領英PYMK推薦系統優化：爲用戶帶來更平等的人脈

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作者 | Qiannan Yin"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"共同作者 | Qiannan Yin、Yan Wang、Divya Venugopalan、Cyrus Diciccio、Heloise Logan、Preetam Nandy、Kinjal Basu、Albert Cui"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"編譯 | 王強"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"編輯 | 劉燕"}]}]},{"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":"最近十年來，個性化推薦技術可能是互聯網世界中發展最快、影響最深遠的技術種類之一。從流媒體視頻到電商購物，從新聞瀏覽到私房音樂，個性化推薦已經深入到了社會生活的諸多領域，極大改變了人們的生活方式與行爲習慣。"}]},{"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":"但飛速擴散的推薦技術也帶來了很多負面效應，引發了多種層面的擔憂和反對聲。相關問題包括個性化推薦強化的“信息繭房”效應、推薦系統有意無意融入的社會性偏見、推薦系統造成的話語權不公等等。"}]},{"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":"有些問題的根源並非是技術層面的，但無論是批評家還是技術社區都希望這些問題能通過技術自身的進步來解決 — 就像是人類工業革命以來曾經解決的無數問題一樣。"}]},{"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":"可喜的是，已經有越來越多的企業，包括很多大量應用推薦系統的頭部IT企業意識到了上述問題，並開始在這一主題中投入資源，全球知名的職場社交平臺領英就是其中的一個代表。"}]},{"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":"strong"}],"text":"領英使用一套推薦系統爲用戶推送適合他們的崗位與社交信息，而最近這家公司的工程團隊對推薦系統做了諸多改進來改善系統的公平性。"}]},{"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":"2021年8月，領英在工程博客上發表了一篇文章，介紹了相關工作的細節。InfoQ獲得了領英工程團隊的許可翻譯這篇文章，並就一些問題採訪了團隊，得到了頗具價值的答覆。下面將譯文全文與採訪內容登出，希望能爲國內同仁提供有價值的參考。"}]},{"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":"領英專注於營造更加公平的環境，爲全球勞動力的每一位創造經濟機會。爲了達成這一目標，一項關鍵環節是讓平臺用戶能夠互相溝通，在線上建立起自己職業的人脈關係網。"}]},{"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":"我們很清楚"},{"type":"link","attrs":{"href":"https:\/\/engineering.linkedin.com\/blog\/2021\/professional-network-checklist","title":"","type":null},"content":[{"type":"text","text":"人脈的重要性"}]},{"type":"text","text":"，因爲它們可以轉化爲有形而高質量的職業機會。考慮到人脈在領英用戶生活中所發揮的重要作用，我們一直在尋找可以幫助改善所有用戶的人脈體驗的方法，例如我們在努力"},{"type":"link","attrs":{"href":"https:\/\/blog.linkedin.com\/2019\/september\/26\/closing-the-network-gap","title":"","type":null},"content":[{"type":"text","text":"縮小人脈鴻溝"}]},{"type":"text","text":"，還在"},{"type":"link","attrs":{"href":"https:\/\/engineering.linkedin.com\/blog\/2021\/professional-network-checklist","title":"","type":null},"content":[{"type":"text","text":"分享很多有數據支持的建議"}]},{"type":"text","text":"，幫助大家擴張自己的人脈。"}]},{"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":"strong"}],"text":"而我們最近的一項工作重心是優化用戶體驗、創造更多平等的聯繫機會，這項工作涉及人們在領英上建立人脈時用到的基礎功能之一：我們的“"},{"type":"link","attrs":{"href":"https:\/\/www.linkedin.com\/help\/linkedin\/topics\/6096\/6118\/29","title":"","type":null},"content":[{"type":"text","text":"猜您認識"}]},{"type":"text","marks":[{"type":"strong"}],"text":"”（People You May Know，PYMK）推薦系統。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"什麼是“猜您認識”系統"}]},{"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":"在領英上建立人脈時，用戶會用到的一項基本功能是一個名爲“猜您認識”（PYMK）的推薦系統。"