Commit acd9b679eb18038ce448f53b5c9a6bd349412a86
1 parent
9eb36bd2
doc
Showing
1 changed file
with
432 additions
and
0 deletions
Show diff stats
| ... | ... | @@ -0,0 +1,432 @@ |
| 1 | +# 在线召回Redis Key完整清单 | |
| 2 | + | |
| 3 | +## 📊 Redis数据统计 | |
| 4 | + | |
| 5 | +- **总Key数量**: 211,232 个 | |
| 6 | +- **数据库**: Redis DB 3 (port: 6479) | |
| 7 | + | |
| 8 | +--- | |
| 9 | + | |
| 10 | +## 🔑 一、商品相似度索引(I2I Recall) | |
| 11 | + | |
| 12 | +用于"相似推荐"场景,根据用户点击的商品召回相似商品。 | |
| 13 | + | |
| 14 | +### 1. C++ Swing算法(高性能版本) | |
| 15 | +**Key模式**: `item:similar:swing_cpp:{item_id}` | |
| 16 | + | |
| 17 | +**数量**: 15,310 个 | |
| 18 | + | |
| 19 | +**示例**: | |
| 20 | +```redis | |
| 21 | +Key: item:similar:swing_cpp:3562865 | |
| 22 | +Value: "3605981:0.0209644,2526246:0.0209644,2348869:0.0209644" | |
| 23 | +``` | |
| 24 | + | |
| 25 | +**格式**: JSON字符串,`similar_id:score` 格式 | |
| 26 | + | |
| 27 | +**用途**: | |
| 28 | +- 详情页"相似推荐" | |
| 29 | +- 基于用户点击历史的个性化推荐 | |
| 30 | +- 高性能生产环境推荐 | |
| 31 | + | |
| 32 | +--- | |
| 33 | + | |
| 34 | +### 2. Python Swing算法(标准版本) | |
| 35 | +**Key模式**: `item:similar:swing:{item_id}` | |
| 36 | + | |
| 37 | +**数量**: 169,409 个 | |
| 38 | + | |
| 39 | +**示例**: | |
| 40 | +```redis | |
| 41 | +Key: item:similar:swing:3141390 | |
| 42 | +Value: "3264320:1.2656,3264284:1.2656,3562128:0.4000,..." | |
| 43 | +``` | |
| 44 | + | |
| 45 | +**格式**: JSON字符串,`similar_id:score` 格式 | |
| 46 | + | |
| 47 | +**用途**: | |
| 48 | +- 详情页"看了又看" | |
| 49 | +- 购物车"相关推荐" | |
| 50 | +- 需要归一化分数的场景 | |
| 51 | + | |
| 52 | +--- | |
| 53 | + | |
| 54 | +### 3. Session W2V算法 | |
| 55 | +**Key模式**: `item:similar:session_w2v:{item_id}` | |
| 56 | + | |
| 57 | +**数量**: 26,503 个 | |
| 58 | + | |
| 59 | +**示例**: | |
| 60 | +```redis | |
| 61 | +Key: item:similar:session_w2v:3402580 | |
| 62 | +Value: "商品序列相似度数据" | |
| 63 | +``` | |
| 64 | + | |
| 65 | +**用途**: | |
| 66 | +- 基于用户会话序列的推荐 | |
| 67 | +- 捕捉用户的浏览路径 | |
| 68 | +- 补充Swing算法的不足 | |
| 69 | + | |
| 70 | +--- | |
| 71 | + | |
| 72 | +## 🎯 二、兴趣聚合索引(Interest Recall) | |
| 73 | + | |
| 74 | +用于"热门推荐"场景,根据用户兴趣维度召回相关商品。 | |
| 75 | + | |
| 76 | +### 当前存在的兴趣聚合Key: | |
| 77 | + | |
| 78 | +#### 1. 平台维度 | |
| 79 | +```redis | |
| 80 | +Key: interest:hot:platform:essaone | |
| 81 | +Value: [3582462, 3582428, 3582384, ...] # 1000个商品 | |
| 82 | +``` | |
| 83 | + | |
| 84 | +#### 2. 客户端平台维度 | |
| 85 | +```redis | |
| 86 | +Key: interest:hot:client_platform:pc | |
| 87 | +Value: [3582462, 3582428, 3582384, ...] # 1000个商品 | |
| 88 | +``` | |
| 89 | + | |
| 90 | +#### 3. 平台+客户端组合维度 | |
| 91 | +```redis | |
| 92 | +Key: interest:hot:platform_client:essaone_pc | |
| 93 | +Value: [平台+客户端的组合热门商品] | |
| 94 | +``` | |
| 95 | + | |
| 96 | +#### 4. 一级分类维度 | |
| 97 | +```redis | |
| 98 | +Key: interest:hot:category_level1:1 | |
| 99 | +Value: [一级分类的热门商品] | |
