{ "source_file": "0930", "source_platform": "知乎专栏", "author": "王喆的机器学习笔记", "title_listed": "重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文", "keywords": "推荐系统,Youtube,DNN", "domain": "机器学习", "content_type": "论文精读", "is_paid_flag": "否", "url": "https://zhuanlan.zhihu.com/p/52169807", "publish_date": "2018", "corpus_stage": "推荐系统", "notes": "YouTube 推荐系统论文精读", "id": "497ff60141d8", "collected_at": "2026-10-02T18:37:52", "title": "重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文 - 知乎", "body_markdown": "[![](https://zhuanlan.zhihu.com/p/52169807)](javascript:void\\(0\\))\n![ZhiHu logo](../images/body-e1fa91dbdb48b8055184.png)\n[](https://www.zhihu.com)\n[关注](https://www.zhihu.com/signin?next=%2Ffollow)[推荐](https://www.zhihu.com/signin?next=%2F)[热榜](https://www.zhihu.com/signin?next=%2Fhot)[专栏](https://www.zhihu.com/signin?next=%2Fcolumn-square)[圈子](https://www.zhihu.com/signin?next=%2Fring-feeds)[AI WorksBeta ](https://www.zhihu.com/project-square)[故事](https://www.zhihu.com/fiore/h5/vip-web)\n​\n[直答](https://zhida.zhihu.com/)\n切换模式\n登录/注册\n重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文\n首发于[王喆的机器学习笔记](https://www.zhihu.com/column/wangzhenotes)\n切换模式\n请您登录后查看更多专业优质内容。\n![重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文](../images/000.img)\n# 重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文\n[![王喆](../images/008.jpg)](https://www.zhihu.com/people/wang-zhe-58)\n[王喆](https://www.zhihu.com/people/wang-zhe-58)\n[​![](../images/body-2dd3e5c360100c555650.png)](https://www.zhihu.com/question/48509984)\n数据挖掘等 3 个话题下的优秀答主\n[收录于 · 王喆的机器学习笔记](https://www.zhihu.com/column/wangzhenotes)\n3252 人赞同了该文章\n这里是**[王喆的机器学习笔记](https://zhuanlan.zhihu.com/wangzhenotes)** ，每隔一到两周我会站在算法工程师的角度讲解一些[计算广告](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E8%AE%A1%E7%AE%97%E5%B9%BF%E5%91%8A&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLorqHnrpflub_lkYoiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.R32lpRLpVqE8LiKPSFsShK7_V8J_ho51fjptxHyszD8&zhida_source=entity)、推荐系统相关的文章。选择文章必须满足一下三个条件：\n * **一是工程导向的；**\n * **二是阿里、facebook、google等一线互联网公司出品的；**\n * **三是前沿或者经典的。**\n\n\n这周我们一起讨论一下Youtube的深度推荐系统论文《[Deep Neural Networks for YouTube Recommendations](https://link.zhihu.com/?target=https%3A//github.com/wzhe06/Reco-papers/blob/master/Recommendation/Deep%2520Neural%2520Networks%2520for%2520YouTube%2520Recommendations.pdf)》，这是2016年的论文，按照今天的标准来看，已经没有什么新颖的地方，我也是两年前读过这篇文章之后就放下了，但前几天重读这篇文章，竟让发现了诸多亮点，几乎处处是套路，处处是经验，不由惊为神文。这篇神文给我留下的深刻印象有两点：\n 1. 这毫无疑问是**工业界论文的典范** ，是我非常推崇的工程导向的，[算法](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=2&q=%E7%AE%97%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnrpfms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoyLCJ6ZF90b2tlbiI6bnVsbH0.NMoDf0BxlY8jt6odGqCYrnDQqjQG8G9_Iv_pW1p39LI&zhida_source=entity)工程师必读的文章；\n 2. 我以为毫不起眼的地方，也藏着Youtube工程师宝贵的工程经验，相比上周介绍的[阿里的深度兴趣网络DIN](https://zhuanlan.zhihu.com/p/51623339)，最重要的价值就在于[Attention机制](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=Attention%E6%9C%BA%E5%88%B6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJBdHRlbnRpb27mnLrliLYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.DSTytNVLeXzD9VdB7GlIZVkzsRUgE-xEA_B0tuInQNM&zhida_source=entity)，**这篇文章你应该精确到句子来体会** ，这是我惊为神文的原因。\n\n\n \n\n废话不多说，下面就跟大家分享一下两次拜读这篇论文的不同体验和收获。\n第一遍读这篇论文的时候，我想所有人都是冲着算法的架构去的，在深度学习推荐系统已经成为各大公司“基本操作”的今天，Youtube在算法架构上并无惊奇之处，我们来快速介绍一下文章中的深度学习推荐系统的算法架构。\n![](../images/004.jpg)\n \n\nYoutube的用户推荐场景自不必多说，作为全球最大的UGC的视频网站，需要在百万量级的视频规模下进行个性化推荐。由于候选视频集合过大，考虑online系统延迟问题，不宜用复杂网络直接进行推荐，所以Youtube采取了两层[深度网络](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%B7%B1%E5%BA%A6%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLmt7HluqbnvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.ulqOvK9k70NnS5qF8UluETvDYo31fFOzteX1eueImCk&zhida_source=entity)完成整个推荐过程：\n 1. 第一层是**Candidate Generation Model** 完成候选视频的快速筛选，这一步候选视频集合由百万降低到了百的量级。\n 2. 第二层是用**Ranking Model** 完成几百个候选视频的精排\n\n\n首先介绍[candidate](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=candidate&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJjYW5kaWRhdGUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.RbQinN4AxTvP83rpXdk1NC2FJEwfXqqHugnaZwUDPzI&zhida_source=entity) generation模型的架构\n![](../images/005.jpg)\nYoutube Candidate Generation Model\n \n\n我们自底而上看这个网络，最底层的输入是用户观看过的video的embedding向量，以及搜索词的embedding向量。至于这个[embedding向量](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=3&q=embedding%E5%90%91%E9%87%8F&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJlbWJlZGRpbmflkJHph48iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjozLCJ6ZF90b2tlbiI6bnVsbH0.oTgCn2tvoshy_V8dedW3-c-y5iCQFUodGnZxq7NFxm4&zhida_source=entity)是怎么生成的，作者的原话是这样的 \n> Inspired by continuous bag of words language models, we learn high dimensional embeddings for each video in a xed vocabulary and feed these embeddings into a [feedforward](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=feedforward&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJmZWVkZm9yd2FyZCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjEwNjUzODQ4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.1SxefkTvTkin-cTjHV6-ImzjoTIzZ39OAB2pS5f-anU&zhida_source=entity) neural network\n所以作者是先用[word2vec](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=word2vec&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJ3b3JkMnZlYyIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjEwNjUzODQ4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.na38Qz-ba3BB3ylNewLUUxhCIX8gdNK6uZBzT-M1u5o&zhida_source=entity)方法对video和search token做了embedding之后再作为输入的，这也是做embedding的“基本操作”，不用过多介绍；当然，除此之外另一种大家应该也比较熟悉，就是通过加一个[embedding层](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=embedding%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJlbWJlZGRpbmflsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.9LFhmf9ZrkvEarNKdPlyeJZ6iofZ3SiEFfxpL3q4Aqw&zhida_source=entity)跟上面的[DNN](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=DNN&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJETk4iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.mHSlAs281_x8GR24icjVBMqwomGzvu71YgnWR3YgN20&zhida_source=entity)一起训练，两种方法孰优孰劣，有什么适用场合，大家可以讨论一下。