}]},{"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":"PYMK是領英平臺的一項老功能，由我們最早開發的一些機器學習（ML）算法提供支持。PYMK的目標是幫助用戶與可能加強他們職業人脈網絡的其他用戶建立聯繫。"}]},{"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":"領英用戶可能出於多種原因希望與平臺上的其他用戶成爲好友，例如聯絡過去的同學或同事，或從好友那裏獲得工作機會或職業建議。你可以在領英的“人脈”選項卡上看到這個模塊。"}]},{"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":"PYMK主要使用經濟圖譜（Economic Graph）和平臺交互等數據來挖掘特徵，並使用ML算法提出相關建議。"}]},{"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":"具體來說，它使用線性和非線性模型的組合來估計兩個用戶之間加爲好友的傾向。這個概率產生一個P(connect)分數，PYMK隨後參考該分數向用戶推薦新的潛在人脈。"}]},{"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":"然而，與其他所有人工智能系統一樣，影響這套系統準確性的一大要素是外部社會因素，例如用戶在平臺外的知名度或特定技術的採用趨勢。這可能會導致AI驅動的產品反映出某些偏見，讓系統更偏好推薦一些人羣，忽視另一些人羣。"}]},{"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":"去年，我們對PYMK底層算法做出了多項更改，以改善所有用戶的PYMK體驗。這些更改讓PYMK成爲了一項更公平的功能："},{"type":"text","marks":[{"type":"strong"}],"text":"無論用戶已有的人脈有多強、平臺使用率多高，系統都會對他們更加一視同仁， 而不是讓平臺的少數“高級用戶”獲得更大比例的資源。"}]},{"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":"此外，儘管我們預計PYMK的一些傳統關鍵參與指標（例如發送的邀請）會因這些更改而減少，但我們實際上看到淨參與度因此有所提升了。"}]},{"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":"我們之前改進領英Feed得到的經驗也是類似的，當時我們優化了"},{"type":"link","attrs":{"href":"https:\/\/engineering.linkedin.com\/blog\/2018\/10\/linkedin-feed-with-creator-side-optimization","title":"","type":null},"content":[{"type":"text","text":"創作者"}]},{"type":"text","text":"，而不僅僅是受衆的體驗：我們更新Feeds時不再只看重病毒式傳播潛力排名，結果平臺的參與度由此提升了。"}]},{"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":"過去一年來，隨着越來越多的崗位轉向了遠程狀態，我們認爲在互聯網上建立人脈和職業社交圈的能力將越來越成爲人們的必要技能。如果像PYMK這樣的AI驅動系統能夠滿足大衆的普遍需求，那麼它們一定會體現出非凡的價值。"}]},{"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":"在這篇文章中我們將介紹其中的兩項更改，看看我們需要解決哪些問題、我們如何實施解決方案，以及迄今爲止我們從工作中看到的成果。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"發送方-接收方PYMK體驗整體優化"}]},{"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":"當一位領英用戶想與平臺上的某人加好友時，她（發送方）需要向對方（接收方）發送邀請。只有在對方接受邀請後纔會建立好友關係。換句話說，除非接收者\/被邀請者有機會查看邀請並想要接受它，否則好友是加不成的。"}]},{"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":"爲了在領英上創造更好的用戶體驗和更健康的人脈生態系統，我們需要確保好友關係雙方都從關係中受益。"}]},{"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":"最近，增長數據科學（Growth Data Science）團隊試圖解決一個初看上去與增進公平性無關的問題：PYMK爲極受歡迎的領英用戶提供的體驗並不理想。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"通過降低曝光提升用戶體驗"}]},{"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":"領英上有一部分用戶收到了大量的好友請求，例如行業中的名人、知名的高級管理人員或大公司的招聘人員等。宏觀來看，得到不成比例的好友請求似乎與我們既定的縮小人脈差距的目標背道而馳。"}]},{"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":"然而，它也可能導致用戶的人脈被看似隨機，或與他們自己的職業沒多大關係的用戶的Feed更新和通知所淹沒。被邀請淹沒的用戶也可能錯過很多有價值的社交機會。這種負面體驗既體現在了這些用戶使用PYMK的相關數據中，也體現在了直接用戶反饋意見裏面。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/51\/f6\/51f6cb4280eb74a28382026c1aa0f6f6.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":"爲了解決這個問題，我們在P(connect)分數的基礎上添加了一個重新排序器，對收到邀請數量過多的接收方的分數做打折\/衰減。換句話說，接收方收到的邀請越多，他要顯示在PYMK結果中所需要的原始分數就越高。