""" i2i - Swing算法实现 基于用户行为的物品相似度计算 参考item_sim.py的数据格式，适配真实数据 """ import sys import os sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))) import pandas as pd import math from collections import defaultdict import argparse import json from datetime import datetime, timedelta from db_service import create_db_connection from offline_tasks.config.offline_config import ( DB_CONFIG, OUTPUT_DIR, I2I_CONFIG, get_time_range, DEFAULT_LOOKBACK_DAYS, DEFAULT_I2I_TOP_N ) def calculate_time_weight(event_time, reference_time, decay_factor=0.95, days_unit=30): """ 计算时间衰减权重 Args: event_time: 事件发生时间 reference_time: 参考时间（通常是当前时间） decay_factor: 衰减因子 days_unit: 衰减单位（天） Returns: 时间权重 """ if pd.isna(event_time): return 1.0 time_diff = (reference_time - event_time).days if time_diff < 0: return 1.0 # 计算衰减权重 periods = time_diff / days_unit weight = math.pow(decay_factor, periods) return weight def swing_algorithm(df, alpha=0.5, time_decay=True, decay_factor=0.95): """ Swing算法实现 Args: df: DataFrame with columns: user_id, item_id, weight, create_time alpha: Swing算法的alpha参数 time_decay: 是否使用时间衰减 decay_factor: 时间衰减因子 Returns: Dict[item_id, List[Tuple(similar_item_id, score)]] """ # 如果使用时间衰减，计算时间权重 reference_time = datetime.now() if time_decay and 'create_time' in df.columns: df['time_weight'] = df['create_time'].apply( lambda x: calculate_time_weight(x, reference_time, decay_factor) ) df['weight'] = df['weight'] * df['time_weight'] # 构建用户-物品倒排索引 user_items = defaultdict(set) item_users = defaultdict(set) item_freq = defaultdict(float) for _, row in df.iterrows(): user_id = row['user_id'] item_id = row['item_id'] weight = row['weight'] user_items[user_id].add(item_id) item_users[item_id].add(user_id) item_freq[item_id] += weight print(f"Total users: {len(user_items)}, Total items: {len(item_users)}") # 计算物品相似度 item_sim_dict = defaultdict(lambda: defaultdict(float)) # 遍历每个物品对 for item_i in item_users: users_i = item_users[item_i] # 找到所有与item_i共现的物品 for item_j in item_users: if item_i >= item_j: # 避免重复计算 continue users_j = item_users[item_j] common_users = users_i & users_j if len(common_users) < 2: continue # 计算Swing相似度 sim_score = 0.0 common_users_list = list(common_users) for idx_u in range(len(common_users_list)): user_u = common_users_list[idx_u] items_u = user_items[user_u] for idx_v in range(idx_u + 1, len(common_users_list)): user_v = common_users_list[idx_v] items_v = user_items[user_v] # 计算用户u和用户v的共同物品数 common_items = items_u & items_v # Swing公式 sim_score += 1.0 / (alpha + len(common_items)) item_sim_dict[item_i][item_j] = sim_score item_sim_dict[item_j][item_i] = sim_score # 对相似度进行归一化并排序 result = {} for item_i in item_sim_dict: sims = item_sim_dict[item_i] # 归一化（可选） # 按相似度排序 sorted_sims = sorted(sims.items(), key=lambda x: -x[1]) result[item_i] = sorted_sims return result def main(): parser = argparse.ArgumentParser(description='Run Swing algorithm for i2i similarity') parser.add_argument('--alpha', type=float, default=I2I_CONFIG['swing']['alpha'], help='Alpha parameter for Swing algorithm') parser.add_argument('--top_n', type=int, default=DEFAULT_I2I_TOP_N, help=f'Top N similar items to output (default: {DEFAULT_I2I_TOP_N})') parser.add_argument('--lookback_days', type=int, default=DEFAULT_LOOKBACK_DAYS, help=f'Number of days to look back for user behavior (default: {DEFAULT_LOOKBACK_DAYS})') parser.add_argument('--time_decay', action='store_true', default=True, help='Use time decay for behavior weights') parser.add_argument('--decay_factor', type=float, default=0.95, help='Time decay factor') parser.add_argument('--output', type=str, default=None, help='Output file path') args = parser.parse_args() # 创建数据库连接 print("Connecting to database...") engine = create_db_connection( DB_CONFIG['host'], DB_CONFIG['port'], DB_CONFIG['database'], DB_CONFIG['username'], DB_CONFIG['password'] ) # 获取时间范围 start_date, end_date = get_time_range(args.lookback_days) print(f"Fetching data from {start_date} to {end_date}...") # SQL查询 - 获取用户行为数据 sql_query = f""" SELECT se.anonymous_id AS user_id, se.item_id, se.event AS event_type, se.create_time, pgs.name AS item_name FROM sensors_events se LEFT JOIN prd_goods_sku pgs ON se.item_id = pgs.id WHERE se.event IN ('contactFactory', 'addToPool', 'addToCart', 'purchase') AND se.create_time >= '{start_date}' AND se.create_time <= '{end_date}' AND se.item_id IS NOT NULL AND se.anonymous_id IS NOT NULL ORDER BY se.create_time """ print("Executing SQL query...") df = pd.read_sql(sql_query, engine) print(f"Fetched {len(df)} records") # 转换create_time为datetime df['create_time'] = pd.to_datetime(df['create_time']) # 定义行为权重 behavior_weights = { 'contactFactory': 5.0, 'addToPool': 2.0, 'addToCart': 3.0, 'purchase': 10.0 } # 添加权重列 df['weight'] = df['event_type'].map(behavior_weights).fillna(1.0) # 运行Swing算法 print("Running Swing algorithm...") result = swing_algorithm( df, alpha=args.alpha, time_decay=args.time_decay, decay_factor=args.decay_factor ) # 创建item_id到name的映射 item_name_map = dict(zip(df['item_id'].unique(), df.groupby('item_id')['item_name'].first())) # 输出结果 output_file = args.output or os.path.join(OUTPUT_DIR, f'i2i_swing_{datetime.now().strftime("%Y%m%d")}.txt') print(f"Writing results to {output_file}...") with open(output_file, 'w', encoding='utf-8') as f: for item_id, sims in result.items(): item_name = item_name_map.get(item_id, 'Unknown') # 只取前N个最相似的商品 top_sims = sims[:args.top_n] if not top_sims: continue # 格式：item_id \t item_name \t similar_item_id1:score1,similar_item_id2:score2,... sim_str = ','.join([f'{sim_id}:{score:.4f}' for sim_id, score in top_sims]) f.write(f'{item_id}\t{item_name}\t{sim_str}\n') print(f"Done! Generated i2i similarities for {len(result)} items") print(f"Output saved to: {output_file}") if __name__ == '__main__': main()