推薦系統的價值觀(三十二)

{"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":"推薦系統作爲滿足人類不確定性需求的一種有效工具，是具有極大價值的，這種價值既體現在提升用戶體驗上，又體現在獲取商業利潤上。對絕大多數公司來說，提升用戶體驗的最終目標也是爲了獲取商業價值。我們在《推薦系統的商業價值》中已經詳細介紹過推薦系統的用戶價值和商業價值，相信讀者還記憶猶新。","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":"公司作爲社會經濟發展到一定階段的產物，獲取商業利潤是公司的本質特徵，也是公司賴以生存的基礎，這是合乎情理的事情。獲取商業利潤的方式有很多，其中推薦系統就是一種非常成熟的、具備極大商業價值的工具。鑑於推薦系統巨大的商業價值，幾乎絕大多數公司將衡量推薦系統價值的標準定義爲獲得更多的商業利潤，將推薦系統的短期商業價值作爲最重要的目標。這樣真的對嗎？對公司的長期健康發展真的有幫助嗎？推薦系統應當發揮怎樣的價值需要我們深入思考。","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":"對待推薦系統的態度直接來自於企業高層(創始人)的判斷和思考。在當前技術條件下，推薦系統還屬於弱人工智能，很多方面還需要藉助人的力量才能做得更好，推薦系統是由人構建的，推薦系統應該具備什麼樣的價值也是人類賦予的，","attrs":{}},{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"企業高層的價值觀直接決定了推薦系統的價值觀。","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":"本章我們就來談談推薦系統的價值觀，即我們構建推薦系統希望達到什麼目標，希望推薦系統可以做到什麼。具體來說，我們會從當前推薦系統存在的問題、推薦系統應當具有的價值觀、在正向價值觀指導下構建推薦系統的思路和方法等3個角度來講解推薦系統價值觀相關的背景、具體體現形式及方法論。期望讀者讀完本文可以從更加全面、更具人文關懷的角度來思考推薦系統對用戶、對企業、對合作夥伴、對社會的影響，而不僅僅將推薦系統看成一種快速變現的方法和手段。","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":"30.1 當前推薦系統存在的問題","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在今日頭條的示範和帶動下，國內絕大多數toC互聯網產品都具備了個性化推薦能力，個性化推薦系統在幫助公司獲取商業價值上立下了汗馬功勞。推薦系統經過這幾年的發展，也逐漸暴露了很多問題。這些問題如果不能很好地預防和解決，從小裏說，會影響用戶體驗，從大里說，會影響公司的業務發展，甚至對社會價值觀的塑造都有極大的副作用。","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":"在這一節我們就來梳理一下推薦系統存在的問題，只有清楚地瞭解推薦系統存在的問題，我們纔可以進行鍼對性地優化，讓推薦系統真正發揮應有的用戶價值、商業價值和社會價值。","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":"30.1.1 過度商業化","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統的危害首先來自於過度的商業化，很多企業漠視法律法規、走紅線，毫無道德底線，將推薦系統作爲斂財獲益的好工具。前幾年出現的魏則西事件，就是一個很好的案例(這個案例雖然屬於搜索問題，但搜索跟推薦本質是類似的，在更大的框架下，搜索問題等價於推薦問題，這種思路作者會《推薦算法工程師的成長之道》中進行說明)，當時在全社會引起了強烈的反響和聲討，百度也迫於輿論壓力下線了所有莆田系醫院的廣告。","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":"在很多APP的信息流推薦中，多少都存在一些尺度比較大的文章和圖片，在文章的評論區更是不堪入目，這種通過激發人的生物本能來獲得流量和粘性的方法是不齒的。這裏要提一下，之所以出現這樣的結果，有兩個原因。一是企業的價值觀就是這樣，是有意爲之的。另外一個原因是算法自身原因導致的，有些人主動搜索點擊這方面的內容，導致一直給他推薦相關的內容。","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":"合理的商業化可以讓平臺和標的物生產方更好地生存下去，是對平臺、標的物生產方、用戶都有正向作用的，有利於整個生態健康發展。過度的商業化可能引發極大的道德風險，對用戶也是非常不友好的。","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":"作爲一個有良知的企業，一定要謹慎考慮商業化，什麼樣的事情可以做，什麼事情不能做，心裏要有一杆秤，需要平衡好各方的利益，同時關注社會影響，努力讓整個生態持續、穩健、健康發展。","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","marks":[{"type":"strong","attrs":{}}],"text":"30.1.2 算法本身的缺陷和不足","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦算法本身的缺陷也會帶來風險，別有用心的標的物提供方會利用算法的特性和漏洞，攻擊推薦系統，對自己提供的標的物進行“刷榜”，惡意提升自己標的物的權重，獲得更多的流量。