| 100 | +``` | |
| 101 | + | |
| 102 | +#### 5. 二级分类维度 | |
| 103 | +```redis | |
| 104 | +Key: interest:hot:category_level2:29 | |
| 105 | +Value: [3482467, 2708138, 1008982, ...] # 1000个商品 | |
| 106 | +``` | |
| 107 | + | |
| 108 | +#### 6. 三级分类维度 | |
| 109 | +```redis | |
| 110 | +Key: interest:hot:category_level3:2040 | |
| 111 | +Value: [三级分类的热门商品] | |
| 112 | +``` | |
| 113 | + | |
| 114 | +#### 7. 四级分类维度 | |
| 115 | +```redis | |
| 116 | +Key: interest:hot:category_level4:2040 | |
| 117 | +Value: [四级分类的热门商品] | |
| 118 | +``` | |
| 119 | + | |
| 120 | +#### 8. 平台+二级分类组合维度 | |
| 121 | +```redis | |
| 122 | +Key: interest:hot:platform_category2:essaone_29 | |
| 123 | +Value: [260个商品] | |
| 124 | +``` | |
| 125 | + | |
| 126 | +#### 9. 平台+三级分类组合维度 | |
| 127 | +```redis | |
| 128 | +Key: interest:hot:platform_category3:essaone_2040 | |
| 129 | +Value: [231个商品] | |
| 130 | +``` | |
| 131 | + | |
| 132 | +#### 10. 供应商维度 | |
| 133 | +```redis | |
| 134 | +Key: interest:hot:supplier:24947 | |
| 135 | +Value: [1566197] # 1个商品 | |
| 136 | +``` | |
| 137 | + | |
| 138 | +--- | |
| 139 | + | |
| 140 | +## 📋 在线召回使用策略 | |
| 141 | + | |
| 142 | +### 场景1: 详情页相似推荐 | |
| 143 | + | |
| 144 | +**召回Key**: | |
| 145 | +```python | |
| 146 | +# 根据当前浏览的商品ID | |
| 147 | +item_id = request.sku_id | |
| 148 | + | |
| 149 | +recall_keys = [ | |
| 150 | + f"item:similar:swing_cpp:{item_id}", # 优先使用高性能版本 | |
| 151 | + f"item:similar:swing:{item_id}", # 降级使用 | |
| 152 | + f"item:similar:session_w2v:{item_id}", # 补充推荐 | |
| 153 | +] | |
| 154 | +``` | |
| 155 | + | |
| 156 | +**召回流程**: | |
| 157 | +1. 优先查询 `swing_cpp`(C++高性能版本) | |
| 158 | +2. 如果没有结果,查询 `swing`(Python标准版本) | |
| 159 | +3. 如果还没结果,查询 `session_w2v`(会话序列版本) | |
| 160 | +4. 最终结果融合多个算法的相似商品 | |
| 161 | + | |
| 162 | +--- | |
| 163 | + | |
| 164 | +### 场景2: 首页热门推荐 | |
| 165 | + | |
| 166 | +**召回Key**: | |
| 167 | +```python | |
| 168 | +# 根据用户特征和上下文 | |
| 169 | +user_profile = get_user_profile(user_id) | |
| 170 | +context = request.context # {platform, client, category, ...} | |
| 171 | + | |
| 172 | +recall_keys = [] | |
| 173 | + | |
| 174 | +# 1. 组合维度推荐(最精准) | |
| 175 | +if context.get('platform') and context.get('category_level2'): | |
| 176 | + recall_keys.append( | |
| 177 | + f"interest:hot:platform_category2:{context['platform']}_{context['category_level2']}" | |
| 178 | + ) | |
| 179 | + | |
| 180 | +# 2. 单一分类维度 | |
| 181 | +if context.get('category_level2'): | |
| 182 | + recall_keys.append(f"interest:hot:category_level2:{context['category_level2']}") | |
| 183 | + | |
| 184 | +# 3. 平台维度 | |
| 185 | +if context.get('platform'): | |
| 186 | + recall_keys.append(f"interest:hot:platform:{context['platform']}") | |
| 187 | + | |
| 188 | +# 4. 客户端平台维度 | |
| 189 | +if context.get('client'): | |
