推薦系統的人工調控策略(二十八)

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"寫在前面：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"大家好，我是強哥，一個熱愛分享的技術狂。目前已有 12 年大數據與AI相關項目經驗， 10 年推薦系統研究及實踐經驗。平時喜歡讀書、暴走和寫作。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"業餘時間專注於輸出大數據、AI等相關文章，目前已經輸出了40萬字的推薦系統系列精品文章，今年 6 月底會出版「構建企業級推薦系統：算法、工程實現與案例分析」一書。如果這些文章能夠幫助你快速入門，實現職場升職加薪，我將不勝歡喜。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"想要獲得更多免費學習資料或內推信息，一定要看到文章最後喔。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"內推信息","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果你正在看相關的招聘信息，請加我微信：liuq4360，我這裏有很多內推資源等着你，歡迎投遞簡歷。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"免費學習資料","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果你想獲得更多免費的學習資料，請關注同名公衆號【數據與智能】，輸入“資料”即可！","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"學習交流羣","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果你想找到組織，和大家一起學習成長，交流經驗，也可以加入我們的學習成長羣。羣裏有老司機帶你飛，另有小哥哥、小姐姐等你來勾搭！加小姐姐微信：epsila，她會帶你入羣。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着移動互聯網的深入發展，推薦系統越來越得到企業界的認可，成爲toC互聯網公司的標配技術。推薦系統藉助機器學習技術，基於對用戶行爲的挖掘，能夠洞察用戶的興趣偏好，自動化地爲用戶生成個性化的內容推薦，整個推薦的過程基本可以做到完全自動化，不需要人工干預。雖然很多方面機器可以比人做得更好，但機器也有自身的問題和缺點。目前的人工智能在情感、應急處理、複雜問題決策等很多方面還根本無法與人相比，這些方面都可以很好地體現人類的價值。在推薦系統中，這一情況也存在，推薦系統需要藉助人工來進行策略調控，以獲得優質的用戶體驗、更好地實現商業目標。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"利用人工對推薦系統進行策略調控，除了用戶體驗的考慮外，還有安全性、商業價值等維度的權衡。本章我們就來講講推薦系統中的策略調控問題。具體來說，我們會從什麼是推薦系統的人工調控、爲什麼要進行人工調控、怎樣進行人工調控、怎樣評估人工調控的價值、人工調控面臨的挑戰、人與機器的有效配合等6個角度來介紹相關知識點。通過本章的分析和講解，希望讀者可以更好地理解人工調控在推薦系統中的作用與價值。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.1 什麼是推薦系統的人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"企業級推薦系統進行推薦的流程一般可以分爲召回、排序、業務調控3個階段(見下面圖1)，這其中的第三個階段業務調控就涉及到人工調控策略，這只是其中一種可行的干預方式，也是比較重要的一種干預手段，後面我們會詳細講解在這裏可以進行哪些干預。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/86/86f214ce4375746e260a472db35a101a.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"size","attrs":{"size":10}}],"text":"圖1：企業級推薦系統三階段pipeline架構","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"一般來說，一切對推薦系統運行過程中的策略和模塊進行人工調整的方法都叫做人工干預。廣義地說，選擇什麼樣的數據集、選擇什麼模型、怎麼定義參數、對模型結果的干預等等都屬於人工干預的範疇。","attrs":{}},{"type":"text","text":"這些干預的過程很多是算法人員進行的干預(如特徵構建、模型選擇、參數選擇等)，","attrs":{}},{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"在本篇文章中我們指的干預是指產品運營人員對推薦系統進行的產品策略、運營策略層面的干預，後面統一稱爲運營幹預。","attrs":{}},{"type":"text","text":"作者在《推薦算法團隊介紹》3.2.3節中對運營團隊對推薦系統的干預進行了簡單介紹，至少包括如下3種干預方式：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(1) 調整位置與展示；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(2) 干預具體的推薦結果；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(3) 對算法邏輯的干預；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這只是一部分干預方式。