\n \n\n[特征向量](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%89%B9%E5%BE%81%E5%90%91%E9%87%8F&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnibnlvoHlkJHph48iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.7w-ZzgYQRSbOlMxsXlV0mBexSnmofP2P1JS3MP1bA94&zhida_source=entity)里面还包括了用户的地理位置的embedding，年龄，性别等。然后把所有这些特征concatenate起来，喂给上层的ReLU[神经网络](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnpZ7nu4_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.tMWwdT2mRoaU4uu0KkBZr0cvv2xlURqim6n8qsS8Acc&zhida_source=entity)。\n三层神经网络过后，我们看到了softmax函数。这里Youtube的同学们把这个问题看作为用户推荐next watch的问题，所以输出应该是一个在所有candidate video上的[概率分布](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%A6%82%E7%8E%87%E5%88%86%E5%B8%83&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLmpoLnjofliIbluIMiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.Rf73Cl_lkjrcWJFmhPJMqltnRf-ll0z1VQQLyUAiIcg&zhida_source=entity)，自然是一个多分类问题。\n好了，这一套深度学习的“基本操作”下来，就构成了Youtube的candidate generation网络，看似平淡无奇，其实还是隐藏着一些问题的，比如\n 1. [架构图](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%9E%B6%E6%9E%84%E5%9B%BE&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLmnrbmnoTlm74iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.6_ljl8wiK7Bp__A6WsAeebBbBlxya-f_1OnfR1bD2_Y&zhida_source=entity)的左上角，为什么在online serving的时候不直接用这套网络进行预测而要使用[nearest neighbor search](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=nearest+neighbor+search&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJuZWFyZXN0IG5laWdoYm9yIHNlYXJjaCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjEwNjUzODQ4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.C30ZnQGgp1VxdLba54DwOLtO-ELmS1wSSGZjCt8gWp4&zhida_source=entity) 的方法？\n 2. 多分类问题中，Youtube的candidate video有百万之巨，意味着有几百万个分类，这必然会影响训练效果和速度，如何改进？\n\n\n这些问题在读第一遍的时候我也没有深想深看，但却是工程实现中必然会遇到的问题，我们随后再深入介绍论文中的解决方法。\n既然得到了几百个候选集合，下一步就是利用ranking模型进行精排序，下面是ranking深度学习网络的架构图。\n![](../images/006.jpg)\nYoutube Ranking Model\n \n\n乍一看上面的ranking model似乎与candidate generation模型没有什么区别，模型架构还是深度学习的“基本操作”，唯一的区别就是[特征工程](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%89%B9%E5%BE%81%E5%B7%A5%E7%A8%8B&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnibnlvoHlt6XnqIsiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.-WtZHpLE3T_k7CV5DkMS-v09PpUzJnlv7EvZGiCRg-A&zhida_source=entity)，那么我们就讲讲特征工程。\n事实上原文也明确说明了，引入另一套DNN作为ranking model的目的就是引入更多描述视频、用户以及二者之间关系的特征，达到对候选视频集合准确排序的目的。\n> During ranking, we have access to many more features describing the video and the user's relationship to the video because only a few hundred videos are being scored rather than the millions scored in candidate generation.\n具体一点，从左至右的特征依次是\n 1. **impression video ID embedding** : 当前要计算的video的embedding\n 2. **watched video IDs average embedding** : 用户观看过的最后N个视频embedding的average pooling\n 3. **language embedding** : 用户语言的embedding和当前视频语言的embedding\n 4. **time since last watch** : 自上次观看同channel视频的时间\n 5. **#previous impressions** : 该视频已经被曝光给该用户的次数\n\n\n上面五个特征中，我想重点谈谈第4个和第5个。因为这两个很好的引入了对用户行为的观察。\n第4个特征背后的思想是\n> We observe that the most important signals are those that describe a user's previous interaction with the item itself and other similar items.\n有一些引入attention的意思，这里是用了**time since last watch** 这个特征来反应用户看同类视频的间隔时间。从用户的角度想一想，假如我们刚看过“DOTA经典回顾”这个channel的视频，我们很大概率是会继续看这个channel的视频的，那么该特征就很好的捕捉到了这一用户行为。\n第5个特征**#previous impressions** 则一定程度上引入了exploration的思想，避免同一个视频持续对同一用户进行无效曝光。尽量增加用户没看过的新视频的曝光可能性。\n \n\n至此，我的第一遍论文阅读就结束了，对Youtube的算法框架有了概念，但总觉得不过如此，没什么太多新颖的地方。。\n但如果真这么想，还是太naive了，与上一篇**[阿里的深度兴趣网络DIN](https://zhuanlan.zhihu.com/p/51623339)** 不同的是，你读懂了DIN的[attention机制](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=attention%E6%9C%BA%E5%88%B6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJhdHRlbnRpb27mnLrliLYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.ht-ebaQ_5WyGHRSQX7LdssGv7PzqBS8nbzfFOUi3vYc&zhida_source=entity)，你就抓住了其论文70%的价值，但这篇文章，**如果你只读懂了Youtube的推荐[系统架构](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%B3%BB%E7%BB%9F%E6%9E%B6%E6%9E%84&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLns7vnu5_mnrbmnoQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.u4Xw1VtDd40LKPG-2vab8PkvI7k7QXfa6sb3ofpcOFQ&zhida_source=entity)，你只抓住了30%的价值**。那么剩下的70%的价值在哪里呢？\n在重读这篇文章的时候，我从一个工程师的角度，始终绷着“如何实现”这根弦，发现这篇论文的[工程价值](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E5%B7%A5%E7%A8%8B%E4%BB%B7%E5%80%BC&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLlt6XnqIvku7flgLwiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.aktJjiUPqgbWXBThF0QBXehpaXhRk86GUHG1hwX9WsQ&zhida_source=entity)之前被我大大忽略了。下面我列出十个文中解决的非常有价值的问题：\n 1. 文中把推荐问题转换成多分类问题，在next watch的场景下，每一个备选video都会是一个分类，因此总共的分类有数百万之巨，这在使用softmax训练时无疑是低效的，这个问题Youtube是如何解决的？\n 2. 