# 推荐系统

经典机器学习系统设计题目，注意把各环节串联起来，形成框架性思考。

例如大规模推荐，为什么需要召回加排序的漏斗结构？是因为item数量大，召回能**快速**把大规模candidate显著减小，常见的双塔模型为什么能够快速召回呢？因为用户塔和物品塔无交叉，物品塔可离线计算，用户塔为了体现兴趣可以实时计算，但一次请求只需计算一个用户向量，通过ANN快速计算索引。 因此也决定了召回模型是late fusion, 而排序模型是early fusion(较早进行特征融合能够提升预测精度).

配合[推荐系统理论](/mle-interview/02_ml/17_recommendation.md)，针对不同领域，如电商、O2O，针对领域提出针对性的优化

* 电商推荐业务：曝光->点击->购买
* 地点约束，例如yelp的饭馆推荐涉及geolocation information
* user的 graph network，例如facebook [newsfeed推荐](/mle-interview/03_system/03_ml/news_feed.md)
* 音乐、[视频](/mle-interview/03_system/03_ml/video_recommendation.md)的embedding，例如spotify音乐推荐
* Ins Story推荐，每条Story是独一无二的并且是有时间性的
* O2O场景广告特点：移动化、本地化、场景化、多样性
* [Point of interest](/mle-interview/03_system/03_ml/poi_recommendation.md)

## 1. requirements

> 推荐系统的核心功能还是推荐的personalization accuracy，diversity

**Functional/use cases**

* what is the product for which we have to build a recommendation system, How are we different from XX? Who is the producer/consumer
* Homepage recommendation, session based next item recommendation (short term interest), or related item recommendations
* Explicit feedback or Implicit feedback （即使有explicit, 一般也会选择implicit）
* it's a new product, or we have some current product built already
* the biggest goal of recommendation system is user engagement. Can I assume that to be the goal?
* Is there any demography or country we are targeting
* Text, image, video
* Ranking / Localization
* 是否有friend, follow. 有的话可以作为一个召回通道

**Non-Functional**

> 一定要clarify: scalability, low latency. 因为这两个non-functional requirement决定了后面怎么设计

* MVP and Non-MVP
* users should have a real time/ near real time / small latency experience Idempotency/ exact-once/ at-least-once/ at-most-once
* reliability: data not get lost/ job not executed in parallel retention policy
* security: store encoded personal data
* consistency: read/ write heavy
* dau/ qps / peak qps

## 2. ML task & MVP

**名词解释**

* 曝光(impression): 文档被用户看到
* 点击率(click-through-rate，CTR): 文档d曝光的前提下，用户点击d的概率
* 交互行为(engagement): 在点击的前提下, 文档的点赞、收藏、转发、关注作者、评论；电商的加购物车、下单、付款

推荐系统的重要性源自**信息过载**与**人们行为的长尾分布**. 目的是Link user with items in a reasonable way.

* 长尾/热门
  * 头部用户精细刻画，准确记忆；占比更大的稀疏长尾，需要很好地泛化
* 记忆/探索
* 稀疏/embedding

![](/files/qKqqYPsFaoT5kIfafMLT)

![](/files/n34n49AAjJ8YNUg22QLw)

## 3. data & labels

* 关键：分布
* 通过客户端以及服务端的实时数据，经过流处理的平台，把用户、商品、场景的信息以及端侧的信息全部都收集全
* 再通过特征工程的很多方法，如归一化、离散化、非线性离散等很多的变换，把特征处理成下游模型处理的方式。
* 处理完后一般会放在特征存储中，如 KV 存储: Redis、阿里的 iGraph 等
* 样本
  * 不均衡
  * 置信度: skip-above(点击的item位置以上的展现可以当做负样本, 最深位置以后的样本过滤掉); 完全无正样本session(可能是碰巧唤醒)
* labels
  * explicit vs. implicit labels
* 延迟转化 Delayed Feedback

## 4. feature

![feature](/files/ieXde1eBHlxDVqbZ7x7I)