一般來說，運營策略的干預包括算法之前的干預、算法過程中的干預、生成推薦結果之後的干預3大類，我們在第3節會進行更細緻的講解。上面提到的運營人員的3種干預屬於對結果和過程的干預。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"大家瞭解了什麼是人工干預，下面我們來分析一下爲什麼要進行人工干預，人工干預到底有什麼目的和價值。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.2 爲什麼要進行人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦算法與人工調控的關係，類似於經濟學中的市場機制(market mechanism)和宏觀調控機制(macro-control mechanism)之間的關係。推薦算法根據用戶的行爲構建模型進行推薦，是用戶在平臺上的自然行爲的有效挖掘，這與市場機制是通過市場競爭配置資源(即資源在市場上通過自由競爭與自由交換來實現配置)的機制是非常類似的。人工調控是通過引進人工策略對推薦系統的運行加以優化、調節、引導，這與宏觀調控機制是在國家層面統一協調下以計劃、財稅、金融手段爲主，通過間接手段調控、引導市場活動也是非常類似的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"人工調控的作用是非常巨大的，也是非常有必要的。就拿肆虐全球的新冠病毒來說，中國進行了大量的政策和人工層面的干預，比如隔離、封路、封交通工具、醫療物資的定向供給等等，通過這些干預很好地控制了疫情，而西方國家由於政治體制的不一樣，很難做到中國這樣的有效干預，因此，疫情控制不容樂觀。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在推薦系統中，人工調控的作用同樣不容小覷，人工調控的主要目的是解決機器學習算法比較難解決的問題而進行的有效策略補充。一般說來，之所以進行人工干預，主要是滿足如下5類需求，下面我們來分別詳細介紹。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.2.1 用戶體驗","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統除了需要精準地挖掘出用戶的興趣，推薦用戶感興趣的標的物外，還需要滿足驚喜度、多樣性、新穎性等需求(讀者可以參考《推薦系統評估》瞭解更多推薦系統評估指標及細節)。這些需求很多都是比較抽象的，機器學習算法很難量化，因而很難做好，需要人工增加一些策略上的補充和控制。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在視覺上也可能需要根據特定情況進行調整，部分就是在雙十一這個特定時間點做的特殊配色和UI，圖中這一區域也是某寶個性化推薦的一種產品形態。通過在雙十一做這樣的調整，烘托出節日的氣氛，提升用戶的視覺體驗，讓用戶更有點擊的衝動。一般在重大節日、重大事件或者運營活動時，都可以做UI方面的調整，以營造氣氛，提升用戶感知度，優化體驗。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對熱點事件的把握、對內容的深度思考和深度關聯，有專業素養的編輯運營人員是強於機器的，通過整合專業人員的理性思考，並將這些思考整合到推薦系統中，有助於提升標的物的瀏覽、點擊、分發與轉化，最終提升用戶的滿意度。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.2.2 安全性","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在某些行業(如視頻、食品等)安全性至關重要，需要對待推薦的標的物的安全性進行人工把關，避免推薦不合適的標的物。比如一般電影都是有分級策略的，電影分級策略是指根據發行的電影中包括的性愛、暴力、毒品、粗俗語言等在內的成人內容的量和程度將其劃分成特定級別，並給每一級定義好允許面對的觀衆羣，以便運營人員有參照地、選擇性地進行內容運營，避免在不合適的時機給不合適的用戶推薦不合適的內容，起到促進所有觀衆身心健康的作用(讀者可以看看參考文獻3瞭解某條和臉book在這方面遇到的麻煩)。絕大部分國家和地區如美國、英國、日本、中國香港等都有完善的電影分級制度。在部分國家，電影分級制度不具有法律效力，但在行業內部具有約束力，只對觀衆起提示的作用，由觀衆實行自我保護。對於這類內容需要制定一些人工的策略，比如在家庭電視上，偏向成人的內容需要在晚上十點以後進行選擇性推薦等，避免小孩看到，影響兒童身心健康。