| 190 | + recall_keys.append(f"interest:hot:client_platform:{context['client']}") | |
| 191 | + | |
| 192 | +# 5. 全局热门(兜底) | |
| 193 | +recall_keys.append("interest:hot:platform:essaone") # 全局热门 | |
| 194 | +``` | |
| 195 | + | |
| 196 | +**召回优先级**: | |
| 197 | +1. 组合维度(如 platform_category2) - 最精准 | |
| 198 | +2. 分类维度(category_level2/3/4) | |
| 199 | +3. 平台维度(platform) | |
| 200 | +4. 全局热门 - 兜底策略 | |
| 201 | + | |
| 202 | +--- | |
| 203 | + | |
| 204 | +### 场景3: 用户点击历史推荐 | |
| 205 | + | |
| 206 | +**召回Key**: | |
| 207 | +```python | |
| 208 | +# 根据用户点击的商品列表 | |
| 209 | +clicked_items = [3141390, 3141390, 3606104, 3606102, ...] | |
| 210 | + | |
| 211 | +all_recalled_items = [] | |
| 212 | + | |
| 213 | +for item_id in clicked_items: | |
| 214 | + # 查询相似商品 | |
| 215 | + similar_items = redis.get(f"item:similar:swing_cpp:{item_id}") | |
| 216 | + if not similar_items: | |
| 217 | + similar_items = redis.get(f"item:similar:swing:{item_id}") | |
| 218 | + | |
| 219 | + all_recalled_items.extend(similar_items) | |
| 220 | + | |
| 221 | +# 去重并排序 | |
| 222 | +final_items = deduplicate_and_rank(all_recalled_items) | |
| 223 | +``` | |
| 224 | + | |
| 225 | +--- | |
| 226 | + | |
| 227 | +### 场景4: 类别偏好推荐 | |
| 228 | + | |
| 229 | +**召回Key**: | |
| 230 | +```python | |
| 231 | +# 根据用户收藏的类别 | |
| 232 | +collected_categories = [1541, 137, 153, 1561, 1689, 1691] | |
| 233 | + | |
| 234 | +recall_keys = [] | |
| 235 | + | |
| 236 | +for cat_id in collected_categories: | |
| 237 | + # 查询该类别的热门商品 | |
| 238 | + category_key = f"interest:hot:category_level2:{cat_id}" | |
| 239 | + recall_keys.append(category_key) | |
| 240 | + | |
| 241 | +# 批量查询 | |
| 242 | +hot_items = redis.mget(recall_keys) | |
| 243 | +``` | |
| 244 | + | |
| 245 | +--- | |
| 246 | + | |
| 247 | +## ⚠️ 注意事项 | |
| 248 | + | |
| 249 | +### 1. Key存在性检查 | |
| 250 | +```python | |
| 251 | +def safe_get_recall(key): | |
| 252 | + """安全获取召回数据,带降级策略""" | |
| 253 | + result = redis.get(key) | |
| 254 | + if not result: | |
| 255 | + # 降级策略:使用更通用的key | |
| 256 | + if ":category_level2:" in key: | |
| 257 | + # 尝试用更高级别的分类 | |
| 258 | + fallback_key = key.replace("category_level2", "category_level1") | |
| 259 | + result = redis.get(fallback_key) | |
| 260 | + elif "platform_category" in key: | |
| 261 | + # 尝试单独的平台或分类 | |
| 262 | + result = redis.get(key.replace("platform_category2:", "category_level2:")) | |
| 263 | + | |
| 264 | + return result or [] | |
| 265 | +``` | |
| 266 | + | |
| 267 | +### 2. 数据格式 | |
| 268 | +- **I2I数据**: JSON字符串,格式为 `"item_id1:score1,item_id2:score2"` | |
| 269 | +- **Interest数据**: JSON数组,格式为 `[item_id1, item_id2, ...]` | |
| 270 | + | |
| 271 | +### 3. 解析方法 | |
| 272 | +```python | |
| 273 | +import json | |
| 274 | + | |
| 275 | +# I2I数据解析 | |
| 276 | +def parse_i2i_data(value): | |
| 277 | + items = [] | |