比如在標題中嵌入各種不實的關鍵詞、惡意刷好評、通過不正當手段刪掉差評、給競爭對手刷差評等等都屬於利用推薦算法的缺陷或特性來提升自己提供的標的物流量而打擊競爭對手。","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":"作爲提供推薦算法的平臺方，需要一直不斷的研究對策和調整算法，通過修正數據、優化推薦算法來評估標的物的真實價值和真正的受歡迎程度，這是一個長期的技術對抗過程。對這些惡意競爭的標的物提供方一定要有比較嚴厲的懲罰措施(限流、封號等)，規範整個標的物生產與消費鏈條。","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":"好的推薦算法需要滿足多個目標，除了商業價值、還有用戶體驗、生態繁榮等多類指標，往往很多目標之間是有衝突的，怎麼權衡多個目標之間的關係，是非常複雜的，連人類都覺得非常棘手，何況處在弱人工智能階段的推薦算法呢。目前多目標優化問題從算法實現上是比較困難的，有待算法理論的突破。","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":"當前推薦算法的可解釋性是比較棘手的問題，特別是像深度學習這種複雜的機器學習算法模型，基本是黑盒模型，可解釋非常差。人類本身是非常關注因果關係的，因果關係比較明確的概念人是更容易接受的。推薦系統沒有很好的可解釋性，用戶也很難信任推薦算法產生的結果。","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":"瞭解GAN(生成對抗網絡)的讀者一定知道，在某些情況下對算法輸入參數做很小的修改，就會產生非常不一樣的輸出結果，讓人不可思議。這導致利用有這類問題的算法構建的推薦系統，可能在某些情況下會給人推薦出極端不靠譜的結果，引起用戶的不適。目前推薦算法尋求的是一種全局最優解，而無法做到每個人都最優，這類問題也需要藉助算法理論的突破和實踐經驗的積累而逐步完善和提升。","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":"所有上面提到的算法本身的不足和缺陷都可能影響到推薦算法目標的達成。即使你的出發點再好，價值觀再正，由於算法自身“力不從心”，可能導致達不到預期的目標，甚至出現不希望看到的結果。","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":"30.1.3 標的物質量問題","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這裏提前對質量做一個解釋，如果標的物是實物，那麼質量就是我們通常意義的質量，比如這件衣服質量好不好，是不是仿冒品。如果標的物是虛擬物品，如視頻、文本等，質量是指內容的精良程度。","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","text":"導致推薦平臺上標的物質量的問題主要有三個方面的原因，一是過度追求商業利益導致質量下降，二是標的物創作/製造者自身專業素養和技術能力不足，三是平臺的規則間接導致了標的物質量下降。","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":"(1) 利益驅動導致的質量問題","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"人都是逐利的，爲了快速地獲利，標的物粗製濫造。對於文本內容的推薦，平臺方給內容創作者的收益是根據點擊量來計算的，因此導致很多內容標題黨盛行，內容創作者通過誇張的用詞、過度渲染來騙取用戶點擊。在電商推薦上，盜版、假貨橫行也是利益導致的後果。","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":"人的本性是對驚奇、情色、賭博等滿足人動物屬性的內容是天然喜歡的，很多內容創造者深諳此道，打擦邊球，製作出相關的內容，以迎合受衆的口味，從而獲得流量變現。","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":"(2) 標的物創作/生產能力不足導致的質量問題","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"不是每個標的物生產者都是專業的，由於自己能力及技術的限制，很多創作者/生產者不具備創作/生產優質標的物的能力，這時要麼創作/生產質量比較差的標的物，要麼進行剽竊/仿冒，這都會導致質量下降。","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":"平臺方必須制定一定的規則，鼓勵標的物生產者花更多的時間，通過提升技能來生產優質標的物，對優質標的物進行適當的流量獎勵，對劣質標的物進行懲罰和打壓。","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":"(3) 平臺規則導致的質量問題","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在(1)中我們就提到，現在很多平臺根據點擊量等指標來對標的物提供方進行獎勵(流量、好位置等)，這些規則比較簡單粗暴，容易滋生出質量差的標的物。","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":"標的物是整個推薦系統生態中最重要的一環，也是最基礎的一環，標的物質量的好壞直接關係到各方的利益和生態的平衡，因此平臺方必須重視標的物質量，需要通過技術手段及人工規則來保證標的物質量，吸引更多的優質標的物提供方進駐平臺。","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":"30.1.4 數據質量的問題","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦算法模型是基於數據來構建的，數據是算法的原料，當然對算法有非常大、甚至是決定性的影響。