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着手機攝像頭技術的成熟及智能手機的普及化，UGC內容是非常重要的一塊內容補充，現在主流的APP基本都提供了用戶上傳內容的功能，比如某手、某站、某寶等，內容的可控性變得越來越困難也越來越重要。UGC內容的安全性把控是這些產品的推薦系統必須要謹慎面對和有效控制的問題。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於這些涉及到安全問題的內容，雖然算法可以做到一定程度的識別(比如通過AI算法進行鑑黃等)，但是由於互聯網信息的非結構特性(特別是圖片、視頻、音頻等)，機器處理難度較大，準確率有待提升，最終還是需要人工來處理。不過機器可以提供很好的輔助，最終減輕人的工作量。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.2.3 商業價值","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有時需要人工制定一些推薦的策略，讓推薦系統可以獲得更大的商業價值。通過制定一些人工的策略，對具備不同購買力的人推薦不同價格的商品，從而獲得更多的商業價值。比如對於商務人士推薦頭等艙而對於一般的人士推薦經濟艙等。這樣做也是道德的，是可以提升用戶體驗的。而根據不同用戶對價格的敏感度不一樣，對同一件商品給不同的人提供不同的價格，從而讓企業獲得更多的商業利潤，這種做法可能就是不道德的，這就是所謂的“大數據殺熟”。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"獲取商業價值是公司生存的基礎，也是運營人員最重要的日常工作和行動目標之一，推薦系統作爲一種成熟的有商業化價值的技術手段，是運營人員在進行商業化決策中的有效工具。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.2.4 運營需要","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有時爲了運營的需要也需要進行人工策略的調控。通過人工策略的引導，讓資源達到某種程度的傾斜，最終讓整個生態更加健康成熟，良性發展。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"比如某寶等平臺方在引進某些新品類的商品時，需要對它們進行一定的資源支持和流量傾斜，這時在推薦策略上會對該類商品增加更多曝光的概率，最終讓該品類獲得更大的市場空間。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"某手作爲短視頻平臺，希望提供普惠的價值觀，所以在某手中的推薦策略是給普通人平等的曝光機會，而抖音更多的是運營導向，只有爆款內容才能得到好的資源，頭部效應更加明顯。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於某些重大事件、節日、運營活動等，也會採取一些策略來對推薦系統進行一定的干預和引導，以配合這些事件和活動。策略的調整既可以是算法策略，也可以是UI交互方式，種類可以非常豐富。某條在推薦中置頂了兩條推薦，都是與新冠病毒疫情相關的新聞，就是在當下特殊時期的人工干預策略。這種策略可以讓更多的人對疫情有更好地瞭解，起到信息普及和告知的作用。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於新功能、新模塊、新產品的引流，也是常用的運營幹預方式，比如某條推薦就會對問答、某音、小視頻等內容進行引流。推薦就是對問答(某空問答)和小視頻(某音)的導流。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.2.5 減少內容運營人員數量，降低人力成本","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果在推薦系統中可以增加足夠多的人工調控策略和手段，讓調控手段更加靈活多樣，將人的優勢和機器的優勢結合起來，那麼所有的位置都可以給到推薦系統(或者說，所有的人工運營板塊都可以整合算法能力)，這樣內容運營人員的工作量就會減少，也不需要這麼多內容運營人員參與了，從而可以降低運營人員的人力成本。這時只需要花更少的錢招聘少量足夠優秀的、對內容有深度理解的高級運營就可以了。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"總之，對於推薦系統來說，人工調控是非常有必要的，不管是提升公司自身利益，還是幫助提升用戶體驗，亦或是構建完善的內容生態，都有極大的價值。既然人工調控這麼有價值，那麼讀者一定想知道該從哪些方面進行調控，其實前面已經零碎地講了一些，下面一節我們來系統性地介紹。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.3 怎樣進行人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"前面兩節我們介紹了什麼是人工調控以及人工調控的目的和價值，本節我們來講解怎麼進行人工調控，有哪些調控的方法和策略，以及可以從哪些維度進行干預。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統是一項非常專業化的系統軟件工程，我們可以大致將推薦系統分爲6個大的模塊(階段)，分別是(生成)數據、(構建)特徵、(訓練)推薦模型、(生成)推薦結果、(渲染)前端展示的結果、(評估)推薦效果(參見下面圖5)。