跳至主要内容

《深度学习推荐系统》By 王喆

 

推荐序言

深度学习/机器学习,爆发于2015年
算法能力+工程能力
1992年,施乐创建了应用协同过滤算法的推荐系统
本书for:推荐、广告、搜索领域的从业者/爱好者

第1章 互联网的增长引擎──推荐系统

  • 不同业务模式的公司定义的具体推荐系统优化目标不同,例如,视频类公司更注重用户观看时长,电商类公司更注重用户的购买转化率(Conversion Rate,CVR )
  • YouTube推荐系统的主要优化目标就是观看时长,而不是CTR
    • 工程论文:Deep Neural Networks for YouTube Recommenders
    • 大致流程:构建深度学习模型,预测用户观看某候选视频的时长,按照预测时长进行候选视频的排序,形成最终的推荐列表
  • 淘宝首页:千人千面(即使是1%的CTR提高,也会到来数十亿元收入增加)
  • 推荐系统逻辑框架:f(U,I,C), User, Item, Context
  • 推荐系统数据部分:实时数据、准实时数据、离线数据
  • 推荐系统模型部分:召回层、排序层、补充策略与算法层(排序层是工业界与学术界研究的核心)

第2章 前深度学习时代──推荐系统的进化之路

  • UserCF: 基于用户相似度进行推荐的协同过滤算法
    • 互联网用户数的增长会导致矩阵维护复杂度急剧上升
    • 用户的历史数据可能稀疏(导致UserCF不适用于酒店预订、大件商品购买等低频应用)
    • 适用于发现热点、以及跟踪热点的趋势
  • ItemCF:基于物品相似度进行推荐的协同过滤算法
    • 更适用于兴趣变化较为稳定的应用
  • 协同过滤的天然缺陷:推荐结果的头部效应较明显
  • 其它模型:矩阵分解、逻辑回归、FM、FFM、GBDT+LR、LS-PLM (细节已经要看不懂了)

第3章 浪潮之巅──深度学习在推荐系统中的应用

  • Google Play的推荐团队,基于对业务场景的深刻理解,提出了Wide&Deep模型
  • 这张属于看不懂那种类型

第4章 Embedding技术在推荐系统中的应用

  • Not Interested

第5章 多角度审视推荐系统

  • 从7个不同的角度切入推荐系统
    • 推荐系统如何选取和处理特征
    • 召回层的主要策略有哪些?
    • 实时性的重要性体现在哪里?有哪些提高实时性的方法?
    • 如何根据具体场景构建推荐模型的优化目标
    • 如何基于用户动机改进模型结构
    • 冷启动问题的解决方法有哪些
    • 什么是探索与利用问题?有哪些主流的解决方法