在candidate generation model的serving过程中，Youtube为什么不直接采用训练时的model进行预测，而是采用了一种[最近邻](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%9C%80%E8%BF%91%E9%82%BB&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLmnIDov5HpgrsiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.yUcya9YkISWlMkvqyi2jR5SN9yQ0yZ4vVgILBgSecAc&zhida_source=entity)搜索的方法？\n 3. Youtube的用户对新视频有偏好，那么在模型构建的过程中如何引入这个feature？\n 4. 在对训练集的预处理过程中，Youtube没有采用原始的用户日志，而是对每个用户提取等数量的训练样本，这是为什么？\n 5. Youtube为什么不采取类似RNN的Sequence model，而是完全摒弃了用户观看历史的[时序特征](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%97%B6%E5%BA%8F%E7%89%B9%E5%BE%81&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLml7bluo_nibnlvoEiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.BpMMoAeP3dr5kGLH6rvV1HS4mcqhLsctzguHGq1SLiM&zhida_source=entity)，把用户最近的浏览历史等同看待，这不会损失有效信息吗？\n 6. 在处理测试集的时候，Youtube为什么不采用经典的随机[留一法](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%95%99%E4%B8%80%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnlZnkuIDms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.Ezk5w_0lX57aQqKwbf0R7ztt3xQFf0WF8oLsgePpu3E&zhida_source=entity)（random holdout），而是一定要把用户最近的一次观看行为作为测试集？\n 7. 在确定优化目标的时候，Youtube为什么不采用经典的[CTR](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=CTR&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJDVFIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.ERW2BzRmRQlBTDuzUV2CUrgd9NM-DepnxKJU4X5PY-c&zhida_source=entity)，或者播放率（Play Rate），而是采用了每次曝光预期播放时间（expected watch time per impression）作为优化目标？\n 8. 在进行[video embedding](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=video+embedding&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJ2aWRlbyBlbWJlZGRpbmciLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.56UATeqf0_67_u-HtmsOhrEvQvJvkfn7hTInMf8NfoQ&zhida_source=entity)的时候，为什么要直接把大量长尾的video直接用[0向量](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=0%E5%90%91%E9%87%8F&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiIw5ZCR6YePIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MTA2NTM4NDgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.IpNz8BWmx97s9-Vs7Lr9sAeVj8WvDpgyKJZxfsjwe-k&zhida_source=entity)代替？\n 9. 针对某些特征，比如#previous impressions，为什么要进行[开方](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E5%BC%80%E6%96%B9&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLlvIDmlrkiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.J9Syt02-MKc2WMATujXtA9GuaeVsFwlT2GOLaM6x224&zhida_source=entity)和平方处理后，当作三个特征输入模型？\n 10. 为什么ranking model不采用经典的[logistic regression](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=logistic+regression&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJsb2dpc3RpYyByZWdyZXNzaW9uIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MTA2NTM4NDgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.aS7XaXmMQ8Az5qGHmxvDXZeLdrQmBkHKCeuPEJX1G0o&zhida_source=entity)当作输出层，而是采用了[weighted logistic regression](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=weighted+logistic+regression&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJ3ZWlnaHRlZCBsb2dpc3RpYyByZWdyZXNzaW9uIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MTA2NTM4NDgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.jHPi2PkUyrcNBLxyg1cLJ6cssbegQibbtHvgaJ4GExk&zhida_source=entity)？\n\n\n \n\n因为我也是在视频推荐领域工作，所以可以很负责任的说以上的十个问题都是非常有价值的。但今天一口气写到这里，感觉有点气力不足了。。大家如果感兴趣的话可以点个赞，我明天再详细分析一下以上十大问题的答案。\n上面问题的解答已经完成啦，大家请参考我的下一篇文章——\n[](https://zhuanlan.zhihu.com/p/52504407)\n好了，这里是**[王喆的机器学习笔记](https://zhuanlan.zhihu.com/wangzhenotes)** 的第二篇文章，水平有限，欢迎大家吐槽，批评，纠错。\n也欢迎大家关注我的同名**微信公众号 王喆的机器学习笔记** （**wangzhenotes** ），或者通过公众号加我的个人微信进行进一步交流讨论，谢谢。\n \n\n参考资料：\n 1. [Deep Neural Networks for YouTube Recommendation](https://link.zhihu.com/?target=https%3A//github.com/wzhe06/Reco-papers/blob/master/Recommendation/Deep%2520Neural%2520Networks%2520for%2520YouTube%2520Recommendations.pdf)\n 2. [Recommender System Paper List](https://link.zhihu.com/?target=https%3A//github.com/wzhe06/Reco-papers)\n 3. [推荐系统中的注意力机制——阿里深度兴趣网络（DIN）](https://zhuanlan.zhihu.com/p/51623339)\n\n\n本文亦收录在我的新书「**[深度学习推荐系统](https://link.zhihu.com/?target=https%3A//union-click.jd.com/jdc%3Fe%3Djdext-1225677006157246464-0%26p%3DAyIGZRprFQEVBVQYWh0yVlgNRQQlW1dCFFlQCxxKQgFHRE5XDVULR0UVARUFVBhaHR1LQglGa0tDQm8QBR5sYE1PBlI9FnsVYjxoI3UOHjdVElsXChMGVRxYJQITBlUfXhYBFAZlK1sQMkRpVRpaFAMTAlUeWCUDIgdREl8UAxcHUxpZEwQiB10fa8OshtLdpIKlp8St%252B87WpdebrIGl3SUyIgRlQA5BS1cbVRhYEQUUAVQaXBUGFARQHF4SBw4GZRtYEwUaaVR1WBQHFg9RHF0RAiIHZRtrSkZPWmUZWhQAEA%253D%253D)** 」中，这本书系统性地整理、介绍了专栏中所有的重点内容，如果您曾在「**[王喆的机器学习笔记](https://zhuanlan.zhihu.com/wangzhenotes)** 」中受益，欢迎订阅，谢谢知友们的支持！\n[ ![](../images/007.jpg)广告 深度学习推荐系统 京东 ¥54.00 去购买​ ](https://union-click.jd.com/jdc?e=jdext-1225685139176988672-0&p=AyIGZRprFQEVBVQYWh0yVlgNRQQlW1dCFFlQCxxKQgFHRE5XDVULR0UVARUFVBhaHR1LQglGa3J1U3EmUC92YWd9Il0cRWV7ZgNJHGUOHjdVElsXChMGVRxYJQITBlUfXhYBFAZlK1sQMkRpVRpaFAMTAlUeWCUDIgdREl8UAxUHVBNZEAIiB10fa8OshtLdpIKlp8St%2B87WpdebrIGl3SUyIgRlQA5BS1cbVRhYEQUbA1UZUxUEFQ9cElwTAQ4GZRtYEwUaaVR1WBQHFg9RHF0RAiIHZRtrSkZPWmUZWhQAEA%3D%3D)\n编辑于 2020-03-23 23:28\n[机器学习](https://www.zhihu.com/topic/19559450)\n[推荐系统](https://www.zhihu.com/topic/19563024)\n[深度学习（Deep Learning）](https://www.zhihu.com/topic/19813032)\n​赞同 3252​​140 条评论​7333 ​181 \n​分享\n​申请转载​\n​\n![](../images/body-7e064b35ba4350074989.jpeg)\n未登录用户\n \n\n网络故障，请检查网络连接\n重新加载\n关于作者\n[![王喆](../images/008.jpg)](https://www.zhihu.com/people/wang-zhe-58)\n[王喆](https://www.zhihu.com/people/wang-zhe-58)\n广告/推荐 Engineering