[那些年，我们追过的 Feature](https://pyemma.github.io/Features-in-Recommendation-System/)

* dense feature
  * log变换
  * 分箱，可以根据业务分箱，分箱思路：等频，等宽，卡方
    * 分桶交叉特征提升泛化能力
* sparse feature
  * onehot
  * embedding
    * [电商场景下的itemid有上亿，embedding](https://zhuanlan.zhihu.com/p/397600084)
* cross feature
  * [特征工程中的特征交叉究竟是什么? - 谢陈龙的回答 - 知乎](https://www.zhihu.com/question/63593481/answer/3176327607)
* sequence feature
* 处理
  * 缺失
  * 标准化
  * 平滑与消偏

## 5. model

> 熟练掌握模型: DNN, DeepFM, DIN, MMOE

**召回**

> * Generates a much smaller subset of candidates from huge corpus with low latency and computational cost
> * ensemble retrieval from different models (rule, filtering, nn). A given model may provide multiple candidate generators, each nominating a different subset of candidates
> * Neighborhood models (itemCF, userCF) are most effective at detecting very localized relationships, but unable to capture the totality of weak signals encompassed in all of a user’s ratings.
> * Latent factor models (MF) are generally effective at estimating overall structure that relates simultaneously to most or all items. However, these models are poor at detecting strong associations among a small set of closely related items.

* 召回系统的要求是，“低延时”与“高精度（precision）
* 多路召回 (ensemble retrieval)
* 负样本选择
  * recommendation as extreme multiclass classification: 每一个正样本，意味着其他所有class都是负样本 -> negative sampling -> generation, accuracy
  * 全局负采样，采样方式: uniform subsampling, negative down sampling
  * Batch内负采样的问题： batch内都是热门物品，导致采样后的负样本中也大都为热门物品，造成对热门物品的过度打压，校正：
  * 负采样带来的问题：CTR预估值漂移, 校正: p / (p + (1-p)/w)
* model
  * 规则：热度高，同一作者、tag
  * content based
    * The model doesn't need any data about other users
  * itemCF
    * ItemCF基于item之间的共现关系计算相似度，item行为越多，就会与更多的item发生共现，进而获得更多的曝光，即推荐系统中的马太效应或长尾效应
    * pro: 无需训练，长于记忆；效果好
    * con: 泛化能力弱；容易产生马太效应，推荐的都是头部和中部产品; Cannot handle fresh items; Hard to include side features for query/item
  * two power
    * arbitrary continuous and categorical features can be easily added to the model, 还有图片、文本等多模态都很容易加到双塔模型
    * 更容易scalable, 处理大规模数据
  * matrix factorization
    * con: worse performance on tail users
    * inability to add content-based features
  * embedding: graph, picture, text
* 局部敏感哈希，KD树
* similarity metrics

**排序**

* 粗排、精排、重排
* multi-task deep learning
* 粗排一致性
* 记忆和泛化
* feature/embedding
* model
  * rule based model, matrix factorization, factorization machine
  * 转化为classification/regression模型
  * 深度学习 (wide and deep learning)
    * pro：泛化能力，增量训练，推理速度
    * con: performance ramp-up, preprocessing
  * GBDT
    * pro: without preprocess, good performance
    * con: large scale, unsuitable for continual learning
* rerank
  * take into account additional constraints for the final ranking
  * DPP多样性算法
* multi-task
  * 稀疏task可以借助其他task学习
  * consistency among tasks
  * richer feature
  * 不好的地方是: 优化，尤其是想优化某个task变得更难
  * shared-bottom，MMoE，PLE