構建推薦系統的數據主要有用戶行爲數據、標的物metadata數據等。","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":"用戶行爲數據在日誌上報時可能存在字段錯誤(比如錯誤地將時間戳賦值給了播放時長字段，導致播放時長很大，不符合邏輯)，用戶由於沒有關電視而去幹別的事情忘記了，導致一直連續播放十幾個小時，還有測試等髒數據的混入，黑客或者競爭對手的攻擊產生的垃圾數據，程序bug引起的錯誤數據等等。","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":"如果是UGC平臺，由於不同標的物提供方填寫的標籤等信息格式及定義不一樣，導致數據質量低。重複數據也是一個比較大的問題，新聞中不同渠道來源的內容有可能是對同一事件的報道(因此內容是重複的)，淘寶中不同商家生產一樣的物品，這些都可以產生重複。","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":"這些低質量數據，如果不加以處理，就會影響算法的效果，對最終的總體目標是有極大副作用的。","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":"30.1.5 過濾氣泡問題","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"個性化推薦存在被人所詬病的“過濾氣泡”(Filter Bubble)問題。Filter Bubble 的概念由 Eli Pariser 提出，他認爲個性化推薦算法會基於用戶的各種信息向其推薦可能感興趣的內容，長此以往用戶會因接收不到與自己相左的觀點，停留在自己的文化和認知泡泡中。另外一個類似的概念是信息繭房，讀者可以學習參考文獻3，進行更深入的瞭解。","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":"相信刷過某條推薦的人有這個感受，當你點某個感興趣的內容時，系統一直會給你推送相關的內容，讓你一直停留在該內容相關的內容中出不來，就像滑進了一個漩渦。前幾年的今日頭條這個問題更加嚴重，這即是形成了過濾氣泡。","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":"一味推薦你喜歡的標的物，你沒有接觸過的、不熟悉的優質標的物可能被排斥而逐漸邊緣化，導致最終只推薦你所熟知的標的物。","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":"推薦系統這種給用戶“灌輸”標的物的方式，導致用戶被動接受，減少了用戶的決策成本，最終用戶不願意決策，導致過濾氣泡的存在。","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","marks":[{"type":"strong","attrs":{}}],"text":"30.1.6 用戶隱私","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統通過收集用戶行爲，並基於用戶行爲給用戶推薦感興趣的標的物。如果用戶看的是一些比較隱私或者不希望別人知道的內容，那麼推薦系統也獲得了用戶的這些特殊嗜好，因此也會給他推薦這類內容，這對用戶來說就是隱私的泄露。舉個例子，比如用戶一直用某條看比較暴露的美女短視頻，那麼當他在跟朋友喫飯刷某條時，容易讓別人看到他的這種他不希望別人知道的喜好。","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":"用戶的偏好就是用戶的隱私，推薦系統知道了用戶的偏好，有些企業就可能利用這些偏好進行不道德的商業化(比如我們知道的大數據殺熟)，從而傷害用戶。","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":"30.1.7 難以量化的目標","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統的價值除了用CTR等比較容易量化的指標來衡量外，還有很多比較難量化的指標，比如多樣性、驚喜度、社會價值等，而當前的機器學習算法還只能處理容易量化的目標函數，對這些比較難量化的目標無能爲力。","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":"上面提到的這些問題都是在推薦系統發展過程中出現並逐步凸顯出來的，只有對這些問題及背後的原因有一個比較清晰的認知，我們纔有機會構建更加優質的推薦系統(讀者可以查看參考文獻1、2進一步瞭解推薦系統存在的問題)，那麼什麼是優質的推薦系統？推薦系統應該具備什麼樣的價值呢？這就是我們下節要討論的主題。","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":"30.2 推薦系統應當具有的價值觀","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在上面一節我們提到了推薦系統存在的問題，這些問題有些是技術工程問題，有些是人的決策導致的問題(如過度商業化就是公司管理層的決策導致的)，這些問題對發揮推薦系統的價值是有極大影響的。","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":"其實學術界關於AI的倫理討論已經非常多了(見參考文獻7、8)，推薦算法作爲AI的一個應用領域，也一直存在這類問題，只不過在工業界大家的關注點都在算法創造商業價值上，而忽視了很多道德風險和人文關懷。","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":"除了商業化指標外，推薦系統應該還需要發揮哪些價值呢？