Director\n[​](https://zhuanlan.zhihu.com/p/96956163)\n科技作者，代表作《深度学习推荐系统》《理工男谈理财》\n​\n[数据挖掘](https://www.zhihu.com/people/wang-zhe-58/creations/19553534)、[机器学习](https://www.zhihu.com/people/wang-zhe-58/creations/19559450)、[深度学习（Deep Learning）](https://www.zhihu.com/people/wang-zhe-58/creations/19813032) 话题的优秀答主\n[回答](https://www.zhihu.com/people/wang-zhe-58/answers)[文章](https://www.zhihu.com/people/wang-zhe-58/posts)[关注者](https://www.zhihu.com/people/wang-zhe-58/followers)\n​关注他​发私信\n大家都在搜\n换一换\n[原央视主持人阿丘回应被通报464 万](https://www.zhihu.com/search?q=%E5%8E%9F%E5%A4%AE%E8%A7%86%E4%B8%BB%E6%8C%81%E4%BA%BA%E9%98%BF%E4%B8%98%E5%9B%9E%E5%BA%94%E8%A2%AB%E9%80%9A%E6%8A%A5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)热\n[中国男足 0-5 巴勒斯坦310 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E7%94%B7%E8%B6%B3+0-5+%E5%B7%B4%E5%8B%92%E6%96%AF%E5%9D%A6&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)热\n[文春曝张本智和私生活292 万](https://www.zhihu.com/search?q=%E6%96%87%E6%98%A5%E6%9B%9D%E5%BC%A0%E6%9C%AC%E6%99%BA%E5%92%8C%E7%A7%81%E7%94%9F%E6%B4%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[孕妇骑车别车被司机踹翻279 万](https://www.zhihu.com/search?q=%E5%AD%95%E5%A6%87%E9%AA%91%E8%BD%A6%E5%88%AB%E8%BD%A6%E8%A2%AB%E5%8F%B8%E6%9C%BA%E8%B8%B9%E7%BF%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)热\n[国庆多地高速服务区电车取号排队充电262 万](https://www.zhihu.com/search?q=%E5%9B%BD%E5%BA%86%E5%A4%9A%E5%9C%B0%E9%AB%98%E9%80%9F%E6%9C%8D%E5%8A%A1%E5%8C%BA%E7%94%B5%E8%BD%A6%E5%8F%96%E5%8F%B7%E6%8E%92%E9%98%9F%E5%85%85%E7%94%B5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[韩国检察厅被撤销258 万](https://www.zhihu.com/search?q=%E9%9F%A9%E5%9B%BD%E6%A3%80%E5%AF%9F%E5%8E%85%E8%A2%AB%E6%92%A4%E9%94%80&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[亚运男足中国 U23 vs 乌兹别克斯坦 U23245 万](https://www.zhihu.com/search?q=%E4%BA%9A%E8%BF%90%E7%94%B7%E8%B6%B3%E4%B8%AD%E5%9B%BD+U23+vs+%E4%B9%8C%E5%85%B9%E5%88%AB%E5%85%8B%E6%96%AF%E5%9D%A6+U23&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)新\n[葡萄牙首次在无C罗情况下打进4球231 万](https://www.zhihu.com/search?q=%E8%91%A1%E8%90%84%E7%89%99%E9%A6%96%E6%AC%A1%E5%9C%A8%E6%97%A0C%E7%BD%97%E6%83%85%E5%86%B5%E4%B8%8B%E6%89%93%E8%BF%9B4%E7%90%83&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[C罗官宣离开国家队集训211 万](https://www.zhihu.com/search?q=C%E7%BD%97%E5%AE%98%E5%AE%A3%E7%A6%BB%E5%BC%80%E5%9B%BD%E5%AE%B6%E9%98%9F%E9%9B%86%E8%AE%AD&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[中国女足 0-1 韩国202 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E5%A5%B3%E8%B6%B3+0-1+%E9%9F%A9%E5%9B%BD&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n_想来知乎工作？请发送邮件到 jobs@zhihu.com_\n打开知乎App\n在「我的页」右上角打开扫一扫\n![](../images/body-0f9788d006ab0f4da762.png)\n其他扫码方式：微信\n下载知乎App\n无障碍模式\n验证码登录\n密码登录\n[开通机构号](https://zhuanlan.zhihu.com/org/signup)\n中国 +86\n获取短信验证码\n忘记密码\n登录/注册\n请您登录后查看更多专业优质内容。\n其他方式登录\n未注册手机验证后自动登录，注册即代表同意[《知乎协议》](https://www.zhihu.com/term/zhihu-terms)[《隐私保护指引》](https://www.zhihu.com/term/privacy)\n扫码下载知乎 App\n关闭二维码\n请您登录后查看更多专业优质内容。\n请您登录后查看更多专业优质内容。\n", "metadata": { "title": "重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文 - 知乎", "description": "这里是 王喆的机器学习笔记，每隔一到两周我会站在算法工程师的角度讲解一些计算广告、推荐系统相关的文章。选择文章必须满足一下三个条件：一是工程导向的；二是阿里、facebook、google等一线互联网公司出品的；…", "keywords": "机器学习,推荐系统,深度学习（Deep Learning）", "author": null, "og:title": "重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文", "og:url": "https://zhuanlan.zhihu.com/p/52169807", "og:description": "这里是 王喆的机器学习笔记，每隔一到两周我会站在算法工程师的角度讲解一些计算广告、推荐系统相关的文章。选择文章必须满足一下三个条件：一是工程导向的；二是阿里、facebook、google等一线互联网公司出品的；…", "og:image": "https://pica.zhimg.com/v2-4dd8f94f0cf5becd7dd5ff4c0fd9393f_720w.jpg?source=172ae18b", "og:type": "article", "og:site_name": "知乎专栏" }, "paywall": { "listed_paid": false, "detected": false, "marker": "", "public_only": false, "policy": "付费/会员内容只保留页面公开可见部分，不绕过付费墙" }, "crawl": { "success": true, "status_code": 200, "error": "", "retried": true }, "code_blocks": [], "images": [ { "status": "http_403", "url": "https://zhuanlan.zhihu.com/p/52169807", "src": "https://zhuanlan.zhihu.com/p/52169807", "alt": "", "downloaded": false, "discovery": "saved_markdown", "retried_at": "2026-10-20T17:59:52Z" } ], "audio": [], "videos": [], "links": [ { "type": "internal", "href": "https://www.zhihu.com" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Ffollow" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2F" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fhot" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fcolumn-square" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fring-feeds" }, { "type": "internal", "href": "https://www.zhihu.com/project-square" }, { "type": "internal", "href": "https://www.zhihu.com/fiore/h5/vip-web" }, { "type": "internal", "href": "https://zhida.zhihu.com/" }, { "type": "internal", "href": "https://www.zhihu.com/column/wangzhenotes" }, { "type": "internal", "href": "https://www.zhihu.com/people/wang-zhe-58" }, { "type": "internal", "href": "https://www.zhihu.com/question/48509984" }, { "type": "internal", "href": "https://zhuanlan.zhihu.com/wangzhenotes" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%AE%97%E6%B3%95%E5%B7%A5%E7%A8%8B%E5%B8%88&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkwNzQsInEiOiLnrpfms5Xlt6XnqIvluIgiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.ZGs_CAF5__BFyzYJrnUprIwG8zglnIcY5F9jsv1cUXA&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E8%AE%A1%E7%AE%97%E5%B9%BF%E5%91%8A&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkwNzQsInEiOiLorqHnrpflub_lkYoiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.lk8BPQwibjP4WZBuH74Y4xEias5WBIzAUWEzyGSkyTk&zhida_source=entity" }, { "type": "internal", "href": "https://link.zhihu.com/?target=https%3A%2F%2Fgithub.com%2Fwzhe06%2FReco-papers%2Fblob%2Fmaster%2FRecommendation%2FDeep%2520Neural%2520Networks%2520for%2520YouTube%2520Recommendations.pdf" }, { "type": "internal", "href": "https://zhuanlan.zhihu.com/p/51623339" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=Attention%E6%9C%BA%E5%88%B6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkwNzQsInEiOiJBdHRlbnRpb27mnLrliLYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.o0ZR-XaewNAFUmcNg3clSGF0_3M8Af01LgabMXf1Ggg&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%B7%B1%E5%BA%A6%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkwNzQsInEiOiLmt7HluqbnvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.JDG5NscijHG_d7AkbvG4axfMbEz6ZOgh6PETY-SR9bE&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=candidate&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkwNzQsInEiOiJjYW5kaWRhdGUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.tuhDL2dOWRkPGKS71YECdU04US5GwBIJQzfM6oP3Axg&zhida_source=entity" }, { "type": "external", "href": "javascript:void(0)" }, { "type": "external", "href": "https://union-click.jd.com/jdc?e=jdext-1225685139176988672-0&p=AyIGZRprFQEVBVQYWh0yVlgNRQQlW1dCFFlQCxxKQgFHRE5XDVULR0UVARUFVBhaHR1LQglGa3J1U3EmUC92YWd9Il0cRWV7ZgNJHGUOHjdVElsXChMGVRxYJQITBlUfXhYBFAZlK1sQMkRpVRpaFAMTAlUeWCUDIgdREl8UAxUHVBNZEAIiB10fa8OshtLdpIKlp8St%2B87WpdebrIGl3SUyIgRlQA5BS1cbVRhYEQUbA1UZUxUEFQ9cElwTAQ4GZRtYEwUaaVR1WBQHFg9RHF0RAiIHZRtrSkZPWmUZWhQAEA%3D%3D" } ], "retry_at": "2026-10-03T08:12:41", "repair_version": 2, "repaired_at": "2026-10-20T17:30:12Z", "source_url": "https://zhuanlan.zhihu.com/p/52169807", "delivery_status": "unverified", "discovery_status": "http_403", "missing_reason": "源页面未重新核验:http_403；已按保留正文补采图片", "repair_error": "", "actual_video_files": [], "expected_saved_body_images": 1, "image_discovery": "saved_markdown", "source_body_markdown": "[![](https://zhuanlan.zhihu.com/p/52169807)](javascript:void\\(0\\))\n![ZhiHu logo](https://static.zhihu.com/heifetz/assets/wechat-share-logo.39ea9ecd.png)\n[](https://www.zhihu.com)\n[关注](https://www.zhihu.com/signin?next=%2Ffollow)[推荐](https://www.zhihu.com/signin?next=%2F)[热榜](https://www.zhihu.com/signin?next=%2Fhot)[专栏](https://www.zhihu.com/signin?next=%2Fcolumn-square)[圈子](https://www.zhihu.com/signin?next=%2Fring-feeds)[AI WorksBeta ](https://www.zhihu.com/project-square)[故事](https://www.zhihu.com/fiore/h5/vip-web)\n​\n[直答](https://zhida.zhihu.com/)\n切换模式\n登录/注册\n重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文\n首发于[王喆的机器学习笔记](https://www.zhihu.com/column/wangzhenotes)\n切换模式\n请您登录后查看更多专业优质内容。\n![重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文](https://pica.zhimg.com/70/v2-4dd8f94f0cf5becd7dd5ff4c0fd9393f_1440w.image?source=172ae18b&biz_tag=Post)\n# 重读Youtube深度学习推荐系统论文，字字珠玑，惊为神文\n[![王喆](https://pica.zhimg.com/08f032fe9b5e5cdf1b435321f1dbc0ec_l.jpg?source=172ae18b)](https://www.zhihu.com/people/wang-zhe-58)\n[王喆](https://www.zhihu.com/people/wang-zhe-58)\n[​![](https://picx.zhimg.com/v2-27bfcba90e66db79ce8768ab807e017e_l.png?source=32738c0c)](https://www.zhihu.com/question/48509984)\n数据挖掘等 3 个话题下的优秀答主\n[收录于 · 王喆的机器学习笔记](https://www.zhihu.com/column/wangzhenotes)\n3252 人赞同了该文章\n这里是**[王喆的机器学习笔记](https://zhuanlan.zhihu.com/wangzhenotes)** ，每隔一到两周我会站在算法工程师的角度讲解一些[计算广告](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E8%AE%A1%E7%AE%97%E5%B9%BF%E5%91%8A&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLorqHnrpflub_lkYoiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.R32lpRLpVqE8LiKPSFsShK7_V8J_ho51fjptxHyszD8&zhida_source=entity)、推荐系统相关的文章。选择文章必须满足一下三个条件：\n * **一是工程导向的；**\n * **二是阿里、facebook、google等一线互联网公司出品的；**\n * **三是前沿或者经典的。**\n\n\n这周我们一起讨论一下Youtube的深度推荐系统论文《[Deep Neural Networks for YouTube Recommendations](https://link.zhihu.com/?target=https%3A//github.com/wzhe06/Reco-papers/blob/master/Recommendation/Deep%2520Neural%2520Networks%2520for%2520YouTube%2520Recommendations.pdf)》，这是2016年的论文，按照今天的标准来看，已经没有什么新颖的地方，我也是两年前读过这篇文章之后就放下了，但前几天重读这篇文章，竟让发现了诸多亮点，几乎处处是套路，处处是经验，不由惊为神文。这篇神文给我留下的深刻印象有两点：\n 1. 这毫无疑问是**工业界论文的典范** ，是我非常推崇的工程导向的，[算法](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=2&q=%E7%AE%97%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnrpfms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoyLCJ6ZF90b2tlbiI6bnVsbH0.NMoDf0BxlY8jt6odGqCYrnDQqjQG8G9_Iv_pW1p39LI&zhida_source=entity)工程师必读的文章；\n 2. 我以为毫不起眼的地方，也藏着Youtube工程师宝贵的工程经验，相比上周介绍的[阿里的深度兴趣网络DIN](https://zhuanlan.zhihu.com/p/51623339)，最重要的价值就在于[Attention机制](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=Attention%E6%9C%BA%E5%88%B6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJBdHRlbnRpb27mnLrliLYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.DSTytNVLeXzD9VdB7GlIZVkzsRUgE-xEA_B0tuInQNM&zhida_source=entity)，**这篇文章你应该精确到句子来体会** ，这是我惊为神文的原因。\n\n\n \n\n废话不多说，下面就跟大家分享一下两次拜读这篇论文的不同体验和收获。\n第一遍读这篇论文的时候，我想所有人都是冲着算法的架构去的，在深度学习推荐系统已经成为各大公司“基本操作”的今天，Youtube在算法架构上并无惊奇之处，我们来快速介绍一下文章中的深度学习推荐系统的算法架构。\n![](https://pic2.zhimg.com/v2-dd9c1428a304330361331a924f2c39fd_1440w.jpg)\n \n\nYoutube的用户推荐场景自不必多说，作为全球最大的UGC的视频网站，需要在百万量级的视频规模下进行个性化推荐。