Chapter 1: 初识推荐系统：你每天都在用的「隐形推荐官」 / Meet Your Daily Invisible Assistant: What is a Recommender System?

昨晚刷短视频刷到凌晨1点的你，是不是突然发现：怎么刚搜完「露营帐篷怎么选」，首页就推了3条不同价位的帐篷测评？上周刚在购物车加了「敏感肌洗面奶」，今天打开电商APP，首页「猜你喜欢」的第一排全是各种温和氨基酸洁面？甚至你上周跟朋友吐槽「想找个靠谱的健身教练」，今天刷内容的时候就刷到了本地健身工作室的体验课推荐？ 别慌，这不是你手机被监听了，是你的「隐形推荐官」正式上线了。 If you stayed up until 1 a.m. scrolling through short videos last night, you might have suddenly noticed: as soon as you searched for "how to choose a camping tent", your homepage pushed 3 reviews of tents at different price points? Last week, you added a "sensitive skin facial cleanser" to your shopping cart, and today when you opened the e-commerce app, the first row of the "Guess You Like" section on the homepage was full of various mild amino acid cleansers? Even when you complained to your friends last week that "I want to find a reliable fitness coach", today when you scroll through content, you see recommendations for local fitness studio experience classes? Don’t panic, your phone is not being monitored. It’s your "invisible recommender assistant" that has officially gone online. 这个你每天接触、却很少主动注意到的技术，就是推荐系统（Recommender System）——它已经悄悄渗透到你生活的几乎所有数字场景里，成了你获取信息、消费内容、购买商品时最常用的「隐形助手」。 This technology that you come into contact with every day but rarely notice actively is the Recommender System. It has quietly penetrated almost all digital scenarios in your life and has become the most commonly used "invisible assistant" for you to obtain information, consume content, and purchase goods. 很多人对推荐系统的第一印象是「高深莫测的AI黑科技」，只有大厂的技术专家才能玩得转，其实完全不是。推荐系统的核心逻辑非常简单：就是把「你可能感兴趣的内容」从海量信息里筛出来，送到你面前。哪怕你没有任何技术基础，看完这本书，也能自己动手写一个简单的推荐系统。 Many people’s first impression of recommender systems is that they are "mysterious AI black technology" that only big tech experts can play with, but this is completely wrong. The core logic of recommender systems is very simple: it filters out "content you may be interested in" from massive amounts of information and sends it to you. Even if you have no technical foundation, after reading this book, you can write a simple recommender system yourself.

1.1 无处不在的「隐形推荐官」

1.1 The ubiquitous "invisible recommender assistant"

如果你仔细回想，你昨天接触到的所有数字内容里，至少70%都是推荐系统推给你的：

• 早上打开抖音/快手刷短视频，前10条里至少有8条是推荐系统根据你之前的观看习惯推的；
• 上班摸鱼逛淘宝/京东，首页的「猜你喜欢」、商品详情页的「看了又看」「买了又买」，全是推荐系统的杰作；
• 中午点外卖，美团/饿了么首页的「为你推荐」店铺和菜品，也是基于你之前的点餐习惯推的；
• 晚上下班听歌，网易云音乐的「每日推荐」、QQ音乐的「个性电台」，都是根据你平时的听歌偏好生成的；
• 甚至你刷BOSS直聘找工作，首页推荐的岗位、刷小红书看穿搭笔记，背后都有推荐系统在工作。

If you think carefully, at least 70% of all the digital content you came into contact with yesterday was pushed to you by a recommender system:

• When you open Douyin/Kuaishou to watch short videos in the morning, at least 8 of the first 10 videos are pushed by the recommender system based on your previous viewing habits;
• When you go to Taobao/JD.com to shop during work hours, the "Guess You Like" on the homepage, "People who watched this also watched" and "People who bought this also bought" on the product detail page are all masterpieces of the recommender system;
• When you order takeout at noon, the "Recommended for You" stores and dishes on the homepage of Meituan/Ele.me are also pushed based on your previous ordering habits;
• When you listen to music after get off work in the evening, NetEase Cloud Music's "Daily Recommendation" and QQ Music's "Personal Radio" are all generated based on your usual music preferences;
• Even when you browse BOSS Zhipin to look for a job, the recommended positions and the outfit notes you browse on Xiaohongshu all have recommender systems working behind them.