第6章 深度学习推荐系统的工程实现

第7章 推荐系统的评估

  • 离线评估与AB测试

第8章 深度学习推荐系统的前沿实践

  • YouTube视频推荐的实践
  • 阿里巴巴商品推荐的实践
  • FB的深度学习推荐系统
  • Airbnb基于Embedding的实时搜索推荐系统

第9章 构建属于你的推荐系统知识框架

  • 构建自己的知识框架
  • 优秀的推荐工程师:兼具深度与广度(知识理论能力、工程能力),其实其它领域同样如此

后记

Popular posts from 产品随想的博客

产品随想 | 周刊 第130期:集结信徒,而非官僚

On Dyson, techno-centric design and social consumption   https://2earth.github.io/website/20250707.html 如何創造偉大的事物   https://ryolu.notion.site/1610a94b9c108079a95be4362afd4a26 集結信徒,而非官僚 Reflections on OpenAI   https://calv.info/openai-reflections 创业架构 Shui   https://github.com/rock-zhang/Shui 好好喝水 https://www.ghibli.jp/works/ 吉卜力作品的高清图 From Skeuomorphic to Liquid Glass: Apple's Strategic Bet on the Post-Touch Future   https://omc345.substack.com/p/from-skeuomorphic-to-liquid-glass 迄今为止关于苹果Liquid Glass变革的解读,最好的一篇 The Nueva School   https://en.wikipedia.org/wiki/The_Nueva_School?useskin=vector 看起来是很酷的一个学校 The Barbican   https://arslan.io/2025/05/12/barbican-estate/ 史蒂夫·乔布斯希望你阅读的 9 本书   https://www.douban.com/doulist/147158849/ 《禅者的初心》里有句话:“做任何事,其实都是在展示内心的天性,这是我们存在的唯一理由。” Chuck Feeney was one of the greatest philanthropists ever   https://www.gatesnotes.com/Remembering-Chuck-Feeney 慈善家 My new deadline: 20 years to give away virtually all my wealt...

Apple's One-Dollar-a-Year Man, By Steve Jobs, 2000

(FORTUNE Magazine) – Now that Steve Jobs has showed his hand on Apple's Internet and system software strategies and dropped the "interim" from his title, other questions loom. He's always denied it, but isn't it true that his old company, Next, did wind up taking over Apple? Will there ever be an encore to the 15-year-old Macintosh? Short of that, does Apple have any plans to jump into the "Internet appliance" fray? Will Apple ever build computers for business people again? And what, pray tell, does Steve think of all these young Internet zillionaires? Let's ask. Practically every technology that your old company, Next, possessed when Apple acquired it in 1997 is now being used by Apple in some strategic way. This must seem like sweet vindication.  The thing about Next was that we produced something that was truly brilliant for an audience that our heart really wasn't into selling to--namely, the enterprise. I suppose if you were wr...

《Becoming Steve Jobs》Chapter 16 Blind Spots, Grudges, and Sharp Elbows

Steve could be pretty thin-skinned when someone prominent criticized the aesthetics of his products. He took great umbrage that Neil would, as Steve put it, “pop off in public like that without coming to talk to us about his technical concerns first.” From that point on he had rebuffed all of Neil’s attempts to smoke the peace pipe. 有趣 He had blind spots, grating behavioral habits, and a tendency to give in to emotional impulse that persisted his entire life. These characteristics are often used to make the case that Steve was an “asshole” or a “jerk,” or perhaps simply “binary”—that odd adjective often used to convey the sense that he was half asshole/half genius from birth to death. These aren’t useful, interesting, or enlightening descriptions. What’s more illuminating is to take a look at the specific ways in which Steve failed to do an effective job of tempering some of his weaknesses and antisocial traits, and to consider how, when, and why some of them continued to flare up even...

A Sister’s Eulogy for Steve Jobs

I grew up as an only child, with a single mother. Because we were poor and because I knew my father had emigrated from Syria, I imagined he looked like Omar Sharif. I hoped he would be rich and kind and would come into our lives (and our not yet furnished apartment) and help us. Later, after I’d met my father, I tried to believe he’d changed his number and left no forwarding address because he was an idealistic revolutionary, plotting a new world for the Arab people. Even as a feminist, my whole life I’d been waiting for a man to love, who could love me. For decades, I’d thought that man would be my father. When I was 25, I met that man and he was my brother. By then, I lived in New York, where I was trying to write my first novel. I had a job at a small magazine in an office the size of a closet, with three other aspiring writers. When one day a lawyer called me — me, the middle-class girl from California who hassled the boss to buy us health insurance — and said his cl...