## 6. evaluation

* offline
  * ranking: map, ndcg
* online
  * CTR

## 7. deploy & service

* parameter server
  * 本质是一个: 分布式键值存储系统
* 召回中user embedding为什么线上实时计算
  * 用户的兴趣短时间内可能会发生显著变化
  * 用户的行为数据通常是稀疏的
  * 缓解用户冷启动
* 召回中item embedding为什么离线计算好
  * 量大，静态
  * 实时性要求
* pagination + sort\_key / video start timestamp / offset, page\_size
* user context/ client info(ios, network condition) diagram
* list high level diagram first, then each component
* different choices/ tradeoffs and give your recommendation
* MQ or no mq
* cache: consistency/ failure/ cold start
* data model/database: 1 master, 2 replica, primary key, user\_id, timestamp, status
* embedding矩阵的lazy初始化和lazy更新
* embedding矩阵对分布式通信的影响

## 8. Monitoring & maintenance

[How to Monitor a Recommender System](https://code.tubitv.com/how-to-monitor-a-recommender-system-6d720c922c90)

注意区分statistical metric和business metric。后者意义更大，但经常无法直接optimize，只能通过ab-testing测试

* 电商：根据业务需要，在 GMV (商品交易总额) 主目标之外，通常还要兼顾 IPV、转化率、人均订单数等多个次目标
* ctr（点击率）和 CVR (Conversion Rate) 转化率
* impression per second
* candidate count (from recall)
* budget burn rate
* 通用metrics：cpu, qps, latency
* 淘宝主搜将 “全域成交 Hit rate” 作为粗排最重要的评价标准，提出两类评价指标，分别描述“粗排->精排损失”和“召回->粗排损失”

## 9. 优化与问答

* 数据采集和处理
  * 如何建立index
* 召回
  * faiss: 使用PCA和PQ(Product quantization乘积量化)两种技术进行向量压缩和编码
  * 向量召回、排序没用实时行为序列特征
* 怎么做counterfactual evaluation
* 怎么deploy?
  * embedding retrieval
  * **real time or batch**, Batch Prediction Versus Online Prediction
* 怎么上线？
  * A/B testing, metric
    * setup A/B testing
    * control group (current system), treatment (system w/ our new model)
    * use a canary release, start with 99% control and 1% treatment
    * compare the key business metrics between the two groups (conversion rate etc.)
    * slowly ramp up the traffic in the treatment group
  * 热启动与冷启动
  * 老汤模型的本质，是样本空间不一致导致的公平性问题
    * 方案一：回滚历史数据，对齐样本空间
    * 方案二：模型热启动warm-up
* 实时化
  * 特征实时化难点：实时数据处理能力（Flink等），并将处理结果实时写入Redis
  * 模型实时化
    * 样本数据流：实时样本落地、实时样本拼接（反馈延时问题）
    * 模型训练：在线学习（资源和稳定性问题）、实时模型纠偏
* 连续特征离散化方法以及为什么需要对连续特征离散化
* FM、FFM的参数量以及时间复杂度
* 多目标模型
  * 参数共享及不共享参数各自的优缺点
* 用户长期兴趣和多兴趣怎么建模
* 如何冷启动
  * 思路一: 尽可能应用side information/多模态进行推荐
  * 思路二: 尽可能用小流量探索出新物料的真实质量，扶持新item
  * [embedding 冷启动](https://zhuanlan.zhihu.com/p/351390011), embedding初始化采用default, 而不是random
* bias
  * 如何解决 position bias
    * 数据入手：点击前的展示作为负例
    * 模型入手：bias特征须通过线性层接入模型；训练时将位置作为特征喂入模型，预测时置0
  * popularity bias
* 对热门的打压 (long tail)
  * 召回：多路召回增加多样性
    * itemCF/swing相似度计算中，把共同点击商品A和B的用户pair对的交集放在分母，对热门用户行为的打击
    * 召回时，对负样本进行下采样，采样概率大致为点击数的0.75次方，对热门进行打压。注意in-batch-negative中，打压过度了
    * log q correction
  * 精排
    * item count取对数
    * 热门特征加入到偏差的网络中，预测时去掉
  * 流量调控
* 稀疏
  * [魔改结构什么时候有效 - 古德猫宁的文章 - 知乎](https://zhuanlan.zhihu.com/p/698600932)
* 模型更新策略, retrain plan
  * 按天增量训练，实时训练(在线学习)，或者 在线学习+天级增量结合
  * embedding的更新
    * [推荐系统embedding过大如何压缩一下？最近正在解决这个问题？ - Jesper的回答 - 知乎](https://www.zhihu.com/question/522006535/answer/2836825253)
  * 全量更新：可以每天更新一次，shuffle, 更新ID embedding 和全连接层，1 epoch。每次更新的还是上一天的全量的模型更新，而不是增量
  * 增量更新：不停做，可以几十分钟更新一次，online learning只更新ID embedding参数, 尽量实时追踪用户兴趣。但其实是有偏的
    * Real-time Logging
    * 在线学习所需的实时特征可通过在线inference系统实时输出
* 怎么加user and item metadata
* 线上评价，线上线性不一致
* model debugging, offline online inconsistency, light ranking, ab test, heavy ranking, two tower
* 统计特征用等宽分桶导致特征值聚集
* 多语言搜索召回率低
* 有些国家节日多，模型T+1更新导致节日后消费数据下降
* 有一些情况下同一用户对不同item的 pctr 是同一个值
* 模型目标和业务目标不一致
* Itemid hash 碰撞率太高
* E\&E
  * embedding: 特征转化为可以学习的向量，模糊查找
  * embedding in sequence: 共现
* 多任务学习
* 多场景学习 (multi-domain/multi-scenario/multi-channel)
  * 不同用户群体（如新老用户）、APP不同频道模块、不同客户端等