其實，人工調控可以在這6個階段中的每一個階段進行控制，下面我們一一說明。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/fa/faadae41d52d80b3f9a887bf5b6ac7dd.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"size","attrs":{"size":10}}],"text":"圖5：可以在推薦系統的6個階段進行人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.3.1 數據層面的人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦算法的數據至少包括兩大類，一類是用戶行爲數據，一類是標的物相關數據，有些公司還可以收集更多的其他數據：包括用戶相關數據、第三方數據等。一般用戶行爲數據用於推薦算法構建模型，具體採用什麼數據、怎麼使用數據構建模型都是算法工程師來決策，主要目的也是爲了推薦算法更加精準，這一塊一般運營人員不會進行干預。但是對於日誌打點需要收集什麼類型的數據，用戶的每類操作對於用戶是否產生點擊行爲的重要性等，運營人員可能會更有感覺。在這方面尋求運營人員的專業建議，對選擇合適的特徵、怎麼構建特徵等是非常有幫助的。爲了收集到更多的有價值的用戶行爲數據，構建更有意義的特徵，訓練出效果更好的模型，需要算法人員跟運營人員多溝通。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於標的物metadata數據，一般用於構建基於內容的推薦模型。可以採用文本、圖片、音視頻等信息來構建模型，算法人員基於目前已有的數據和技術能力來自己控制怎麼選擇和利用這些數據。其實很多數據是需要藉助運營人員來補充和完善的，比如最典型的標的物的標籤，就需要藉助內容運營人員的專業能力進行規範和統一化，構建完善的標籤體系。完善的標籤更利於構建質量好的內容推薦模型，像某條、某x等都有龐大的編輯團隊對內容進行標籤化。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"運營人員雖然對模型構建過程不需要進行干預，但對於能夠推薦什麼樣的標的物、在什麼範圍內推薦標的物，運營人員是需要進行把控的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基於特殊場景、安全性、標的物質量上的考慮，運營人員一般需要控制可以推薦的標的物池，在這個池子中進行標的物的推薦。前面提到的視頻安全性中，就有這樣的訴求。再舉個例子，在視頻行業中，在首頁推薦的視頻的海報圖一般要很清晰，很多老電影海報圖質量是比較差的，這時運營人員就可以選擇海報圖質量高的視頻(如果視頻metadata中沒有海報圖質量這個屬性，可以基於年代來粗略篩選，最近十幾年拍攝的視頻一般海報質量會比幾十年之前的好很多)，在首頁只能推薦這類有高質量海報圖的視頻。這種屬於正向選擇推薦池，反向操作也是可行的，剔除掉不滿足一定需求的標的物，在剩下的標的物中進行推薦，這屬於黑名單策略。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"一般提供內容的平臺方，內容來源於第三方，這時推薦系統的一個重要作用是要維護好整個生態的穩定平衡。需要保證提供優質內容的生產方獲得更多的曝光機會，而生產劣質內容(如低俗、標題黨、低質量、暴力、等)受到一定程度的限制和懲罰。這就需要採用一定的規則和策略對它們進行干預，這種干預可以採用人工調控的方式(比如選擇對什麼類型的標的物進行打壓，打壓的粒度等等，都需要人工制定規則和策略)。像快手這種提供普惠價值觀(讓每一個生活都被看見，見參考文獻1)的APP，需要保證每個人提供的視頻都可以被曝光，只要你的內容足夠優質，你也可以成爲熱門，這裏面肯定有很多人工的策略在裏面，這種普惠的價值觀其實就是一種最強的、價值觀層面的人工策略。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"數據和內容是整個推薦系統的核心基礎，推薦系統給予運營人員一定範圍的控制能力，並多結合他們的行業經驗和對內容深度把握的優勢，是可以讓推薦系統變得更好的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.3.2 特徵層面的人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在這一階段，通過特徵工程，我們基於數據構建出模型可以直接使用的特徵。特徵是給算法用的，運營人員在這方面的調控可能更多是建議性質的。運營人員與用戶距離近，更熟悉用戶，更懂業務，可能更知道哪些特徵對模型優化的指標是有正向價值的，哪些特徵是沒有什麼幫助的，特徵怎麼進行交叉更有價值等。總之，運營人員更知道怎麼構建合適的業務特徵。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果算法平臺可以提供一個自動化、可視化地構建特徵的工具，那麼運營人員通過適當的培訓是可以作爲(業務)特徵生產者的。下面的圖6就包含特徵構建的可拖拽模塊，有了這樣的工具，運營人員就可以發揮出他們業務敏感度和專業度的價值。