下面作者結合自己的理解和感悟，從如下4個方面來說明推薦系統應該具備的價值觀。","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","marks":[{"type":"strong","attrs":{}}],"text":"30.2.1 追求正當的商業價值","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"獲取商業利益是公司生存之本，對公司來說商業化確實是非常重要的，也是管理層需要持續關注和努力的方向。但不能通過利用技術手段來惡意傷害用戶、造成社會負面影響來獲得商業利潤。","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":"大家都知道的大數據殺熟(基於用戶的行爲和購買力，故意給用戶推薦比較貴的東西，或者單獨給該用戶定一個高價)就是一種不正當的獲取商業利益的方式，這裏面還有故意欺騙用戶的成分。","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，追求商業價值也不應該違背國家法律、法規、公序良俗。比如通過盜鏈或者破解正版視頻網站的播放器或者鏈接，聚合到自己的視頻平臺並通過付費的方式獲取商業利益，這就是違反法律的，侵犯了視頻網站的版權。生產假冒僞劣產品在電商平臺上售賣也是非法的。故意推薦一些惡俗的內容給用戶，通過提升流量來獲得更多的廣告收益這也是不道德的。","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":"這些通過不正當的手段來進行商業化的行爲，最終會讓企業站在輿論的風口浪尖上，搞不好會葬送企業的前程。","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":"30.2.2 關注用戶體驗與成長","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統不但要從技術角度去思考，還要從人文社科方向尋找靈感。推薦系統的神經元裏要植入人文關懷，推薦系統要以人爲本，迴歸人性，建立與人的信任關係。因此推薦系統需要關注用戶體驗和用戶成長。","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":"首先，推薦系統需要給用戶足夠好的體驗，界面美觀簡潔、操作方便，能夠給用戶提供優質的推薦結果，讓用戶在交互的時候是放鬆的，同時推薦系統需要一定的機制避免用戶沉浸其中。國內目前有一些餐廳對顧客點菜是有要求的，會根據就餐人數來了解你點菜的情況，不讓你多點菜，避免浪費，我覺得這就是一個非常好的控制手段。現在很多遊戲需要身份信息登錄，並且對使用時長有控制，快手也有青少年模式，這些都是人文關懷的舉措。推薦系統也需要在這些方面有所作爲。","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":"推薦系統不光要有CTR等與商業化相關的指標，更應該包含多樣性、驚喜度、新穎性等與用戶體驗相關的指標，用戶如果能夠獲得超出預期的體驗，一定會更加信任你的產品，信任你的推薦系統。","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","text":"給用戶推薦多樣性的內容，而不是一味滿足用戶的興趣，可以讓推薦系統突破過濾氣泡的魔咒。通過用戶在新內容形式(用戶過去沒有探索過的)的反饋來拓展用戶新的興趣點，這對用戶是一個學習成長的過程，讓用戶可以獲得新的知識、新的體驗，是對未知領域的一種探索。","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":"除了用戶體驗外，推薦系統要對用戶進行引導而非灌輸，給用戶足夠多的操作方式讓用戶進行自主探索，激發用戶的探索欲，而不是被動”享受“，這也是培養用戶好奇心的過程。這裏面可能會涉及到跟用戶的多倫交互過程，這是目前的工業級推薦系統不具備的能力，也是未來可行的發展方向。","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":"30.2.3 考慮整個生態系統的繁榮與長期發展","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統涉及到平臺方、用戶、標的物提供方三方，缺一不可，這三方組成一個小的生態系統，互相之間合作與依賴。標的物提供方是內容創作方，是整個生態系統的核心，是系統賴以生存的基礎。用戶是該生態系統的消費者，與平臺是一種價值交換的關係，用戶通過在平臺上消費標的物來養活整個平臺(包括平臺方和標的物提供方)，而平臺方提供高效、便捷、精準匹配的信息分發渠道/通道來連接用戶和標的物提供方。","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":"上面是生態系統三方之間的價值關係網的梳理，好的推薦系統一定要考慮整個生態系統的平衡與穩健發展，缺少任何一方，整個系統都將無法運轉。","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":"我們不光需要這個生態系統短期繁榮，我們更應該努力促進系統長期穩定發展。這就要求平臺方有延遲滿足感，管理層不能只做一些只考慮短期利益的事情，而要關注長期利益，考慮多方共贏，構建正和博弈系統，只有這樣平臺方纔能基業長青。","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":"30.2.4 弘揚正向的價值觀","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦產品一般是面向C端用戶的，在移動互聯網紅利逐漸消失的當下，每個人都可以通過智能手機連接到網絡。網絡對人的工作生活的影響是全方位的，像某寶、某條、度娘等APP日活(DAU)都超過幾億，對社會生活及信息輿論有極大的影響。