| 278 | + for pair in value.split(','): | |
| 279 | + if ':' in pair: | |
| 280 | + item_id, score = pair.split(':') | |
| 281 | + items.append([int(item_id), float(score)]) | |
| 282 | + return items | |
| 283 | + | |
| 284 | +# Interest数据解析 | |
| 285 | +def parse_interest_data(value): | |
| 286 | + return json.loads(value) | |
| 287 | +``` | |
| 288 | + | |
| 289 | +### 4. 推荐数量 | |
| 290 | +- 每个相似推荐key通常包含10-50个商品 | |
| 291 | +- 每个兴趣聚合key通常包含100-1000个商品 | |
| 292 | +- 最终推荐数量建议:取Top 20-50个商品 | |
| 293 | + | |
| 294 | +--- | |
| 295 | + | |
| 296 | +## 📊 Key覆盖率统计 | |
| 297 | + | |
| 298 | +| Key类型 | 数量 | 覆盖率 | 使用优先级 | | |
| 299 | +|---------|------|--------|-----------| | |
| 300 | +| item:similar:swing | 170,116 | ~80% | ⭐⭐⭐⭐⭐ | | |
| 301 | +| item:similar:content_name | 129,103 | ~70% | ⭐⭐⭐⭐ | | |
| 302 | +| item:similar:session_w2v | 52,339 | ~30% | ⭐⭐⭐ | | |
| 303 | +| item:similar:deepwalk | 49,052 | ~30% | ⭐⭐⭐ | | |
| 304 | +| item:similar:swing_cpp | 15,310 | ~10% | ⭐⭐⭐⭐⭐ | | |
| 305 | +| item:similar:content_pic | 12,118 | ~7% | ⭐⭐ | | |
| 306 | +| interest:hot:* | 10 | ~5% | ⭐⭐ | | |
| 307 | + | |
| 308 | +**说明**: | |
| 309 | +- Swing覆盖商品约10万+,应优先使用 | |
| 310 | +- 兴趣聚合数据较少,建议作为补充召回策略 | |
| 311 | + | |
| 312 | +--- | |
| 313 | + | |
| 314 | +## 🔧 修复建议 | |
| 315 | + | |
| 316 | +### 缺失的关键数据: | |
| 317 | +1. **DeepWalk索引**: `item:similar:deepwalk:{item_id}` - 缺失 | |
| 318 | +2. **内容相似度**: `item:similar:content_name:{item_id}` - 缺失 | |
| 319 | +3. **图片相似度**: `item:similar:content_pic:{item_id}` - 缺失 | |
| 320 | +4. **更多兴趣聚合维度**: | |
| 321 | + - `interest:hot:category_level2:{id}` - 缺失(1541, 137, 153等) | |
| 322 | + - `interest:cart:*` - 缺失 | |
| 323 | + - `interest:new:*` - 缺失 | |
| 324 | + - `interest:global:*` - 缺失 | |
| 325 | + | |
| 326 | +### 建议操作: | |
| 327 | +```bash | |
| 328 | +# 1. 运行兴趣聚合任务生成更多维度数据 | |
| 329 | +cd /home/tw/recommendation/offline_tasks | |
| 330 | +python3 scripts/interest_aggregation.py | |
| 331 | + | |
| 332 | +# 2. 加载更多兴趣聚合数据到Redis | |
| 333 | +python3 scripts/load_index_to_redis.py --load-interest | |
| 334 | + | |
| 335 | +# 3. 定期更新数据(每天) | |
| 336 | +crontab -e | |
| 337 | +# 添加: 0 3 * * * /home/tw/recommendation/offline_tasks/update_indices.sh | |
| 338 | +``` | |
| 339 | + | |
| 340 | +--- | |
| 341 | + | |
| 342 | +## 📞 使用示例 | |
| 343 | + | |
| 344 | +### Python召回示例 | |
| 345 | + | |
| 346 | +```python | |
| 347 | +import redis | |
| 348 | +import json | |
| 349 | + | |
| 350 | +redis_client = redis.Redis( | |
| 351 | + host='localhost', | |
| 352 | + port=6479, | |
| 353 | + db=3, | |
| 354 | + password='BMfv5aI31kgHWtlx', | |
| 355 | + decode_responses=True | |
| 356 | +) | |
| 357 | + | |
| 358 | +def recall_similar_items(item_id): | |
| 359 | + """召回相似商品""" | |
| 360 | + # 按优先级查询 | |
| 361 | + keys = [ | |
