123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230 |
- import os
- import yaml
- from common_func import get_db_engine
- import pandas as pd
- from sklearn.preprocessing import StandardScaler
- from sklearn.neighbors import NearestNeighbors
- import numpy as np
- import datetime
- import sys
- import traceback
- import pickle
- import logging
- from concurrent_log import ConcurrentTimedRotatingFileHandler
- log_formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s', '%Y/%m/%d %I:%M:%S %p')
- log_handler = ConcurrentTimedRotatingFileHandler("/data/pythonProject/similar_material/logs/get_similar_material.log", when="midnight",
- backupCount=100)
- log_handler.setFormatter(log_formatter)
- logger = logging.getLogger('historical_material_logger')
- logger.addHandler(log_handler)
- logger.setLevel(logging.DEBUG)
- logger.info("historical_material_logger started!")
- print('id of historical_material_logger %s' % id(logger))
- def scaler_feature(data, axis=0):
- """
- :param data: pd.DataFrame
- :param axis: 0-横向归一化 , 1-纵向归一化
- :return: z = (x - u) / s
- """
- scaler = StandardScaler()
- scaler_data = None
- if axis == 0:
- scaler_data = scaler.fit_transform(data.T).T
- if axis == 1:
- scaler_data = scaler.fit_transform(data)
- scaler_df = pd.DataFrame(scaler_data)
- scaler_df.index = data.index
- scaler_df.columns = [col + '_scaler' for col in data.columns]
- return scaler_df
- def get_segment_features(table, signature, dim_col, project_id, col_prefix):
- # 1 获取原始数据
- get_org_data_sql = """
- select signature, aclick, bclick, activation, charge, %s from %s
- where signature in %s and project_id = %s
- """ % (dim_col, table, tuple(signature), project_id)
- org_df = pd.read_sql(get_org_data_sql, read_engine)
- # 2 计算各分段的 行为率、激活率、激活成本,并进行横向归一化
- g = org_df.groupby(['signature', dim_col]).agg({'activation': sum, 'aclick': sum, 'bclick': sum, 'charge': sum}).unstack()
- # 该维度下有哪些分段值:如 gender维度下有:男、女、其他 , 需要去掉 '其他',因为业务含义不明确,没法解释模型效果
- segment_cols = list(g['charge'].columns)
- if '其他' in segment_cols:
- segment_cols.remove('其他')
- segment_fea_df = pd.DataFrame()
- for col in segment_cols:
- segment_fea_df[col + '_bclick_aclick_pct'] = g[('bclick', col)] / g[('aclick', col)]
- segment_fea_df[col + '_activation_bclick_pct'] = g[('activation', col)] / g[('bclick', col)]
- segment_fea_df[col + '_cost'] = g[('charge', col)] / g[('activation', col)]
- # 3 计算累计激活中各分段的激活个数占比,并进行横向归一化
- for col in segment_cols:
- segment_fea_df[col + '_activation_pct'] = g[('activation', col)] / (g[[('activation', item) for item in segment_cols]].sum(axis=1))
- # 4 对 np.nan 和 np.inf 进行处理
- # 4-1 cost
- # np.inf : 分子>0/分母=0, 表示有消耗但是没有激活数,填充为 999
- # np.nan :分子=0/分母=0, 表示没有消耗也没有激活数,填充为 -1
- cost_cols = [col for col in segment_fea_df.columns if 'cost' in col]
- segment_fea_df[cost_cols] = segment_fea_df[cost_cols].replace({np.nan: -1, np.inf: 999})
- # 4-2 占比类
- # np.inf :分子>0/分母=0 ,如行为率:表示有点击但没有曝光,属于数据bug情况。需要剔除该素材并写入日志
- # np.nan :分子=0/分母=0 ,如激活率:表示没有激活也没有点击,填充为 0
- pct_cols = [col for col in segment_fea_df.columns if 'pct' in col]
- segment_fea_df[pct_cols] = segment_fea_df[pct_cols].replace({np.nan: 0})
- inf_df = np.isinf(segment_fea_df[pct_cols]).sum(axis=1)
- remove_sig_lst = inf_df[inf_df >= 1].index
- if len(remove_sig_lst):
- logger.error('素材的人群报表数据存在异常(如:点击数据大于0,但是曝光数据等于0),以下素材不参与相似素材计算:%s' % remove_sig_lst)
- print(remove_sig_lst)
- segment_fea_df = segment_fea_df[~segment_fea_df.index.isin(remove_sig_lst)]
- # 5 对特征进行横向归一化
- segment_fea_scaler_df = pd.DataFrame()
- for col in ['bclick_aclick_pct', 'activation_bclick_pct', 'cost', 'activation_pct']:
- tmp_df = segment_fea_df.filter(like=col)
- scaler_df = scaler_feature(tmp_df, axis=0)
- segment_fea_scaler_df = pd.concat([segment_fea_scaler_df, scaler_df], axis=1)
- # 6 col_prefix and index
- segment_fea_scaler_df.columns = [col_prefix + col for col in segment_fea_scaler_df.columns]
- segment_fea_scaler_df.index = g.index
- return segment_fea_scaler_df
- def get_signature_features(signature, project_id):
- """
- 计算素材整体的 行为率、激活率、激活成本
- :param project_id:
- :param signature:
- :return:
- """
- get_data_sql = """ select signature, sum(aclick) aclick, sum(bclick) bclick, sum(activation) activation, sum(charge) charge
- from ctop_kuaishou_audience_report_daily_age_material
- where signature in %s and project_id = %s