這些產品都將推薦作爲核心功能點，放到最核心的位置，毫不誇張地說，推薦做得好不好，直接關係到產品的生死存亡，也直接影響者整個社會輿論和社會價值觀。像這類影響力巨大的產品，一定要有足夠的社會責任感，在商業利益之外，必須弘揚正向的社會價值觀。","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":"像百度搜素提供的競價排名，我個人覺得就是一種非常不好的價值觀。這讓出錢多的人獲得了好的流量，如果這個出錢多的人投的廣告關聯的產品或者服務是不好的，會對個人和社會造成極大的危害，魏則西事件就是這種有害性的一次集中爆發。","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":"作爲一個有良知、正義感的企業，在打造推薦系統時一定要心正，弘揚正向的價值觀，起到價值宣導的作用，在某種程度上要做到政治正確、道義正確，而不能完全被商業化的思維控制。","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":"某條在推薦中置頂了兩條與疫情相關的信息，這天剛剛是武漢解封之時，在這個關鍵的利好時間點將這一重要信息告知大衆是非常有價值的。而作爲反面的案例，2020年4月8日手機百度APP的推薦模塊被迫下架整改，在這之前的幾天，我在用百度的推薦模塊時，確實看到很多推薦下面的評論都是“很黃很暴力的”，並且有幾個評論的賬號直接與色情相關，大面積出現這種情況不被約談整改纔怪呢。","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":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這一節我們從推薦系統應該具備的價值觀的角度來講解構建推薦系統應該思考的方向。當然，不同行業由於數據來源、服務的對象、提供服務的類別不一樣，在具體怎麼體現推薦產品的價值觀時是需要謹慎思考的。","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":"有了牢固可靠的價值觀，我們就要利用價值觀來判斷一切行動是否值得做、能夠做。怎樣基於上述價值觀來構建推薦系統，我們需要從哪些維度來努力和思考，這就是下一節要講的內容。","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":"30.3 在正向價值觀指導下構建推薦系統的思路和方法","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有了正確的價值觀，這些價值觀就是指導我們構建推薦系統的指導原則，我們在遇到任何困難時，就有章法可循，不會亂了方寸。基於30.2節提到的價值觀來構建推薦系統時，我們可以從如下幾個角度來努力，最終減輕、避免、甚至解決30.1節提到的推薦系統存在的問題。","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":"30.3.1 努力提供高質量的標的物，標的物是整個平臺的核心","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"通過前面的分析，我們知道標的物纔是(提供推薦系統服務的)平臺方的核心競爭力。好標的物是獲得用戶喜愛的前提，只有標的物質量好，用戶才願意留下來，才願意在平臺上付費/消費，整個生態系統纔能有效運轉。標的物就像生態位中的草，只有水草豐茂，才能供養牛羊。","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":"平臺方一定要有正義感，要制定一系列規則和策略對提供優質標的物的供應方進行鼓勵和獎賞，對提供劣質標的物的第三方進行打壓和懲罰。在這裏可以充分發揮機器學習算法和人工審覈干預的作用，提升標的物監控和審覈的效率。可以先用機器學習算法召回質量可能有問題的標的物，再借助人工二次審覈，既減輕了工作量，標的物質量又有保證。","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":"通過不斷的優化完善，當產品具備了一套比較完善的標的物質量標準和引導規則時，這套規則會引導着標的物生產鏈路朝着更好更健康的方向發展，這時系統是具備一定的自愈能力的，最終會進化出一個更加健康的生態。","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":"作爲推薦系統推薦給用戶的標的物，好的標的物對用戶體驗、對用戶信息獲取等各個方面都是有好處的。有了好的標的物及標的物metadata數據，推薦系統也可以更好地獲得構建模型的原材料(基於內容的推薦算法需要標的物的metadata信息)，最終構建出更加可靠的推薦系統。","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":"30.3.2 注入人的因素，讓推薦系統更有溫度、更有情感","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"機器學習算法是對現實的一種抽象和簡化，雖然可以解決很多問題，在很多方面甚至超越了人類，但算法還是有短板的，在複雜情況下的決策，在涉及到藝術與情感方面的問題時是根本無法跟人相提並論的。","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":"推薦系統需要朝着更有人情味、更有溫度的方向發展。這一方面需要算法能力的提升(比如微軟的小冰試圖打造一個更加智能化、人性化的助手)，需要獲取更多的數據和信息(特別是場景化信息、上下文信息)，以期更好地理解用戶當前的意圖，這是非常難的，也是一個需要長期努力的方向。","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":"另外一種更加實用化、更加可落地的策略是通過在推薦算法中整合人工策略和邏輯，讓推薦系統具備一定的人文關懷(我們在第26章《推薦系統的人工調控策略》中對人工調控的策略和方法進行過非常詳盡的介紹)。