由于候选视频集合过大，考虑online系统延迟问题，不宜用复杂网络直接进行推荐，所以Youtube采取了两层[深度网络](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%B7%B1%E5%BA%A6%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLmt7HluqbnvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.ulqOvK9k70NnS5qF8UluETvDYo31fFOzteX1eueImCk&zhida_source=entity)完成整个推荐过程：\n 1. 第一层是**Candidate Generation Model** 完成候选视频的快速筛选，这一步候选视频集合由百万降低到了百的量级。\n 2. 第二层是用**Ranking Model** 完成几百个候选视频的精排\n\n\n首先介绍[candidate](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=candidate&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJjYW5kaWRhdGUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.RbQinN4AxTvP83rpXdk1NC2FJEwfXqqHugnaZwUDPzI&zhida_source=entity) generation模型的架构\n![](https://pic2.zhimg.com/v2-b6998c31a1b29125d8da403b9a7b29bd_1440w.jpg)\nYoutube Candidate Generation Model\n \n\n我们自底而上看这个网络，最底层的输入是用户观看过的video的embedding向量，以及搜索词的embedding向量。至于这个[embedding向量](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=3&q=embedding%E5%90%91%E9%87%8F&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJlbWJlZGRpbmflkJHph48iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjozLCJ6ZF90b2tlbiI6bnVsbH0.oTgCn2tvoshy_V8dedW3-c-y5iCQFUodGnZxq7NFxm4&zhida_source=entity)是怎么生成的，作者的原话是这样的 \n> Inspired by continuous bag of words language models, we learn high dimensional embeddings for each video in a xed vocabulary and feed these embeddings into a [feedforward](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=feedforward&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJmZWVkZm9yd2FyZCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjEwNjUzODQ4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.1SxefkTvTkin-cTjHV6-ImzjoTIzZ39OAB2pS5f-anU&zhida_source=entity) neural network\n所以作者是先用[word2vec](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=word2vec&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJ3b3JkMnZlYyIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjEwNjUzODQ4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.na38Qz-ba3BB3ylNewLUUxhCIX8gdNK6uZBzT-M1u5o&zhida_source=entity)方法对video和search token做了embedding之后再作为输入的，这也是做embedding的“基本操作”，不用过多介绍；当然，除此之外另一种大家应该也比较熟悉，就是通过加一个[embedding层](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=embedding%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJlbWJlZGRpbmflsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.9LFhmf9ZrkvEarNKdPlyeJZ6iofZ3SiEFfxpL3q4Aqw&zhida_source=entity)跟上面的[DNN](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=DNN&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJETk4iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.mHSlAs281_x8GR24icjVBMqwomGzvu71YgnWR3YgN20&zhida_source=entity)一起训练，两种方法孰优孰劣，有什么适用场合，大家可以讨论一下。\n \n\n[特征向量](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%89%B9%E5%BE%81%E5%90%91%E9%87%8F&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnibnlvoHlkJHph48iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.7w-ZzgYQRSbOlMxsXlV0mBexSnmofP2P1JS3MP1bA94&zhida_source=entity)里面还包括了用户的地理位置的embedding，年龄，性别等。然后把所有这些特征concatenate起来，喂给上层的ReLU[神经网络](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnpZ7nu4_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.tMWwdT2mRoaU4uu0KkBZr0cvv2xlURqim6n8qsS8Acc&zhida_source=entity)。\n三层神经网络过后，我们看到了softmax函数。这里Youtube的同学们把这个问题看作为用户推荐next watch的问题，所以输出应该是一个在所有candidate video上的[概率分布](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%A6%82%E7%8E%87%E5%88%86%E5%B8%83&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLmpoLnjofliIbluIMiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.Rf73Cl_lkjrcWJFmhPJMqltnRf-ll0z1VQQLyUAiIcg&zhida_source=entity)，自然是一个多分类问题。\n好了，这一套深度学习的“基本操作”下来，就构成了Youtube的candidate generation网络，看似平淡无奇，其实还是隐藏着一些问题的，比如\n 1. [架构图](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%9E%B6%E6%9E%84%E5%9B%BE&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLmnrbmnoTlm74iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.6_ljl8wiK7Bp__A6WsAeebBbBlxya-f_1OnfR1bD2_Y&zhida_source=entity)的左上角，为什么在online serving的时候不直接用这套网络进行预测而要使用[nearest neighbor search](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=nearest+neighbor+search&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJuZWFyZXN0IG5laWdoYm9yIHNlYXJjaCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjEwNjUzODQ4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.C30ZnQGgp1VxdLba54DwOLtO-ELmS1wSSGZjCt8gWp4&zhida_source=entity) 的方法？\n 2. 多分类问题中，Youtube的candidate video有百万之巨，意味着有几百万个分类，这必然会影响训练效果和速度，如何改进？\n\n\n这些问题在读第一遍的时候我也没有深想深看，但却是工程实现中必然会遇到的问题，我们随后再深入介绍论文中的解决方法。\n既然得到了几百个候选集合，下一步就是利用ranking模型进行精排序，下面是ranking深度学习网络的架构图。\n![](https://pic3.zhimg.com/v2-bd824fe6c1f5beb1d1eb896955734fdc_1440w.jpg)\nYoutube Ranking Model\n \n\n乍一看上面的ranking model似乎与candidate generation模型没有什么区别，模型架构还是深度学习的“基本操作”，唯一的区别就是[特征工程](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%89%B9%E5%BE%81%E5%B7%A5%E7%A8%8B&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnibnlvoHlt6XnqIsiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.-WtZHpLE3T_k7CV5DkMS-v09PpUzJnlv7EvZGiCRg-A&zhida_source=entity)，那么我们就讲讲特征工程。\n事实上原文也明确说明了，引入另一套DNN作为ranking model的目的就是引入更多描述视频、用户以及二者之间关系的特征，达到对候选视频集合准确排序的目的。\n> During ranking, we have access to many more features describing the video and the user's relationship to the video because only a few hundred videos are being scored rather than the millions scored in candidate generation.\n具体一点，从左至右的特征依次是\n 1. **impression video ID embedding** : 当前要计算的video的embedding\n 2. **watched video IDs average embedding** : 用户观看过的最后N个视频embedding的average pooling\n 3. **language embedding** : 用户语言的embedding和当前视频语言的embedding\n 4. **time since last watch** : 自上次观看同channel视频的时间\n 5. **#previous impressions** : 该视频已经被曝光给该用户的次数\n\n\n上面五个特征中，我想重点谈谈第4个和第5个。