产品爱好者周刊 第26期:PRISM, XKeyscore, Trust No One

  Products Gitea - Git with a cup of tea   https://gitea.io/en-us/ A painless self-hosted Git service. 自建Git服务,避免GitHub隐私侵犯 https://github.com/objective-see/LuLu LuLu is the free macOS firewall 监视Mac的出站流量,且阻断 OverSight   https://github.com/objective-see/OverSight OverSight monitors a mac's mic and webcam, alerting the user when the internal mic is activated, or whenever a process accesses the webcam. 监视是否有应用调用Mac的麦克风、摄像头 Mozilla Hubs   https://github.com/mozilla/hubs The client-side code for Mozilla Hubs, an online 3D collaboration platform that works for desktop, mobile, and VR platforms. 开源的多人虚拟空间,Mozilla打造,企业级VR诉求 数字移民   https://shuziyimin.org 关于内容源、工具的推荐,适合刚接入国际的新人 SimpleLogin   https://simplelogin.io/ 匿名邮箱工具,转发用,Michael Bazzell推荐 Telegram 群组、频道、机器人 - 汇总分享   https://congcong0806.github.io/2018/04/24/Telegram/#机器人-bot https://archive.ph/iJMBj 献给那些将来到Telegram的朋友 Design Patrick Wardle   https://www.instagram.com/patrickwardle/?hl=en 他的IG,摄影也精彩,审美...

黑客技术论坛推荐

原文来自知乎, 世界各大黑客技术论坛TOP排行榜 ,表示感谢 这份名单基本上囊括了目前世界上各大最佳黑客技术论坛。

产品随想 | 周刊 第127期:晨光只开一刻钟,但比千年松,并无甚不同

Cherry Studio   https://github.com/CherryHQ/cherry-studio Cherry Studio is a desktop client that supports for multiple LLM providers. Support deepseek-r1 Aalto Repository beta   https://repo.aalto.fi/ Images, sounds and videos from Aalto University 这个系列,价值极高 Nokia Design Archive   https://nokiadesignarchive.aalto.fi/ 芬兰这个国家很了不起 对话影石刘靖康:两代未出现划时代的产品,就会沦为平庸的品牌   https://www.geekpark.net/news/308996 还挺喜欢这个创始人的,有一种海盗的内涵 从哈佛、明星创业者到酷家乐副总裁,苏奇的传奇   https://app.modaiyun.com/mdy/article/3FO4K4W0M259 WHO关于猫狗咬伤、抓伤的处理建议 动物咬伤: https://www.who.int/zh/news-room/fact-sheets/detail/animal-bites 狂犬病: https://www.who.int/zh/news-room/fact-sheets/detail/rabies 关于狂犬病的10个事实: https://www.who.int/zh/news-room/facts-in-pictures/detail/rabies INDIGO 新年直播(2025)   https://www.youtube.com/live/ZIgPvSDGAfY 对2024年AI发展的回顾部分特别好 Artab   https://github.com/get-artab/artab Get Inspired by the World's Greatest Artworks Every Time You Open a New Tab. Extension Available for Chrome, Edge, and...

巴菲特致股东信-1980年

 笔记: 会计中对于下属股权公司的记账方式有3种: 持股50%以上,全部并入 持续20%--50%,则按持股比例并入 持股20%以下,则以实际收到的利润返还,计入报表 这种会计方式,会导致伯克希尔旗下,不少的企业,未能暴露实际的收益情况 对伯克希尔而言,对盈余的认定并非取决于持股比例是100%,50%,20%,5%或是1%,盈余的真正价值在于其将来再投资所能产生的效益 我们宁愿将所赚的盈余继续交由不受我们控制的人好好发挥,也不希望转由我们自己来浪费 高通货膨胀等于是对投入的资本额外课了一次税 翻译: https://xueqiu.com/6217262310/131837878 https://archive.ph/XMX5n  原文: Buffett’s Letters To Berkshire Shareholders 1980 巴菲特致股东的信 1980 年 Operating earnings improved to $41.9 million in 1980 from $36.0 million in 1979, but return on beginning equity capital (with securities valued at cost) fell to 17.8% from 18.6%. We believe the latter yardstick to be the most appropriate measure of single-year managerial economic performance. Informed use of that yardstick, however, requires an understanding of many factors, including accounting policies, historical ca...