通過不斷嘗試，調和AI算法與人的智慧，相互約束和補位，將\"人\"的價值發揮到最大，讓推薦引擎不只是迎合用戶，而是嘗試引領用戶、感動用戶，從而做到真正的以人爲本。","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":"像圖1中置頂疫情信息、當你看了很長時間系統給你提醒讓你休息一會兒、給你推薦你期望學習的新知識點等等，這些都以人爲本的做法。什麼叫做以人爲本，這本身就是一個比較難以界定的概念，更難以量化，需要算法工程師和產品經理多琢磨，更好地瞭解人的社會化需求和情感需求，而不僅僅是滿足人的生物本能需求。","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":"30.3.3 關注數據質量、關注用戶隱私","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們在30.1.4節中提到了數據質量問題，數據作爲推薦算法的原材料，質量的好壞直接決定了算法的效果，不好的數據會讓推薦模型偏離原來既定的價值取向，走向“邪路”，比如一個提供假冒僞劣產品的淘寶店主，如果通過不正當手段來刷好評、刷流量，讓自己的店鋪和商品排名靠前，獲得更多收益，這就損害了消費者的利益，對平臺的健康發展也極爲不利。因此，數據質量在推薦系統中起着至關重要的作用。","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":"在數據處理上，需要通過有效的手段過濾掉不合理、不合法的數據，識別潛在的針對產品的惡意攻擊，不讓外部干擾污染整個數據源。","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":"另外一個比較重要的問題是關於用戶隱私的，這類問題屢屢發生(今年上半年萬豪酒店出現了泄露520萬住客信息的極大安全事件)，隨着人們安全意識的覺醒，大家對隱私問題會越來越重視。推薦系統需要蒐集用戶各方面的信息(用戶自身的信息以及用戶的行爲偏好信息)來獲得精準的推薦，推薦系統利用的信息是非常有隱私性的。","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":"處理好隱私問題，除了需要企業加強數據安全外，另外一種可行的方法是給用戶一定的控制權，讓用戶自由選擇是否可以讓企業收集行爲數據、利用行爲信息進行推薦。即使用戶同意了企業可以使用自己的行爲數據，用戶在以後的任何一個時間點有權要求企業刪除對用戶興趣的建模，迴歸到信息爲零的狀態，甚至可以再次拒絕讓企業收集信息。企業獲得了用戶的這些偏隱私的信息也需要用在正道上，不能利用這些信息獲取一些非法的收益，比如將用戶的興趣偏好售賣給其它第三方或者利用這些偏好進行不正當的、有傾向性的營銷。","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":"目前中國在用戶隱私上的法律制度還不夠健全，隨着這方面問題的暴露及處理這類事件經驗的積累，未來一定會有比較規範的關於互聯網隱私方面的法律出臺的。今年7月2日國家出臺了《數據安全法(草案)》，這算是從國家層面開始正式將數據安全列入法律管控，數據安全有法可依了。企業應該提前做好預防和規劃，重視用戶的信息安全，採取合適的保護用戶敏感信息的方法。","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":"作爲用戶，也需要特別關注自己的信息安全。在註冊一個新APP時，謹慎選擇授權，比如授權訪問位置、通信錄、攝像頭、麥克風等。","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":"目前很多APP雖然說有用戶隱私協議，但是基本都是霸王協議，用戶沒有選擇權，用戶只有放棄自己的隱私才能使用該APP，這不是真正的將隱私權交給用戶。主動讓用戶自己選擇是否可以利用用戶數據，看似對收集數據不利，其實這是一種非常好的人文關懷舉措，會極大地提升用戶的好感，增加用戶對產品的信任度。同時這也倒逼企業通過其它的技術手段在保護用戶隱私的同時對用戶行爲建模，比如下面要提到的聯邦機器學習。","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":"Google在2016年提出的聯邦學習(參見參考文獻4)就是一種很好的嘗試，聯邦機器學習是一個機器學習框架，能有效幫助多個機構在滿足用戶隱私保護、數據安全和政府法規的要求下，進行數據使用和機器學習建模。聯邦機器學習在推薦系統上的應用也是一個非常值得探索的方向，在這方面已經有企業在嘗試和應用了，其中參考文獻5中就有關於聯邦推薦算法在微衆銀行中的應用，同時微衆銀行還開源了全球首個企業級的聯邦學習框架FATE(見參考文獻6)，並且內置了聯邦推薦學習算法，有興趣的讀者可以作爲很好的學習材料。","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":"30.3.4 幫助用戶獲取更多的差異化信息，讓用戶更好地學習成長","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統是一種被動獲取信息的技術手段，如果一直推薦用戶感興趣的標的物，會存在過濾氣泡的問題，這不利於用戶獲取新的信息和知識。推薦系統一定要打破這種固有思維邏輯，給用戶提供多樣化的推薦，雖然這樣可能短期對商業化有負向影響，但是這是讓推薦系統具備人文關懷的必要，作者相信這種做法長久來看是“有利可圖”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":"其實，每個人由於出生、見識、閱歷、地理位置、社交固化等各種原因，接觸到的知識是有限的，每個人都生活在一個自己親手打造的深井中，我們只能觀察到一片有限的天空。