- group by signature
- """ % (tuple(signature), project_id)
- sig_org_df = pd.read_sql(get_data_sql, read_engine)
- sig_fea_df = pd.DataFrame()
- sig_fea_df['cost'] = sig_org_df['charge'] / sig_org_df['activation']
- sig_fea_df['bclick_aclick_pct'] = sig_org_df['bclick'] / sig_org_df['aclick']
- sig_fea_df['activation_bclick_pct'] = sig_org_df['activation'] / sig_org_df['bclick']
- # 对特征进行纵向的归一化
- sig_fea_df.index = sig_org_df['signature'].values
- sig_fea_scaler_df = scaler_feature(sig_fea_df, axis=1)
- return sig_fea_scaler_df
- if __name__ == '__main__':
- try:
- # 1、读取配置文件
- with open('/data/pythonProject/similar_material/config/config.yaml', mode='r', encoding='utf-8') as f:
- config = yaml.load(f.read(), Loader=yaml.FullLoader)
- # 1-1 数据库连接引擎,依据开发环境/生产环境 进行切换
- # 读数据库引擎使用生产数据库,写数据库引擎依据系统环境进行切换(测试数据库/生产数据库)
- if os.getenv('LYY_DEV', 'unknown') == 'dev':
- write_engine = get_db_engine(config['devDB'])
- else:
- write_engine = get_db_engine(config['productDB'])
- read_engine = get_db_engine(config['productDB'])
- # 2、按项目计算相似素材
- project_id_lst = config['projectId']
- for project_id in project_id_lst:
- # 获取历史所有达标的素材 - - 暂定为激活个数达到100个的素材
- get_sig_sql = """select signature, sum(activation) activation
- from ctop_kuaishou_audience_report_daily_age_material
- where project_id = %s
- group by signature
- having sum(activation) >= 100
- """ % project_id
- sig_df = pd.read_sql(get_sig_sql, read_engine)
- sig_df = sig_df[~sig_df.signature.isnull()]
- sig_lst = list(sig_df.signature.values)
- if sig_lst is None:
- logger.info("没有达标的素材,系统退出!")
- sys.exit(0)
- if len(sig_lst) == 1:
- logger.info("达标的素材只有一个,没法计算得到与之相似的素材,系统退出!")
- sys.exit(1)
- # 2、构造特征
- # 2-1 获取素材整体的特征情况:行为率、激活率、激活成本
- sig_feature_df = get_signature_features(sig_lst, project_id)
- # 2-2 获取各维度(age,gender)各分段(男、女等)的行为率、激活率、激活成本、激活占比,并进行纵向归一化
- # table_dim_tuple 中的元素:表名称、分段名称、构造特征名称的前缀
- table_dim_tuple = [('ctop_kuaishou_audience_report_daily_age_material', 'age_segment', 'age_'),
- ('ctop_kuaishou_audience_report_daily_gender_material', 'gender', 'gender_')]
- segment_feature_df = pd.DataFrame()
- for table, dim_col, col_prefix in table_dim_tuple:
- tmp_df = get_segment_features(table, sig_lst, dim_col, project_id, col_prefix)
- segment_feature_df = pd.concat([segment_feature_df, tmp_df], axis=1)
- # 2-3 素材整体的特征 和 各维度各分段的特征 进行合并,得到最终的特征DataFrame
- final_feature_df = pd.concat([sig_feature_df, segment_feature_df], axis=1)
- # 部分素材在 age 或者 gender 维度上没有数据,特征 concat 时,会导致空值
- # 去掉这部分的素材,防止在调用 NearestNeighbors 时,报错
- null_df = final_feature_df.isnull().sum(axis=1)
- exist_null_fea_sig_lst = null_df[null_df >= 1].index
- # 记录存在部分维度数据缺失的素材
- if len(exist_null_fea_sig_lst) > 0:
- logger.info("以下这些素材存在部分维度数据缺失:%s" % exist_null_fea_sig_lst)
- final_feature_df = final_feature_df[~final_feature_df.index.isin(exist_null_fea_sig_lst)]
- # 3 Finding the Nearest Neighbors
- k = config['nearestNeighbors']['k']
- distance_threshold = config['nearestNeighbors']['distanceThreshold']
- X = final_feature_df.values
- nbrs = NearestNeighbors(n_neighbors=k, algorithm='ball_tree').fit(X)
- distances, indices = nbrs.kneighbors(X)
- # 3-1 两点之间的距离超过阈值时,不作为相邻素材
- # mask 中,为 True 的元素替换为 NaN
- mask = distances > distance_threshold
- indices_mask_df = pd.DataFrame(indices).mask(mask)
- indices_mask_df = indices_mask_df.astype(dtype=pd.Int64Dtype())
- # 3-2 把 index 替换为具体的 signatures
- sig_lst = final_feature_df.index.values
- indices_mask_df.index = sig_lst
- indices_mask_df = indices_mask_df.applymap(lambda x: sig_lst[x] if not pd.isnull(x) else x)
- res_lst = []
- for index, row in indices_mask_df.iterrows():
- others = [item for item in row if (not pd.isnull(item)) and (item != index)]
- res_lst.append([index] + others)
- res_df = pd.DataFrame(res_lst)
- res_df.columns = ['signature'] + ['similar_signature_' + str(i + 1) for i in range(res_df.shape[1] - 1)]
- columns = ['signature'] + ['similar_signature_' + str(i + 1) for i in range(k-1)]
- res_df = res_df.reindex(columns=columns)
- # 4 结果入库
- res_df['stat_date'] = datetime.date.today().strftime('%Y-%m-%d')
- res_df['project_id'] = project_id
- res_df.to_sql(name="ctop_ai_kuaishou_signature_similar",
- con=write_engine,
- if_exists='append',
- index=False)
- logger.info("project_id = %s 的素材相似度计算完成!" % project_id)
- except:
- logger.error("异常信息为:%s" % (traceback.format_exc()))
|