## Reference

**精读**

* [Recommendations: What and Why?](https://developers.google.com/machine-learning/recommendation/overview)

**扩展**

* [Best Practices for Building and Deploying Recommender Systems](https://docs.nvidia.com/deeplearning/performance/recsys-best-practices/index.html)
* <https://github.com/Doragd/Algorithm-Practice-in-Industry>
* [超详细：完整的推荐系统架构设计](https://xie.infoq.cn/article/e1db36aecf60b4da29f56eeb4)
* [推荐系统--完整的架构设计和算法(协同过滤、隐语义)](https://zhuanlan.zhihu.com/p/81752025)
* <https://www.6aiq.com/article/1553963227373>
* [闲鱼搜广推类技术文章汇总](https://zhuanlan.zhihu.com/p/603997107)
* [Improving Deep Learning for Ranking Stays at Airbnb](https://medium.com/airbnb-engineering/improving-deep-learning-for-ranking-stays-at-airbnb-959097638bde)
* [Machine Learning-Powered Search Ranking of Airbnb Experiences](https://medium.com/airbnb-engineering/machine-learning-powered-search-ranking-of-airbnb-experiences-110b4b1a0789)
* [wish冷启动](https://kojinoshiba.com/recsys-cold-start/)
* [阿里-推荐系统综述](https://mp.weixin.qq.com/s/e9xjwefYk2toN9CGK5uPXg)
* [推荐算法优化闲聊](https://zhuanlan.zhihu.com/p/665102155)
* [负样本为王：评Facebook的向量化召回算法](https://zhuanlan.zhihu.com/p/165064102)
* [探讨特征工程的方法论](https://zhuanlan.zhihu.com/p/466685415)
* [现在互联网公司还有做特征工程的工作吗？](https://www.zhihu.com/question/512722857)
* [Recommender Systems, Not Just Recommender Models](https://medium.com/nvidia-merlin/recommender-systems-not-just-recommender-models-485c161c755e)
* [在工业界，应用 Multi-Armed Bandit 的例子多吗？ - 曾文俊的回答 - 知乎](https://www.zhihu.com/question/293811863/answer/843666533)
* [强化学习在美团“猜你喜欢”的实践](https://tech.meituan.com/2018/11/15/reinforcement-learning-in-mt-recommend-system.html)
* [MLSYS-深度推荐系统](https://openmlsys.github.io/chapter_recommender_system/system_architecture.html)
* <https://research.facebook.com/blog/2018/5/the-facebook-field-guide-to-machine-learning-video-series/>
* [推荐系统架构](https://www.zhihu.com/people/yan-yiceng/posts)
* [ML Systems Design Interview Guide](http://patrickhalina.com/posts/ml-systems-design-interview-guide/)
* [CTR/推荐系统 特征工程文章汇总](https://zhuanlan.zhihu.com/p/564479089)
  * [推荐系统简介之特征工程笔记 - Shard Zhang的文章 - 知乎](https://zhuanlan.zhihu.com/p/671291077)
* [CTR/推荐系统中embedding应用概述文章汇总](https://zhuanlan.zhihu.com/p/369236413)
* [CTR/推荐系统中多任务/多目标学习应用概述文章汇总](https://zhuanlan.zhihu.com/p/369289378)
* [CTR/推荐系统冷启动文章汇总](https://zhuanlan.zhihu.com/p/530948093)
* [CTR/推荐系统 冷启动Exploitation & Exploration文章汇总](https://zhuanlan.zhihu.com/p/531020134)
* [CTR/推荐系统中Debias应用概述文章汇总](https://zhuanlan.zhihu.com/p/518175104)
* [CTR/推荐系统线上线下不一致讨论文章汇总](https://zhuanlan.zhihu.com/p/518353056)
* [CTR/推荐系统 转化延迟文章汇总](https://zhuanlan.zhihu.com/p/531949459)
  * [推荐系统简介之标签拼接和延迟反馈 - Shard Zhang的文章 - 知乎](https://zhuanlan.zhihu.com/p/668885759)
* [都说数据是上限，推荐系统ctr模型中，构造正负样本有哪些实用的trick？ - 武侠超人的回答 - 知乎](https://www.zhihu.com/question/324986054/answer/1751584807)
* [一文读懂「Parameter Server」的分布式机器学习训练原理 - 王喆的文章 - 知乎](https://zhuanlan.zhihu.com/p/82116922)
* [Persia: An Open, Hybrid System Scaling Deep Learning-based Recommenders up to 100 Trillion Parameters](https://arxiv.org/pdf/2111.05897)
* [推荐系统Serving架构分析 - Peter的文章 - 知乎](https://zhuanlan.zhihu.com/p/335116835)
* [推荐系统(10):样本拼接工程实践 - yanianthe的文章 - 知乎](https://zhuanlan.zhihu.com/p/594275446)