推薦系統提供的信息如果只滿足了人的生物本能需求，是遠遠不夠的。算法要讓個人獲得突破，就一定要給用戶推薦用戶不知道的東西，引發用戶的好奇心和求知慾，激發用戶對未知領域的興趣。通過讓用戶獲取差異化信息，讓用戶對世界有更多維度的瞭解，這有利於用戶的認知升級和學習成長。作爲一個有社會責任感的企業，需要對用戶成長負責，必要的時候需要引領用戶，而不是一味迎合用戶。","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":"給用戶提供多樣化的推薦，可以在算法中增加隨機策略，給用戶一些探索性的推薦結果，如強化學習中的EE(Exploration and Exploitation，探索-利用)策略就是一種非常好的方法。另外，利用知識圖譜來拓展用戶的興趣空間也是一個比較有前景的研究方向。","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":"30.3.5 給用戶更多的控制權，讓用戶能夠自主抉擇","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們在前面也提到了用戶可以選擇系統是否可以收集用戶的行爲信息，用戶可以決定是否清空當前興趣重新給用戶推薦，這些都是用戶的選擇權。","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":"在與推薦系統進行交互的過程中也需要給用戶更多的選擇權，比如是否可以主動選擇過濾掉某類內容，甚至未來推薦系統可以讓用戶自己制定給自己的推薦規則，基於該規則來爲用戶生成個性化推薦。","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":"在推薦的交互方式及展示方式等方面也需要給用戶提供足夠的控制空間，這樣才能真正做到以人爲本，以用戶爲中心。","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":"推薦系統只是一種獲取信息的方式，產品需要提供像搜索、篩選、導航等其他讓用戶主動獲取信息的方式，並鼓勵用戶從多個渠道來獲取信息，這樣可以從另外的角度來影響推薦系統的行爲，有效避免過濾氣泡帶來的問題。提供更多的用戶主動探索的可能，也有利於用戶自我學習成長。","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":"30.3.6 更加多元化的目標，在用戶、標的物提供方、平臺之間達到利益平衡","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們在30.1節中就已經講到，目前很多公司做推薦系統最核心的目標是商業化，通過推薦獲得商業利益，算法優化的是CTR點擊率。過度的商業化，導致推薦系統目標單一，唯利是圖。短期來看對企業是很有利的，但是這種策略是損害公司長期利益的。","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":"就像30.2節中提到的，推薦系統需要追求在多目標下的均衡發展。推薦系統涉及到的參與方有三個：用戶、標的物提供方、提供產品與推薦服務的平臺方，這三方都有各自的利益訴求，推薦系統一定要照顧到三者的利益。","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":"對於用戶來說，除了給用戶提供精準的推薦外，還需要考慮到用戶多種類的知識獲取，拓展用戶的認知空間，給用戶更多的自主控制權，保護用戶的數據安全與隱私。","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":"對於標的物提供方，推薦系統在算法機制上就要鼓勵優質標的物提供方創造更好的內容，給優質標的物更多的流量支持，適當的時候還需要增加人工的策略和規則，彌補算法做的不夠好的點。另外，還需要在平臺中制定適當的規範來對標的物提供方進行約束，讓標的物提供方知道平臺是鼓勵生產/製造優質標的物而打擊粗製濫造的。","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":"對於平臺方，商業利益需要考慮，但是不能操之過急，要細水長流。平臺方爲了生存，需要很早就考慮商業變現，但是一定要給自己制定一些規則和發展規劃，對自己加以限制和控制，什麼事情可以做，什麼事情不能做，要有自己的原則和道德底線(比如不做醫療、保險等的廣告)。當你在推薦系統迭代過程中始終考慮到了用戶利益和標的物提供方利益，你的推薦產品纔是健康發展的。","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":"上面提到的6個思路及其中的一些方法更多的是指導性的，具體怎麼實施還需要根據自己產品和行業特性進行有針對性的選擇和細化，並有所側重。","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":"推薦系統由於具備巨大的商業價值而得到企業界的追捧，但商業價值只是推薦系統價值的一部分，推薦系統有非常多的目標值得去思考和優化。作者希望推薦系統的從業人員學會更多的人文關懷。","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":"推薦系統本身是沒有價值觀的，是我們人類賦予了它鮮活的生命，讓它具備了某種價值取向。推薦系統需要克服非常多的問題，弘揚更多的社會價值，這就要求人類將自身的價值觀通過規則、算法甚至人工介入的方式更好地整合到推薦系統中。","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":"本章我們基於目前推薦系統存在的問題，結合作者自己的理解和感悟，比較主觀地引出推薦系統應該具有的價值觀，並基於該價值觀來說明從哪些維度來構建推薦系統，可以讓價值觀得到最好的體現。