因果推斷在阿里飛豬廣告算法中的實踐

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在飛豬搜索CPC廣告業務中，廣告策略不僅需要考慮CPC消耗和廣告主ROI，平臺整體營收 ( 即CPC消耗+自然交易抽傭 ) 也是不能忽略的優化目標。傳統上基於廣告pCTR、pCVR、bid等因子的策略算法僅僅從廣告曝光本身來對廣告主、平臺和用戶的利益進行優化，難以準確調優這一平臺整體目標。"}]},{"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":"我們通過引入因果推斷技術，將廣告投放建模爲對搜索產品的干預 ( intervention )，直接預測廣告投放與否對業務目標產生的uplift效應，作爲下游優化問題的線性獎勵 ( rewards ) 或約束 ( constraints )，以支持各類線上策略。"}]},{"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":"在本文，我們從其中預算分配策略的視角，介紹飛豬廣告算法如何利用廣告效應模型進行業務目標優化，同時也介紹模型底層特徵 ( 如CTR、CVR ) 的一些建模經驗。主要內容包括："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平臺視角下的廣告預算分配"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"因果推斷技術"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"因果推斷在廣告策略中的實踐"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"未來思考"}]}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"平臺視角下的廣告預算分配"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1. 預算分配：多目標優化問題"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"廣告策略（或者說預算分配），本質上是一個多目標優化的問題。下面結合一個示例簡述一下這類優化問題的目標和約束。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/3b\/3b9a92081c5d383d8880dbf0614aa964.png","alt":"圖片","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":"從平臺角度上來講，預算分配的優化目標往往是包含多種複雜的因素的函數："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平臺變現效率（即ecpm）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"廣告點擊率（ctr"},{"type":"sub","content":[{"type":"text","text":"ad"}]},{"type":"text","text":"）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"廣告轉化率（cvr"},{"type":"sub","content":[{"type":"text","text":"ad"}]},{"type":"text","text":"）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"廣告主收入（gmv"},{"type":"sub","content":[{"type":"text","text":"ad"}]},{"type":"text","text":"）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平臺利潤（revenue"},{"type":"sub","content":[{"type":"text","text":"ad"}]},{"type":"text","text":"）"}]}]}]},{"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":"同理，從約束角度看，主要包含以下約束："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"廣告主總預算（budget）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"投放計劃的投資回收率（ROI）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"點擊率下限（ctr"},{"type":"sub","content":[{"type":"text","text":"thres"}]},{"type":"text","text":"）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"廣告主收入下限（gmv"},{"type":"sub","content":[{"type":"text","text":"thres"}]},{"type":"text","text":"）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平臺利潤下限（revenue"},{"type":"sub","content":[{"type":"text","text":"thres"}]},{"type":"text","text":"）"}]}]}]},{"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，對於平臺中的不同的業務，不同的用戶，甚至不同產品不同的投放單元，具體的目標和約束也會存在較大差異。針對不同的目標和約束，需要使用不同的因子加以控制。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2. 常用控制方法和控制對象"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"常用的控制算法包括："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"PID：最常用的一種傳統的工程方法"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對偶法：基於線性規劃求解拉格朗日系數，得到影子成本，通過影子成本進行線上控制"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"線性插值法：通過歷史數據擬合控制變量與目標變量的關係"}]}]}]},{"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":"以上提到的幾種方法有一個共同的特點：都基於線性模型，因此，要想達到有效的控制，就需要線性因子。"}]},{"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":"常用的控制對象包括："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"廣告主的出價（bidding）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"CTR\/CVR准入下限"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"參競概率"}]}]}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"3. 使用因果推斷的語言建模"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"廣告策略問題，常用的策略因子包括："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"CTR預估（pCTR）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"CVR預估（pCVR）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"出價勝出概率分佈（Bidding Landscape）"}]}]}]},{"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，以上的策略因子具備侷限性。