写给大家看的中文排版指南

作者:Hindy 原文地址: http:// zhuanlan.zhihu.com/uici rcle/20506092 前言:很遗憾,我们的周围充斥着大量排版丑陋的文章。我国的字体排印与日本、美国等设计强国差距实在太大。我希望能够做些力所能及的小事,让更多人意识到“设计”的价值和其必要性,创造更美好的视觉环境。本文旨在帮助普及、提升大家对文字排版的认识,让大家在平时的学习工作中能有更专业的文字排版素养。 必看人群: 设计师、编辑、作家、撰稿人、教师、学生 目录: 1. 中文排版 1.1 引号 1.2 省略号与破折号 1.3 行首行尾禁则 2. 西文排版基础 2.1 西文撰写基础 2.2 西文标点相关 2.3 斜体的用法 2.4 大小写的区别 3. 中西文混排 3.1 基础原则 3.2 标点相关 1. 中文排版 1.1 引号 我国国家标准要求弯引号,个人建议使用直角引号。 示例:你竟然喜欢“苹果表”? 引号中再用引号使用双直角引号。 示例:我问他,“你竟然喜欢‘苹果表’?” 当引号表示讽刺、反语暗示时,使用弯引号(用法参考“西文排版”部分)。 示例:说真的,我也很 “喜欢”“苹果表”哦。 1.2 省略号(删节号)与破折号 省略号占两个汉字空间,包含六个点。 正确示例:中国设计还有太长路要走…… 错误示例:中国设计还有太长路要走… 破折号占两个汉字空间。 示例:中国设计还有太长路要走──加油罢。 1.3 行首行尾禁则 点号(顿号、逗号、句号等)、结束引号、结束括号等,不能出现在一行的开头。 错误示例: 排版时注意某些 符号不能在行首 ,别弄错了。 正确示例: 排版时注意某些 符号不能在行首, 别弄错了。 开始引号、开始括号、开始双书名号等,不能出现在一行的结尾。 错误示例: 她对我们说:“ 这书太赞了。” 正确示例: 她对我们说: “这书太赞了。” 2. 西文排版基础 2.1 西文撰写基础 句首字母大写。 单词间留空格。 示例:Have a question? 2.2 西文标点相关 点号后加一个空格(如逗号、句号等)。 示例:Hello everyone! Welcome to my blog....

360T7 刷机步骤及固件

https://cmi.hanwckf.top/p/360t7-firmware/   360T7的固件支持由immortalwrt-mt798x项目提供支持,请参考: https://cmi.hanwckf.top/p/immortalwrt-mt798x https://github.com/hanwckf/immortalwrt-mt798x 刷机步骤 参考 此处 的办法开启原厂固件的UART和telnet功能 在以下链接下载360T7测试固件(纯净版,无任何插件) https://wwd.lanzout.com/b0bt9idwd 密码:ezex (此固件已过时,请选择其它更新的固件) 接下来将刷入修改版uboot。修改版uboot的优点有: 固件分区可达108MB,原厂uboot只能使用36M 自带一个简单的webui恢复页面 到以下仓库的Release页面下载uboot,目前暂时仅支持360T7,后续将支持更多mt798x路由器。 推荐使用 mt7981_360t7-fip-fixed-parts.bin , fixed-parts 代表uboot分区表在编译期间固定,不会随着uboot环境变量变化。 https://github.com/hanwckf/bl-mt798x/releases/latest 将 mt7981_360t7-fip-fixed-parts.bin 通过HFS等方式上传到路由器,使用以下命令刷入uboot mtd write mt7981_360t7-fip-fixed-parts.bin fip 确认刷入完毕后,拔掉路由器电源。然后将电脑的IP地址设置为固定的 192.168.1.2 ,接着按住路由器的RESET按钮后通电开机,等待8s后用浏览器进入 192.168.1.1 在uboot恢复页面选择要刷入的固件。immortalwrt-mt798x目前编译两个版本的360T7固件。 建议修改版uboot直接使用 immortalwrt-mediatek-mt7981-mt7981-360-t7-108M-squashfs-factory.bin ,两种固件区别如下: mt7981-360-t7-108M 为108M固件分区,原厂uboot不可启动,需要修改版u...