這裏面很多價值觀是比較抽象，不易於用算法來解決的，因此人的作用就凸顯出來，人與算法的有機結合纔是推薦系統的未來迭代範式。","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","text":"抽象的價值觀怎麼通過推薦系統表達出來，這是一個非常值得思考的問題，也是當前推薦行業非常缺失的，大家更多地關注了推薦系統的商業價值而忽略了其他價值。","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":"本章通過作者自己的思考，對推薦系統應當具備的價值進行了全方位的梳理，希望給推薦算法從業者、算法產品、運營人員、企業管理者提供不一樣的思考視角。","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. [推薦系統有哪些坑？] ","attrs":{}},{"type":"link","attrs":{"href":"https://www.zhihu.com/question/28247353","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://www.zhihu.com/question/28247353","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2. [推薦系統有什麼危害？] ","attrs":{}},{"type":"link","attrs":{"href":"https://www.zhihu.com/question/385821370/answer/1135280697","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://www.zhihu.com/question/385821370/answer/1135280697","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"3. [如何理解“信息繭房” ?] ","attrs":{}},{"type":"link","attrs":{"href":"https://www.zhihu.com/question/58195189/answer/684118964","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://www.zhihu.com/question/58195189/answer/684118964","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"4. [聯邦機器學習] ","attrs":{}},{"type":"link","attrs":{"href":"https://baike.baidu.com/item/%25E8%2581%2594%25E9%2582%25A6%25E6%259C%25BA%25E5%2599%25A8%25E5%25AD%25A6%25E4%25B9%25A0/23618046?fr=aladdin","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://baike.baidu.com/item/%E8%81%94%E9%82%A6%E6%9C%BA%E5%99%A8%E5%AD%A6%E4%B9%A0/23618046?fr=aladdin","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"5. [聯邦學習用於推薦場景] ","attrs":{}},{"type":"link","attrs":{"href":"https://zhuanlan.zhihu.com/p/97826564","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://zhuanlan.zhihu.com/p/97826564","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"6. [工業級聯邦學習開源框架FATE] ","attrs":{}},{"type":"link","attrs":{"href":"https://github.com/FederatedAI/FATE","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://github.com/FederatedAI/FATE","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"7. [算法倫理：現狀與困境] ","attrs":{}},{"type":"link","attrs":{"href":"https://zhuanlan.zhihu.com/p/108567166","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://zhuanlan.zhihu.com/p/108567166","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"8. [人工智能的六大倫理原則] ","attrs":{}},{"type":"link","attrs":{"href":"https://zhuanlan.zhihu.com/p/72952747","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://zhuanlan.zhihu.com/p/72952747","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":{}}]}]}