因为这两个很好的引入了对用户行为的观察。\n第4个特征背后的思想是\n> We observe that the most important signals are those that describe a user's previous interaction with the item itself and other similar items.\n有一些引入attention的意思，这里是用了**time since last watch** 这个特征来反应用户看同类视频的间隔时间。从用户的角度想一想，假如我们刚看过“DOTA经典回顾”这个channel的视频，我们很大概率是会继续看这个channel的视频的，那么该特征就很好的捕捉到了这一用户行为。\n第5个特征**#previous impressions** 则一定程度上引入了exploration的思想，避免同一个视频持续对同一用户进行无效曝光。尽量增加用户没看过的新视频的曝光可能性。\n \n\n至此，我的第一遍论文阅读就结束了，对Youtube的算法框架有了概念，但总觉得不过如此，没什么太多新颖的地方。。\n但如果真这么想，还是太naive了，与上一篇**[阿里的深度兴趣网络DIN](https://zhuanlan.zhihu.com/p/51623339)** 不同的是，你读懂了DIN的[attention机制](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=attention%E6%9C%BA%E5%88%B6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJhdHRlbnRpb27mnLrliLYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.ht-ebaQ_5WyGHRSQX7LdssGv7PzqBS8nbzfFOUi3vYc&zhida_source=entity)，你就抓住了其论文70%的价值，但这篇文章，**如果你只读懂了Youtube的推荐[系统架构](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%B3%BB%E7%BB%9F%E6%9E%B6%E6%9E%84&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLns7vnu5_mnrbmnoQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.u4Xw1VtDd40LKPG-2vab8PkvI7k7QXfa6sb3ofpcOFQ&zhida_source=entity)，你只抓住了30%的价值**。那么剩下的70%的价值在哪里呢？\n在重读这篇文章的时候，我从一个工程师的角度，始终绷着“如何实现”这根弦，发现这篇论文的[工程价值](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E5%B7%A5%E7%A8%8B%E4%BB%B7%E5%80%BC&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLlt6XnqIvku7flgLwiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.aktJjiUPqgbWXBThF0QBXehpaXhRk86GUHG1hwX9WsQ&zhida_source=entity)之前被我大大忽略了。下面我列出十个文中解决的非常有价值的问题：\n 1. 文中把推荐问题转换成多分类问题，在next watch的场景下，每一个备选video都会是一个分类，因此总共的分类有数百万之巨，这在使用softmax训练时无疑是低效的，这个问题Youtube是如何解决的？\n 2. 在candidate generation model的serving过程中，Youtube为什么不直接采用训练时的model进行预测，而是采用了一种[最近邻](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%9C%80%E8%BF%91%E9%82%BB&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLmnIDov5HpgrsiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.yUcya9YkISWlMkvqyi2jR5SN9yQ0yZ4vVgILBgSecAc&zhida_source=entity)搜索的方法？\n 3. Youtube的用户对新视频有偏好，那么在模型构建的过程中如何引入这个feature？\n 4. 在对训练集的预处理过程中，Youtube没有采用原始的用户日志，而是对每个用户提取等数量的训练样本，这是为什么？\n 5. Youtube为什么不采取类似RNN的Sequence model，而是完全摒弃了用户观看历史的[时序特征](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E6%97%B6%E5%BA%8F%E7%89%B9%E5%BE%81&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLml7bluo_nibnlvoEiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.BpMMoAeP3dr5kGLH6rvV1HS4mcqhLsctzguHGq1SLiM&zhida_source=entity)，把用户最近的浏览历史等同看待，这不会损失有效信息吗？\n 6. 在处理测试集的时候，Youtube为什么不采用经典的随机[留一法](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E7%95%99%E4%B8%80%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLnlZnkuIDms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.Ezk5w_0lX57aQqKwbf0R7ztt3xQFf0WF8oLsgePpu3E&zhida_source=entity)（random holdout），而是一定要把用户最近的一次观看行为作为测试集？\n 7. 在确定优化目标的时候，Youtube为什么不采用经典的[CTR](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=CTR&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJDVFIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.ERW2BzRmRQlBTDuzUV2CUrgd9NM-DepnxKJU4X5PY-c&zhida_source=entity)，或者播放率（Play Rate），而是采用了每次曝光预期播放时间（expected watch time per impression）作为优化目标？\n 8. 在进行[video embedding](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=video+embedding&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJ2aWRlbyBlbWJlZGRpbmciLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.56UATeqf0_67_u-HtmsOhrEvQvJvkfn7hTInMf8NfoQ&zhida_source=entity)的时候，为什么要直接把大量长尾的video直接用[0向量](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=0%E5%90%91%E9%87%8F&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiIw5ZCR6YePIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MTA2NTM4NDgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.IpNz8BWmx97s9-Vs7Lr9sAeVj8WvDpgyKJZxfsjwe-k&zhida_source=entity)代替？\n 9. 针对某些特征，比如#previous impressions，为什么要进行[开方](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=%E5%BC%80%E6%96%B9&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiLlvIDmlrkiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoxMDY1Mzg0OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.J9Syt02-MKc2WMATujXtA9GuaeVsFwlT2GOLaM6x224&zhida_source=entity)和平方处理后，当作三个特征输入模型？\n 10. 为什么ranking model不采用经典的[logistic regression](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=logistic+regression&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJsb2dpc3RpYyByZWdyZXNzaW9uIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MTA2NTM4NDgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.aS7XaXmMQ8Az5qGHmxvDXZeLdrQmBkHKCeuPEJX1G0o&zhida_source=entity)当作输出层，而是采用了[weighted logistic regression](https://zhida.zhihu.com/search?content_id=10653848&content_type=Article&match_order=1&q=weighted+logistic+regression&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExODc5NjEsInEiOiJ3ZWlnaHRlZCBsb2dpc3RpYyByZWdyZXNzaW9uIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MTA2NTM4NDgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.jHPi2PkUyrcNBLxyg1cLJ6cssbegQibbtHvgaJ4GExk&zhida_source=entity)？\n\n\n \n\n因为我也是在视频推荐领域工作，所以可以很负责任的说以上的十个问题都是非常有价值的。但今天一口气写到这里，感觉有点气力不足了。。大家如果感兴趣的话可以点个赞，我明天再详细分析一下以上十大问题的答案。\n上面问题的解答已经完成啦，大家请参考我的下一篇文章——\n[](https://zhuanlan.zhihu.com/p/52504407)\n好了，这里是**[王喆的机器学习笔记](https://zhuanlan.zhihu.com/wangzhenotes)** 的第二篇文章，水平有限，欢迎大家吐槽，批评，纠错。