = get_db_engine(config['productDB'])
  203. # 2、 参与定向组合的维度
  204. target_dim = [key for key in config['bayesDim'].keys() if config['bayesDim'][key]['isOn']]
  205. # 4、计算贝叶斯组合入库
  206. for project_id in config['projectId']:
  207. # 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)
  208. sql = '''
  209. select signature from ctop_kuaishou_report_daily_material
  210. where account_id in (select account_id from ctop_user_allocation where project_id = %s)
  211. and datediff(now(),stat_date) <= %s
  212. group by signature
  213. having sum(activation) >= %s
  214. ''' % (project_id,
  215. config['activeMaterialFilterRule']['days'],
  216. config['activeMaterialFilterRule']['activation'])
  217. df = pd.read_sql(sql, product_engine)
  218. active_signatures = df[~df.signature.isnull()].signature.values
  219. # 4.2 在素材人群报表筛选里面近一个月累计100个激活
  220. sql = """
  221. select signature from ctop_kuaishou_audience_report_daily_age_material
  222. where signature in %s
  223. and datediff(now(),stat_date) <= %s
  224. group by signature
  225. having sum(activation) >= %s
  226. """ % (tuple(active_signatures),
  227. config['getBaysCombineMaterialFilterRule']['days'],
  228. config['getBaysCombineMaterialFilterRule']['activation'])
  229. df = pd.read_sql(sql, product_engine)
  230. signature_lst = df['signature'].values
  231. # 这行代码用于测试
  232. # signature_lst = ['4a56bc00ce51420f3a460ab51c6f5110']
  233. # 4.2 计算指定素材的贝叶斯特征
  234. for sig in signature_lst:
  235. try:
  236. for t_type in config['targetType']:
  237. bayes_feature_lst = []
  238. for dimension in target_dim:
  239. cls = BayesFeatures(sig, t_type, dimension, config['bayesDim'][dimension])
  240. # 计算滑窗组合
  241. cls.get_window_combine()
  242. cls.get_bayes_feature()
  243. bayes_feature_lst.append(cls.bayes_feature)
  244. # 依据特征值,计算多维度的组合预估值
  245. cls = BayesCombine(sig, project_id, t_type, bayes_feature_lst)
  246. cls.get_bayes_estimate()
  247. cls.write_to_db()
  248. logger.info('project_id=%s, signature= %s 完成贝叶斯组合计算!' % (project_id, sig))
  249. except:
  250. logger.error('project_id=%s, signature=%s 贝叶斯组合计算出错,异常信息为%s!' % (project_id, sig, traceback.format_exc()))
  251. logger.info('project_id is %s 完成贝叶斯组合计算!' % project_id)