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.3.3 模型層面的人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們知道企業級推薦系統的算法模塊一般包含召回、排序兩個階段(圖1紅色方塊中前兩個模塊)。對於召回階段可以有多種召回策略，內容運營人員根據自己對當下熱點內容的把握和深度理解，結合當下的熱點事件，可以人工整理一些優質的內容池，作爲一種召回策略，這種召回方式是可以整合到整個召回策略中的，從而實現了在召回階段中整合運營人員的專業能力。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基於產品發展或者商業化上的考慮，也需要對模型進行調控。前面提到的推薦系統中需要對新功能、新模塊、新產品進行引流。在信息流推薦中，還需要插入廣告，利用推薦來獲取商業利益。這兩種方式的干預都需要運營人員參與，這種干預涉及到多種類別內容的召回，算是對模型的一種干預。其實這裏也涉及到其他方面的干預，比如控制廣告的次數、控制對新模塊導流的比例等屬於結果層面的控制，在後面不再贅述了。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"另外，如果推薦系統工程體系做得比較好的話，各個算法組件是可以模塊化的，每個算法抽象爲一個算子，算子的輸入輸出採用一定的數據交互協議規範化，這樣就可以做到可視化、拖拽式地進行建模，作者團隊也是採用這種思路做的，最終實現了一套模塊化的推薦系統框架Doraemon，可以像搭積木一樣構建推薦算法體系(我們在《推薦系統的工程實現》中對Doraemon框架進行了深入的講解)，不過還沒有做到可視化、可拖拽的構建模型，這也不是作者團隊當前階段主要考慮的事情，因此價值不大。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"某雲的PAI機器學習平臺(包含推薦相關算子)，思路也是一樣的，並且做到了可視化、可拖拽式建模。做到這個程度了，運營人員只要懂算法的基本原理，就可以自己利用該平臺進行機器學習模型的構建與測試，可以自行完成機器學習模型的訓練，並部署到業務中。藉助AB測試能力，通過不斷迭代提升，最終產生業務價值。這種方式給不懂技術的運營人員提供了操作模型、調控模型的可能，通過技術賦能，讓人人都可以成爲推薦算法工程師。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.3.4 結果層面的人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在推薦結果層面的調控應該是最立竿見影、最直接有效的調控方式了。常用的調控方式有基於運營需要或者特殊事件置頂部分標的物，這在前面已經做過介紹。在視頻行業可能還存在監管的需要，推薦系統必須具備直接下線某個節目的能力，其他行業也會由於版權的問題，需要具備下線某個或者某一組標的物的能力。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這種控制的粒度不只限於某個標的物，還需要具備根據某個標籤或者特徵下線或者上線一批標的物的能力。比如由於侵權，平臺方需要下線某個歌手所有的音樂。對標籤或者特徵的控制除了上下線這種比較極端的情況，還可能需要控制優先級、數量等。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對結果層面的其他調控還包括對某類標的物人工定義不同的權重從而影響標的物的最終排序。這種調控常見於運營活動中，對於重點推的某個品類的標的物或者某個標的物生產方的標的物進行有針對性性的提權。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"除了控制權重，還可以對標的物的比例進行一定的控制，比如冷熱節目的比例、標籤的比例、類型的比例等。在家庭智能電視上的視頻推薦場景中，如果家裏有老人小孩的話，在首頁的推薦就需要保證老人喜歡看的抗戰類、小孩喜歡看的動畫片等多個類別的節目都存在。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上面提到的權重、比例等控制，一般會放到排序後的業務調控階段(參加圖1中紅色方框中的業務調控模塊)，通過算法工程師將這些業務控制邏輯轉化爲規則或者策略整合到算法體系中。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"爲了提升最終推薦結果的多樣性和驚喜度，有時也需要加入一些運營人員制定的隨機策略，對推薦標的物進行多維度的打散。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.3.5 展示層面的人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統鏈路中最後一環是UI展示，當前端獲取到推薦結果時，通過渲染引擎將推薦結果展示出來呈現給用戶，用戶就看到了推薦的標的物。對於視覺展示方法也是可以進行人工控制的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們在前文提到了一種調控方式，即是配色、配圖、文案等的調整。