**代码**

* [fun-rec](https://github.com/datawhalechina/fun-rec)
* [facebook-DLRM](https://github.com/pytorch/torchrec/blob/main/torchrec/models/dlrm.py)
* [torch-rechub](https://github.com/datawhalechina/torch-rechub/tree/main)
* [EasyRec](https://github.com/alibaba/EasyRec)
* <https://github.com/wzhe06/SparrowRecSys>
* <https://github.com/twitter/the-algorithm>
* [spotify\_mpd\_two\_tower](https://github.com/jswortz/spotify_mpd_two_tower)
* <https://github.com/jiawei-chen/RecDebiasing>
* [Multitask-Recommendation-Library](https://github.com/easezyc/Multitask-Recommendation-Library)
* <https://github.com/rixwew/pytorch-fm/tree/master/torchfm/model>
* <https://github.com/LongmaoTeamTf/deep_recommenders/tree/master>
* TF1：Microsoft recommendation
* TF2.5+：GitHub - ZiyaoGeng/RecLearn: Recommender Learning with Tensorflow2.x
* Pytorch：RecBole推荐系统框架
* <https://github.com/CheckChe0803/flink-recommandSystem-demo>

**部署**

* MOBIUS: Towards the Next Generation of Query-Ad Matchingin Baidu’s Sponsored Search
* COLD: Towards the Next Generation of Pre-Ranking System
* [DeepRec](https://deeprec.readthedocs.io/zh/latest/index.html): 支持淘宝搜索、猜你喜欢、定向、直通车等核心业务，支撑着千亿特征、万亿样本超大规模的稀疏训练


---

# Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://longxingtan.gitbook.io/mle-interview/03_system/03_ml/recommendation.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.