| 362 | + f"item:similar:swing_cpp:{item_id}", | |
| 363 | + f"item:similar:swing:{item_id}", | |
| 364 | + f"item:similar:session_w2v:{item_id}" | |
| 365 | + ] | |
| 366 | + | |
| 367 | + for key in keys: | |
| 368 | + value = redis_client.get(key) | |
| 369 | + if value: | |
| 370 | + # 解析数据 | |
| 371 | + items = [] | |
| 372 | + for pair in value.split(','): | |
| 373 | + if ':' in pair: | |
| 374 | + similar_id, score = pair.split(':') | |
| 375 | + items.append([int(similar_id), float(score)]) | |
| 376 | + return items | |
| 377 | + | |
| 378 | + return [] | |
| 379 | + | |
| 380 | +def recall_hot_items(platform, category_level2): | |
| 381 | + """召回热门商品""" | |
| 382 | + key = f"interest:hot:platform_category2:{platform}_{category_level2}" | |
| 383 | + value = redis_client.get(key) | |
| 384 | + | |
| 385 | + if value: | |
| 386 | + return json.loads(value) | |
| 387 | + | |
| 388 | + # 降级策略 | |
| 389 | + category_key = f"interest:hot:category_level2:{category_level2}" | |
| 390 | + value = redis_client.get(category_key) | |
| 391 | + if value: | |
| 392 | + return json.loads(value) | |
| 393 | + | |
| 394 | + # 最终降级:全局热门 | |
| 395 | + global_key = f"interest:hot:platform:{platform}" | |
| 396 | + value = redis_client.get(global_key) | |
| 397 | + if value: | |
| 398 | + return json.loads(value) | |
| 399 | + | |
| 400 | + return [] | |
| 401 | +``` | |
| 402 | + | |
| 403 | +--- | |
| 404 | + | |
| 405 | +## 🎯 总结 | |
| 406 | + | |
| 407 | +**当前Redis中有数据的Key类型**: | |
| 408 | +1. ✅ `item:similar:swing:{item_id}` - 170,116个 | |
| 409 | +2. ✅ `item:similar:content_name:{item_id}` - 129,103个 | |
| 410 | +3. ✅ `item:similar:session_w2v:{item_id}` - 52,339个 | |
| 411 | +4. ✅ `item:similar:deepwalk:{item_id}` - 49,052个 | |
| 412 | +5. ✅ `item:similar:swing_cpp:{item_id}` - 15,310个 | |
| 413 | +6. ✅ `item:similar:content_pic:{item_id}` - 12,118个 | |
| 414 | +7. ✅ `interest:hot:platform:{platform}` - 1个 | |
| 415 | +8. ✅ `interest:hot:client_platform:{client}` - 1个 | |
| 416 | +9. ✅ `interest:hot:platform_client:{platform}_{client}` - 1个 | |
| 417 | +10. ✅ `interest:hot:category_level1:{id}` - 1个 | |
| 418 | +11. ✅ `interest:hot:category_level2:{id}` - 1个 | |
| 419 | +12. ✅ `interest:hot:category_level3:{id}` - 1个 | |
| 420 | +13. ✅ `interest:hot:category_level4:{id}` - 1个 | |
| 421 | +14. ✅ `interest:hot:platform_category2:{platform}_{id}` - 1个 | |
| 422 | +15. ✅ `interest:hot:platform_category3:{platform}_{id}` - 1个 | |
| 423 | +16. ✅ `interest:hot:supplier:{id}` - 1个 | |
| 424 | + | |
| 425 | +**缺失但应该有的Key**: | |
| 426 | +- ❌ 更多类别的 `interest:hot:category_level2:{id}` | |
| 427 | +- ❌ `interest:cart:*` (加购热门) | |
| 428 | +- ❌ `interest:new:*` (新品) | |
| 429 | +- ❌ `interest:global:*` (全局热门) | |
| 430 | + | |
| 431 | +**建议**: 优先使用现有的高质量数据,同时尽快补充缺失的兴趣聚合数据。 | |
| 432 | + | ... | ... |