这些场景里的推荐系统，本质上都是在做同一件事：在海量信息里，找到你最可能需要的那部分，减少你找信息的成本。 The recommender systems in these scenarios are essentially doing the same thing: finding the part you may need most from massive amounts of information, reducing your cost of finding information. 你可以把它理解成你每个数字产品的专属管家：它知道你喜欢看猫视频还是健身教程，知道你喜欢买平价护肤品还是高端数码产品，知道你喜欢听周杰伦还是古典音乐，甚至知道你今天是不是刚分手，需要看点搞笑视频转移注意力。你不需要主动告诉它你的喜好，它通过你平时的行为数据就能慢慢「摸透」你的偏好，然后给你推最合适的内容。 You can think of it as your exclusive steward for each digital product: it knows whether you like watching cat videos or fitness tutorials, whether you like buying affordable skin care products or high-end digital products, whether you like listening to Jay Chou or classical music, and even whether you just broke up today and need to watch some funny videos to distract yourself. You don't need to actively tell it your preferences, it can slowly "figure out" your preferences through your usual behavioral data, and then push the most suitable content to you.

1.2 推荐系统解决的核心问题：信息过载

1.2 The core problem solved by recommender systems: Information Overload

要理解推荐系统的价值，首先要明白它要解决的核心问题：信息过载（Information Overload）。 To understand the value of recommender systems, we must first understand the core problem they solve: Information Overload. 根据市场研究机构Statista的统计，2024年全球每天产生的数字数据量达到328.77TB，相当于每天新增超过1000万小时的高清视频内容；仅YouTube平台，每分钟就有超过500小时的新视频被上传；国内头部电商平台的商品总量已经突破10亿件，覆盖从日用百货到工业设备的几乎所有品类。如果没有任何筛选机制，你哪怕24小时不眠不休地浏览，穷尽一生也无法看完平台上所有你感兴趣的内容。 According to statistics from market research firm Statista, the global daily digital data volume in 2024 will reach 328.77 TB, equivalent to more than 10 million hours of new high-definition video content added every day; on the YouTube platform alone, more than 500 hours of new videos are uploaded every minute; the total number of products on domestic leading e-commerce platforms has exceeded 1 billion, covering almost all product categories from daily necessities to industrial equipment. If there is no screening mechanism, even if you browse 24 hours a day without sleep, you will not be able to read all the content you are interested in on the platform in your lifetime. 这种「信息供给速度远远超过人类处理能力」的现象，就是信息过载。在互联网出现之前，信息过载的问题其实已经存在了：比如图书馆有几百万本书，你不可能全看完，所以会有图书馆管理员给你做荐书；报纸杂志有几百个版面，你不可能全看，所以会有编辑给你做内容精选。但互联网的出现让信息过载的问题放大了成千上万倍：以前图书馆的书再多，你一辈子也看不完，但至少你知道书在哪里；现在互联网上的内容多到你根本找不到你想要的那部分。 This phenomenon of "information supply speed far exceeding human processing capacity" is Information Overload. Before the advent of the Internet, the problem of information overload actually existed: for example, there are millions of books in the library, you can't read them all, so there will be librarians to recommend books for you; there are hundreds of pages of newspapers and magazines, you can't read them all, so there will be editors to do content selection for you. But the advent of the Internet has amplified the problem of information overload by thousands of times: there may be more books in the library than you can read in your lifetime, but at least you know where the books are; now there is so much content on the Internet that you can't find the part you want at all. 推荐系统就是解决这个问题的核心工具：它通过算法，把海量信息里和你最相关的部分筛出来，送到你面前，既节省了你的时间，也提升了你的体验。 Recommender systems are the core tool to solve this problem: through algorithms, it filters out the most relevant part from massive amounts of information and sends it to you, which not only saves your time, but also improves your experience. 从价值维度来看，推荐系统是典型的「双赢工具」： From a value perspective, recommender systems are typical "win-win tools": 1. 对用户的价值：节省找信息的时间，提升内容消费的体验。比如你想找一款适合送爸妈的保健品，不用在几千款商品里翻几个小时，首页直接给你推适合中老年、口碑好的产品，可能10分钟就找到了；你平时喜欢听周杰伦的歌，打开音乐APP直接给你推他的新歌和相似风格的歌曲，不用你自己搜。 **Value for

对用户的价值：节省找信息的时间，提升内容消费的体验。比如你想找一款适合送爸妈的保健品，不用在几千款商品里翻几个小时，首页直接给你推适合中老年、口碑好的产品，可能10分钟就找到了；你平时喜欢听周杰伦的歌，打开音乐APP直接给你推他的新歌和相似风格的歌曲，不用你自己搜。 Value for users: Save time on finding information and improve the experience of content consumption. For example, if you want to find a health supplement suitable for your parents, you don't need to spend a few hours flipping through thousands of products. The homepage will directly push you products suitable for middle-aged and elderly people with good reputation, and you may find it in just 10 minutes; if you usually like listening to Jay Chou's songs, when you open the music app, it will directly push you his new songs and songs of similar styles, without you searching for them yourself. > 💡 侧边栏：根据艾瑞咨询2024年的调研数据，使用个性化推荐的电商用户，平均找商品的时间比不用推荐的用户少62%，商品购买的转化率高47%；内容平台的用户平均使用时长也提升了38%。这就是推荐系统最直接的用户价值。 >