除了這些外，還可以進行的人工調控有如下幾種：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(1) 字體顏色與大小的調控；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(2) 相互位置的調控，包括兩個推薦標的物之間的距離，上下兩行之間的距離；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(3) 交互形態的調控；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(4) 展示的海報圖大小或者形狀的調控；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(5) 模塊位置的調控；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(6) 推薦標的物數量多少的調控；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"(7) 實時推薦中節目刷新的頻次、一次更新的條數的控制等；","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這些調控都是需要後端提供一套完善的內容編排系統、需要前端提供展示支持的，否則是無法做到的。這些調控也是有限度的，很多都依賴於所擁有的的資源，比如只做了兩種不同大小的海報圖，那麼只能支持這兩種海報圖之間的切換。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"下面我們拿某貓的產品舉一些例子讓讀者更好地理解。下面圖7是電視貓的首頁推薦，其中可人工調控的是左邊導航欄的標題、圖片等(例如，現在是疫情期間，增加了一個戰疫情的tab)，中間的海報圖有橫條的長方形還有豎直的長方形，橫條的長方形是豎直長方形的兩倍大小，對於某個節目是可以選擇這兩種UI的(只要這個視頻具備這兩類UI)。對於下面的興趣推薦，興趣推薦這四個字是可以調整的，下面興趣推薦有多少行、選擇什麼樣的海報圖、興趣推薦在產品中所在的位置等都是可以人工干預調整的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"某貓電影頻道的主題推薦，其中愛情和驚悚片是兩個用戶感興趣的主題，屬於主題個性化推薦，這裏面可以人工調控的有：主題的個數(這裏是2個主題)、每一個主題包含幾行(這裏是1行)、每一行包含多少個節目(這裏是6個)、主題在上下的位置等。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.3.6 效果層面的人工調控","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"企業在產品中引入推薦系統是有目的，期望通過推薦系統更好地進行內容分發、提升用戶體驗、促進用戶活躍、增強用戶粘性、產生更多的商業價值等(我們已經在《推薦系統的商業價值》中詳細介紹了推薦系統商業價值方面的知識)。這些目標在公司不同階段有不同的重要性和優先級，有些目標之間是互相沖突的，需要進行權衡。怎麼定義這些目標，在不同階段以什麼目標爲重，這些都是運營人員可以控制的。特別是在運營驅動的公司中，這些指標可能就是運營團隊來背的，因此一定是運營人員來決定的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上面我們從推薦系統業務流的角度從6個維度介紹了運營人員可以控制的部分及具體的控制策略。其實控制策略可以更廣泛，在產品形態等其他方面運營人員也可以進行干預，(產品)運營人員可以確定推薦產品形態的數量，決定哪些產品形態是最重要的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"調控的目的一定是優化用戶體驗、提升標的物曝光、點擊和轉化的，甚至是產生更多的商業價值。那麼人工控制到底起到了作用嗎？我們怎麼評估人工調控的價值呢？下面我們就來探討這個問題。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.4 怎樣評估人工調控的價值","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"前面提到了很多進行人工調控的方法和策略，我們期望人工調控是可以給推薦系統帶來巨大價值的，在第二節也講到了人工調控的價值，那麼我們怎麼來評估人工調控的價值呢？一般我們至少可以從如下4個角度來評估人工調控的價值。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.4.1 從宏觀指標上的趨勢變化來看待人工調控的價值","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於每一類產品都會有一些反應產品整體價值的較緩慢變化的宏觀指標，這些指標是公司非常看重的業務指標。拿視頻行業來說，日人均播放時長是一個比較重要的指標。我們可以基於過去一段時間的運營實踐和數據統計分析，確定某個需要人工干預的推薦模塊(或者在人工運營中整合個性化運營能力的模塊)的基準指標值，後續持續運營與優化，通過不斷提升基準值讓產品做得更好。如果在人工干預運營期間有算法迭代優化的話，這裏面可能需要區分出到底是人工運營產生的價值還是算法優化的價值了，這就需要藉助下面提到的AB測試。