知识卡片
候选生成与排序分担规模
两阶段推荐先从大库找出较小候选集合,再利用更细特征评分排序。候选生成控制覆盖和计算规模,排序决定最终列表;未进入候选的物品无法靠后续排序补回。书中YouTube图为历史方案示意,不代表当前线上架构。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
The item generation process captures events from the user's YouTube history, where a smaller subset of history is retrieved. The procedure gives a general personalization using CF. Subsequently, the ranking process presents a more refined list of recommendations by assigning a score for each video, and videos with the highest score are recommended to the user.
参考来源
- 位置:《Artificial Intelligence and Data Science in Recommendation System: Current Trends, Technologies, and Applications》第03章《深度推荐与评估指标》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Artificial Intelligence and Data Science in Recommendation System: Current Trends, Technologies, and Applications/_epub-src/OEBPS/Text/9789815136746-C3.html)
- 结论依据:摘录中关于候选生成与排序分担规模的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:The item generation process captures events from the user's YouTube history, where a smaller subset of history is retrieved. The procedure gives a general personalization using CF. Subsequently, the ranking process presents a more refined list of recommendations by assigning a score for each video, and videos with the highest score are recommended to the user.