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/67\/18\/67f99fa4f18dbb00777b587a237aef18.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":"舉一個簡單的例子來幫助大家理解這一系統的工作原理。在上圖中我們有一個排序好的推薦列表，右上角的數字顯示了用戶已經收到的邀請數量。在我們衰減分數之前，候選人A排名第一。"}]},{"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":"爲了重新排序用戶，我們計算newScore = score * df，其中score是pConnect，df是[0,1]中的衰減因子。假設我們希望降低在過去一週內收到超過10個邀請的接收方的優先級。衰減後，A的新分數變小，因此A的排名低於B、C和D。在實際應用中，我們使用分段線性函數來計算給定收到邀請數量時對應的衰減因子。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"使用A\/B測試評估我們的方法"}]},{"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":"在領英，我們廣泛使用A\/B測試來評估我們大多數產品和AI模型的表現。這種方法也用來測試我們的PYMK模型，其中用戶被隨機分配到不同的對照組，分別會看到來自不同模型的建議。"}]},{"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":"如果一個模型有更好的推薦效果，用戶就會發送更多的邀請。這是對發送方的影響，可以直接從A\/B測試結果中讀出。"}]},{"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":"然而，PYMK模型的影響不僅限於發送方。用戶收到邀請後，會來到領英查看並接受邀請。這是對接收方的影響，但不太容易衡量，因爲某些接收方可以收到多個發送方的邀請，因此很難將影響歸因於特定的發送方。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/0f\/1e\/0fee7e8276b5096168dc3469fcf0521e.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":"爲了克服這一挑戰，我們開發了一個歸因框架來將接收方的會話歸因於正確的發送方。例如，如果接收方收到通知說發送方A想要連接，然後來領英接受邀請，則會話將歸因於發送方A。如果接收方主動上領英但直接接受來自發送方B的邀請，則會話將歸因於發送方B。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/91\/a0\/9141949fbf531c3ae0ddf54529656fa0.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"實驗結果"}]},{"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":"加入PYMK衰減機制後，我們希望用戶體驗得到改善，因爲我們已經讓用戶不會收到太多邀請了。另一方面，我們也預計用戶參與度會略有下降，因爲發送的邀請會相應減少。"}]},{"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":"我們假設如果發送的邀請變少，用戶的活躍度就會降低。然而，爲了獲得更好的用戶體驗，我們仍然願意推進這一更改。"}]},{"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":"更改實現後，我們觀察到用戶痛點的緩解符合預期。根據我們的調查結果，我們將平臺上過載的接收方（過去一週收到過多邀請的用戶）數量減少了50%，並注意到發送和接收好友請求的用戶體驗有所改善。"}]},{"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":"另一方面，與預期相反，我們實際上提升了用戶參與度，這是違反直覺的。雖然發送的好友請求確實減少了1%，但我們觀察到來自接收方的會話增加了1%。這是因爲原本收到邀請較少的用戶收到了更多的邀請，對他們來說，邀請在提升參與度方面的效果更明顯。來自發送方的會話影響是中性的。因此總體而言，會話提升了1%。"}]},{"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":"從這些實驗和對PYMK的更改來看，很明顯推薦的分佈（而不僅僅是推薦帶來的人脈結構）對整體平臺的健康狀況非常重要。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"協助參與度較低用戶的干預措施"}]},{"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":"領英考慮的另一關鍵領域是爲處於職業生涯不同階段，或可能還不熟悉領英（或一般性的在線社交網絡）的某些功能的用戶提供良好的體驗。"}]},{"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":"對於這些用戶來說，PYMK一直以來都被證明是一項潛能巨大的功能：用戶可以更快地找到同事和同齡人，並找到可能成爲朋友、潛在導師或未來合作者的他人。"}]},{"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":"直觀地說，與擁有數百個甚至數千個好友的用戶相比，好友數很少的用戶可能會因多添加一個好友而獲得更大的增量效用。"}]},{"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":"像許多AI算法一樣，支持我們PYMK推薦系統的算法會從導向成功匹配的推薦中學習經驗。常駐用戶（FM，在領英上參與度更高的用戶）在用於訓練這些算法的數據中往往比不那麼活躍的用戶，也就是非常駐用戶（IM）具有更大的代表性。"}]},{"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":"在ML應用程序中，由於訓練數據中的表徵不能均勻分佈，算法可能會對某些羣體產生偏見。在PYMK的情況下，我們觀察到常駐用戶（由於他們在網站上的較高活動水平而在訓練數據中有更大代表性）被推薦給其他用戶的機率往往更大。