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.4.2 通過科學的AB測試來評估人工調控的價值","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們知道AB測試是一種科學的評估工具，廣泛運用於互聯網公司的產品迭代中，通過AB測試也可以很好地評估人工調控的價值。通過將用戶流量分爲AB兩組，一組是無人工調控的，另外一組是包含人工調控的，通過一段時間的用戶使用，收集用戶行爲數據，在關鍵指標(我們在第15章《推薦系統評估》有對評估指標詳細的介紹)上對比這兩組指標值的差異，就可以評估出人工調控對關鍵指標的影響和價值。AB測試一般用於評估比較複雜的人工干預，特別是對模型層面和特徵層面的干預，通過AB測試是比較好的評估方式。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.4.3 通過用戶調研來評估人工的價值","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上面1、2中評估的都是一些宏觀的商業化指標或者用戶體驗指標，實際上這些指標高並不等價於用戶體驗真正好。並且很多指標也是無法用1、2這種方法評估出來的，如安全性等。某種程度上用戶的真實體驗是至關重要的，它決定了用戶對你的產品的忠誠度和滿意度。這方面的洞察可以通過用戶調研來獲得。我們可以通過問卷、電話訪談、面對面交流等方式來了解用戶對推薦系統的看法，以及對不同人工策略的建議等。對這方面知識的掌握和了解，可以更好地幫助運營人員和推薦算法人員瞭解用戶的心理，從而迭代出更符合用戶心理預期的推薦產品。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.4.4 通過抽查來評估人工的價值","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"運營人員可以對推薦依賴的數據進行控制，比如運營人員打的標籤，可以通過抽查，或者交叉驗證等方式來評估內容運營人員人工生成的標籤數據的質量。對於其他可以直接影響推薦結果的控制(結果層面和展示層面的控制)，一般是可見即所得的，因此是可以直接在UI界面上查看調整效果的，這樣就可以進行人工肉眼評估了，如果有什麼重大問題是可以立馬發現的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.5 人工調控面臨的挑戰","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"人工調控是在整個推薦系統的生命週期中引入人的因素，特別是運營人員的知識和經驗，讓推薦系統更加靈活可控，更好地提升用戶體驗、獲取商業價值。但推薦系統是一個複雜的系統工程，怎麼引入人工調控，人工可以怎樣調控都會面臨着比較大的挑戰，在本節我們來進行簡單說明。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.5.1 知識層面的脫節，溝通不暢","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"一般運營人員更懂用戶、更懂業務，而對算法方面的知識有限，而推薦算法人員熟悉算法和工程體系，但是缺乏對用戶的瞭解和研究，離業務也比較遠，業務敏感度較低。要想讓推薦系統整合運營人員的經驗，首先在溝通上就可能存在障礙，算法人員很難讓運營人員理解調控的影響，運營人員也很難讓算法人員體會到人工調控的價值。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"算法人員與運營人員分屬於不同的團隊，也可能會存在一定的利益衝突，這更加劇了溝通困難。這需要公司引入各種管理和機制層面的保障，努力讓雙方可以多溝通，互相學習。算法人員需要了解一些業務知識，深入瞭解用戶，對運營的一些操作流程有所熟悉。運營人員同樣需要了解一些工程和算法的基本原理。只有這樣，雙方纔可能溝通順暢，最終迭代出一套真正有價值的可運營的推薦系統。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.5.2 推薦系統流程長、算法結構複雜，很難精確評估影響範圍","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統本身非常複雜，包含非常多的模塊，控制流程長。同時很多推薦算法，如深度學習等，本身就是一個黑盒模型，根據輸入的調整與變化很難知道對結果的具體影響。這兩點導致了很難知道調控影響的範圍和結果，也無法做到可見即所得。很多時候需要藉助多年的實踐經驗及AB測試等科學工具來評估運營控制的成效。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.5.3 提供一個供運營人員方便操控的界面是一件複雜的事情","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"前面一節提到運營控制涉及到推薦系統的各個模塊，方方面面都可以進行人工干預，那麼怎麼讓人工干預更好地落地就是很有挑戰的事。怎麼給運營人員提供一個可操作的界面是非常關鍵的，操作還必須要響應及時、流暢，這樣可以提升運營調控的效率，他們也更能夠理解怎麼去進行調整，也更容易進行調整，如果能夠做到可見即所得，那麼就是一個比較好的操作界面了。另外，爲了安全起見，運營人員的操作需要進行記錄，方便對操作歷史的追查，同時在操作出錯時還需要能夠方便地回退到操作前的狀態。