\n也欢迎大家关注我的同名**微信公众号 王喆的机器学习笔记** （**wangzhenotes** ），或者通过公众号加我的个人微信进行进一步交流讨论，谢谢。\n \n\n参考资料：\n 1. [Deep Neural Networks for YouTube Recommendation](https://link.zhihu.com/?target=https%3A//github.com/wzhe06/Reco-papers/blob/master/Recommendation/Deep%2520Neural%2520Networks%2520for%2520YouTube%2520Recommendations.pdf)\n 2. [Recommender System Paper List](https://link.zhihu.com/?target=https%3A//github.com/wzhe06/Reco-papers)\n 3. [推荐系统中的注意力机制——阿里深度兴趣网络（DIN）](https://zhuanlan.zhihu.com/p/51623339)\n\n\n本文亦收录在我的新书「**[深度学习推荐系统](https://link.zhihu.com/?target=https%3A//union-click.jd.com/jdc%3Fe%3Djdext-1225677006157246464-0%26p%3DAyIGZRprFQEVBVQYWh0yVlgNRQQlW1dCFFlQCxxKQgFHRE5XDVULR0UVARUFVBhaHR1LQglGa0tDQm8QBR5sYE1PBlI9FnsVYjxoI3UOHjdVElsXChMGVRxYJQITBlUfXhYBFAZlK1sQMkRpVRpaFAMTAlUeWCUDIgdREl8UAxcHUxpZEwQiB10fa8OshtLdpIKlp8St%252B87WpdebrIGl3SUyIgRlQA5BS1cbVRhYEQUUAVQaXBUGFARQHF4SBw4GZRtYEwUaaVR1WBQHFg9RHF0RAiIHZRtrSkZPWmUZWhQAEA%253D%253D)** 」中，这本书系统性地整理、介绍了专栏中所有的重点内容，如果您曾在「**[王喆的机器学习笔记](https://zhuanlan.zhihu.com/wangzhenotes)** 」中受益，欢迎订阅，谢谢知友们的支持！\n[ ![](https://pic1.zhimg.com/v2-ead48e2a477d1a607231b807a58b94fb_720w.jpg?source=b555e01d)广告 深度学习推荐系统 京东 ¥54.00 去购买​ ](https://union-click.jd.com/jdc?e=jdext-1225685139176988672-0&p=AyIGZRprFQEVBVQYWh0yVlgNRQQlW1dCFFlQCxxKQgFHRE5XDVULR0UVARUFVBhaHR1LQglGa3J1U3EmUC92YWd9Il0cRWV7ZgNJHGUOHjdVElsXChMGVRxYJQITBlUfXhYBFAZlK1sQMkRpVRpaFAMTAlUeWCUDIgdREl8UAxUHVBNZEAIiB10fa8OshtLdpIKlp8St%2B87WpdebrIGl3SUyIgRlQA5BS1cbVRhYEQUbA1UZUxUEFQ9cElwTAQ4GZRtYEwUaaVR1WBQHFg9RHF0RAiIHZRtrSkZPWmUZWhQAEA%3D%3D)\n编辑于 2020-03-23 23:28\n[机器学习](https://www.zhihu.com/topic/19559450)\n[推荐系统](https://www.zhihu.com/topic/19563024)\n[深度学习（Deep Learning）](https://www.zhihu.com/topic/19813032)\n​赞同 3252​​140 条评论​7333 ​181 \n​分享\n​申请转载​\n​\n![](https://pic1.zhimg.com/v2-d41c2ceaed8f51999522f903672a521f_l.jpeg)\n未登录用户\n \n\n网络故障，请检查网络连接\n重新加载\n关于作者\n[![王喆](https://pica.zhimg.com/08f032fe9b5e5cdf1b435321f1dbc0ec_l.jpg?source=172ae18b)](https://www.zhihu.com/people/wang-zhe-58)\n[王喆](https://www.zhihu.com/people/wang-zhe-58)\n广告/推荐 Engineering Director\n[​](https://zhuanlan.zhihu.com/p/96956163)\n科技作者，代表作《深度学习推荐系统》《理工男谈理财》\n​\n[数据挖掘](https://www.zhihu.com/people/wang-zhe-58/creations/19553534)、[机器学习](https://www.zhihu.com/people/wang-zhe-58/creations/19559450)、[深度学习（Deep Learning）](https://www.zhihu.com/people/wang-zhe-58/creations/19813032) 话题的优秀答主\n[回答](https://www.zhihu.com/people/wang-zhe-58/answers)[文章](https://www.zhihu.com/people/wang-zhe-58/posts)[关注者](https://www.zhihu.com/people/wang-zhe-58/followers)\n​关注他​发私信\n大家都在搜\n换一换\n[原央视主持人阿丘回应被通报464 万](https://www.zhihu.com/search?q=%E5%8E%9F%E5%A4%AE%E8%A7%86%E4%B8%BB%E6%8C%81%E4%BA%BA%E9%98%BF%E4%B8%98%E5%9B%9E%E5%BA%94%E8%A2%AB%E9%80%9A%E6%8A%A5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)热\n[中国男足 0-5 巴勒斯坦310 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E7%94%B7%E8%B6%B3+0-5+%E5%B7%B4%E5%8B%92%E6%96%AF%E5%9D%A6&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)热\n[文春曝张本智和私生活292 万](https://www.zhihu.com/search?q=%E6%96%87%E6%98%A5%E6%9B%9D%E5%BC%A0%E6%9C%AC%E6%99%BA%E5%92%8C%E7%A7%81%E7%94%9F%E6%B4%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[孕妇骑车别车被司机踹翻279 万](https://www.zhihu.com/search?q=%E5%AD%95%E5%A6%87%E9%AA%91%E8%BD%A6%E5%88%AB%E8%BD%A6%E8%A2%AB%E5%8F%B8%E6%9C%BA%E8%B8%B9%E7%BF%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)热\n[国庆多地高速服务区电车取号排队充电262 万](https://www.zhihu.com/search?q=%E5%9B%BD%E5%BA%86%E5%A4%9A%E5%9C%B0%E9%AB%98%E9%80%9F%E6%9C%8D%E5%8A%A1%E5%8C%BA%E7%94%B5%E8%BD%A6%E5%8F%96%E5%8F%B7%E6%8E%92%E9%98%9F%E5%85%85%E7%94%B5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[韩国检察厅被撤销258 万](https://www.zhihu.com/search?q=%E9%9F%A9%E5%9B%BD%E6%A3%80%E5%AF%9F%E5%8E%85%E8%A2%AB%E6%92%A4%E9%94%80&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[亚运男足中国 U23 vs 乌兹别克斯坦 U23245 万](https://www.zhihu.com/search?q=%E4%BA%9A%E8%BF%90%E7%94%B7%E8%B6%B3%E4%B8%AD%E5%9B%BD+U23+vs+%E4%B9%8C%E5%85%B9%E5%88%AB%E5%85%8B%E6%96%AF%E5%9D%A6+U23&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)新\n[葡萄牙首次在无C罗情况下打进4球231 万](https://www.zhihu.com/search?q=%E8%91%A1%E8%90%84%E7%89%99%E9%A6%96%E6%AC%A1%E5%9C%A8%E6%97%A0C%E7%BD%97%E6%83%85%E5%86%B5%E4%B8%8B%E6%89%93%E8%BF%9B4%E7%90%83&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[C罗官宣离开国家队集训211 万](https://www.zhihu.com/search?q=C%E7%BD%97%E5%AE%98%E5%AE%A3%E7%A6%BB%E5%BC%80%E5%9B%BD%E5%AE%B6%E9%98%9F%E9%9B%86%E8%AE%AD&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n[中国女足 0-1 韩国202 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E5%A5%B3%E8%B6%B3+0-1+%E9%9F%A9%E5%9B%BD&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=921e19eb-3168-4997-806b-79dace45372e)\n_想来知乎工作？请发送邮件到 jobs@zhihu.com_\n打开知乎App\n在「我的页」右上角打开扫一扫\n![](https://picx.zhimg.com/v2-9e41ea16bdfbe9cf4896617ecad5b4ca.png)\n其他扫码方式：微信\n下载知乎App\n无障碍模式\n验证码登录\n密码登录\n[开通机构号](https://zhuanlan.zhihu.com/org/signup)\n中国 +86\n获取短信验证码\n忘记密码\n登录/注册\n请您登录后查看更多专业优质内容。\n其他方式登录\n未注册手机验证后自动登录，注册即代表同意[《知乎协议》](https://www.zhihu.com/term/zhihu-terms)[《隐私保护指引》](https://www.zhihu.com/term/privacy)\n扫码下载知乎 App\n关闭二维码\n请您登录后查看更多专业优质内容。\n请您登录后查看更多专业优质内容。\n", "body_base_path": "attachments/497ff60141d8/pages/", "body_markdown_path": "attachments/497ff60141d8/pages/archived-body.md", "local_html": "attachments/497ff60141d8/pages/archived-body.html", "body_source": "prior_capture" }