隨後，這些用戶可以添加更多好友，讓他們在訓練數據中的權重得到進一步提升。"}]},{"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":"那麼我們如何確保PYMK公平地代表兩個羣體的用戶，並避免強化網絡行爲中的現有偏見呢？"}]},{"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":"正如我們之前將領英Fairness Toolkit（LiFT）用於PYMK公平用例的"},{"type":"link","attrs":{"href":"https:\/\/engineering.linkedin.com\/blog\/2021\/using-the-linkedin-fairness-toolkit-large-scale-ai","title":"","type":null},"content":[{"type":"text","text":"博客文章"}]},{"type":"text","text":"中所討論的那樣，我們已經開發並測試了根據機會均等和均等概率重新排序好友推薦的方法。"}]},{"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":"這裏我們應用到了推薦系統中，給常駐與非常駐用戶設定同樣的代表性水平以解決這個PYMK問題。由此以來，我們看到發送給IM的邀請增加了5.44%，IM建立的好友關係增加了4.8%，同時FM的相應指標維持不變。"}]},{"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":"這很有趣，因爲通常，當邀請從FM組轉移到IM組時，我們預計後者的指標會增加，而前者的指標會減少。然而，我們觀察到FM的指標維持不變，IM的指標增長，這表明推薦質量總體上有所提高。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"總結"}]},{"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":"如前所述，我們負責任的AI和設計工作的一個目標是縮小經濟機會差距，不因用戶在何處長大、在哪裏上學或工作而對他們抱有偏見。"}]},{"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":"另一個目標是識別和減輕我們產品中的系統性不平等和不公平的偏見，反映我們身爲全球平臺的責任心。"}]},{"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":"鑑於對常駐和非常駐用戶的調整已經取得了成果，我們團隊還在做短期和長期測試，以瞭解調整PYMK推薦結果中的曝光度能否提升機會平等性。例如，我們未來的工作將使用受保護的屬性（如性別）來對比不同羣體。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"問答"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"InfoQ：對接受過多邀請的會員減少曝光率是很好的想法，但是如何應對潛在的惡意利用行爲呢？例如，有些人可能通過大量垃圾邀請來誤導算法減少對特定成員的曝光率。"}]}]},{"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":"strong"}],"text":"Kinjal Basu，機器學習高級軟件工程師："},{"type":"text","text":" 我們邀請系統從設計上就基本杜絕了垃圾邀請轟炸的可能性。比如說一位發送方只能給一位接收方發一次邀請。另外我們有一套機制來讓用戶控制誰能給自己發邀請，並對他們收到的邀請提出反饋意見，我們還有反垃圾AI模型來預防這種行爲。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"InfoQ：有些會員可能並不在意自己的邀請過載，他們能否選擇繼續利用之前的算法，繼續收到更多邀請呢？"}]}]},{"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":"strong"}],"text":"HeloiseLogan，機器學習軟件工程師："},{"type":"text","text":" 用戶當然能繼續通過PYMK以外的渠道收到邀請。例如，他們可以主動搜索別人來發出邀請。PYMK推薦系統希望能鼓勵大家建立對彼此有用的人脈。我們的目標更看重質量而非數量。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"InfoQ：領英是如何區分會員不同行爲對活躍度的影響呢？（例如，搜索一項職位、發起一次邀請、發表一段評論等行爲的活躍度打分是否有很大不同？如果有不同，如何確保這種權重差異的公平性）"}]}]},{"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":"strong"}],"text":"Qinnan Yin，數據科學家："},{"type":"text","text":" 我們會進行各種分析，以瞭解不同用戶行爲在不同情況下的影響。發送和接受邀請只是其中一個場景。我們全面評估了不同用戶行爲的公平性（針對不同產品，如工作推薦、新聞Feed排名、PYMK等）。由於這些產品的性質非常不同，用戶行爲可能會有顯著差異。我們在開放能夠縮小可能存在的代表性差距的技術時考慮了這些差異。"}]},{"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":"strong"}],"text":"原文鏈接："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/engineering.linkedin.com\/blog\/2021\/optimizing-pymk-for-equity-in-network-creation","title":"","type":null},"content":[{"type":"text","text":"https:\/\/engineering.linkedin.com\/blog\/2021\/optimizing-pymk-for-equity-in-network-creation"}]}]}]}