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"雖然本節我們講了很多人工調控面臨的困難和挑戰，但是人工調控是非常有價值的，也是不可或缺的。在當前機器是無法取代人的價值的，最好的方式是人與機器良好協作，通過協同將兩者的價值最大程度的發揮出來。下面我們花一點篇幅來探討人和機器有效協作的問題。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"26.6 人與機器的有效協作","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"前面幾節我們講了人工調控的方法、價值、以及人工調控面臨的困難。目前AI技術還只能在簡單領域超越人類，在推薦系統領域，在很多方面(第三節的內容)還是需要人工的干預才能做得更好，人和機器只有更好地緊密配合才能產生最大的價值(參考資料2中提到某音利用人和機器對內容進行雙重審覈，這是人和機器協作很好的案例)。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"機器(推薦算法)最大的價值是可以做到全天候、無人工干預(當模型部署上線後基本不需要人工幫助了，在模型構建和訓練中是需要算法工程師參與的)、自動化、近實時地爲用戶提供個性化的內容推薦。機器也不會受到情緒的影響，判斷是完全理性客觀的。同時機器可以做到很低的邊際成本。這裏面提到的點都是人工不具備或者做的不夠好的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"運營人員最大的長處在於對行業知識的深刻洞察、對趨勢的判斷與把握、對複雜因果關係本質的分析、對人性的洞察、對情感的關懷、對跨領域概念的連接等，人更有創造力，這些都是機器很難做到、做好的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統是一種運營的工具，最終服務的是人，最瞭解人類自身的還是自己。因此，在推薦系統中整合人的因素、整合人的決策策略，可以讓推薦系統更加人性化，更有溫度，最終讓用戶可以感受到更多的人文關懷，從而增強用戶對推薦產品的情感聯繫，提升用戶體驗的同時，保證了高粘性。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"綜上所述，一個好的、有溫度的推薦系統，一定是人工和算法有效配合的產物，只有發揮兩者的優勢，互相補足，才能讓推薦系統更加完美。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"總結","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本章我們對推薦系統的人工調控進行了全面的介紹。我們介紹了什麼是人工調控：一切對推薦系統的人工干預都算人工調控，但本章指的調控主要聚焦在運營人員對推薦系統的干預。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"人工調控是非常有價值的，它的價值體現在用戶體驗、安全性、商業、運營需要、節省人力成本等多個維度。人工干預的形式是多種多樣的，我們可以對推薦系統流程中的各個方面進行人工干預。對於人工干預，我們需要明確它的價值，也需要知道做好人工干預是一件很困難的事情。人和機器都有各自的優缺點，只有很好地結合兩者的優勢，才能打造出更加精準、更加人性化的推薦產品。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作者認爲人工智能和人類在構造上的不同(機器是物理性構造、人類是生物化學構造)決定了機器永遠也無法替代人。在信息爆炸和科技快速發展的當代，那些善於利用和挖掘人與機器協同價值的公司才能在競爭中獲得生存的主動權。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"參考文獻","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"1. [書]《快手是什麼》","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2. 圖解抖音推薦算法","attrs":{}},{"type":"link","attrs":{"href":"https://mp.weixin.qq.com/s/EuQBuezHo5w7nBq0LQTsFg","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://mp.weixin.qq.com/s/EuQBuezHo5w7nBq0LQTsFg","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"3. [YouTube和今日頭條很委屈：色情暴力的鍋推薦系統該不該背] ","attrs":{}},{"type":"link","attrs":{"href":"https://mp.weixin.qq.com/s/jhuFVLOnbjO0-J27B3cYiA","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://mp.weixin.qq.com/s/jhuFVLOnbjO0-J27B3cYiA","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}}]}