以GMV爲例，對平臺GMV進行控制，僅使用pCTR、pCVR等策略因子是不夠的。因爲從業務角度看，在電商領域，原生廣告和產品（包括搜索、信息流等）的相似度很高，因此廣告投放的分佈和對應原生產品的分佈是相似的；這樣帶來的結果就是，當廣告的CTR\/CVR高的時候，自然產品的CTR\/CVR往往也會很高；我們不能基於較高的CTR\/CVR預估結果推測廣告投放的真實效果。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/07\/071aa1441167d58f85c533e301cb01c1.png","alt":"圖片","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":"基於此，引入本文的主題——使用因果推斷建模。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/30\/301d114450f6e7b81a8542af100370a9.png","alt":"圖片","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":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"公式中的參數T可定義成是否投放廣告的indicator變量"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"模型控制目標是平臺的收入（即GMV）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"建模問題定義爲廣告投放對平臺KPI的影響，即在給定特徵的前提下，投放廣告的GMV減去不投放廣告的GMV得到的收益"}]}]}]},{"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":"只有通過建立上述模型，才能優化廣告投放的真實效果，實現平臺和廣告主的共贏。該模型涉及到的因果推斷技術，將在下一章節詳細闡述。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"因果推斷技術"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1. 基本問題"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/aa\/aaa4f4c9b88cb2aa7816fd8997d144de.png","alt":"圖片","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":"將前文所述的模型抽象成上圖所示：我們需要建模的是T對Y的影響（T即是否投放廣告這個動作；Y是平臺的某項KPI，如GMV）。"}]},{"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":"然而影響平臺建模的因素很多，比如不同活躍度的用戶、不同的時間段、不同的行業等，都會影響廣告在搜索推薦中的轉化效率。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2. 影響因素的控制和約束"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如何在這種複雜的數據關係中提取出廣告投放對平臺的GMV產生的影響？一個較爲樸素的方法就是控制變量法，即人爲控制和約束除了廣告投放之外的其他相關因子。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/6f\/6ff939443b1ab2dc383213a81eb3d7e0.png","alt":"圖片","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":"具體來說，有兩類實現路徑："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"① 隨機實驗（切斷X->T）：假設T是隨機分配的，即T與任何變量獨立；"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"② 特徵工程（切斷X->T，X->Y）：假設X包含所有的confounders。"}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"定義confounders：同時影響T和Y的變量"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"若給定X，則不同treatment group下影響Y的協變量分佈相同，即X->Y是天然的隨機分配"}]}]}]},{"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":"基於以上這兩條假設，從數學上理解，其核心是比較投放廣告與否的前提下，平臺KPI的期望："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/19\/198286e17d251721c527e23156e216be.png","alt":"圖片","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":"由於X與Y\/T獨立，故添加條件概率不影響最終的期望值；這樣，待求解的期望值等同於實際樣本分佈的期望。從因果推斷角度看，這是個比較理想的情況：從觀察數據上得到的模型，和我們期望的因果效應相等的時候，就可以直接使用機器學習模型解決問題。"}]},{"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":"拋開數學公式，從常識角度也能解釋以上假設：每組特徵下可視爲隨機實驗，得到局部的無偏估計；那麼任意樣本上的效應，可以在相應的X的分佈上積分得到。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"3. 從機器學習視角審視"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"前文簡述了一些因果推斷領域的技術背景，下面從機器學習視角審視這個問題，將會遇到兩個挑戰："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/92\/9206eb4339f63a6a7942eb75832a13c5.png","alt":"圖片","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":"strong"}],"text":"樣本結構方面，對兩組樣本（T=0和T=1）的分佈要求更加嚴格："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於一般的預測問題，只需要保證樣本整體無偏，即可預估"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於因果推斷問題，除了樣本整體無偏，還需要要求兩組樣本的結構相近；否則會存在誤差累計效應"}]}]}]},{"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":"對變量T的推斷準確度要求更高，這並不是機器學習所擅長的工作："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"傳統的統計模型，通過對數據引入假設，可推斷參數分佈"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"大數據時代的算法模型，往往關注預測精度而非統計推斷能力"}]}]}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"4. 樣本結構問題"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基於前文提到的兩個挑戰，首先講述樣本結構問題的處理辦法。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/4c\/4c8a6a01d699edeb5760499b87534282.png","alt":"圖片","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":"從業務角度列舉一個簡單示例：控制目標Y是歸一化（即單UV轉化）後的平臺GMV；處理組和控制組的用戶分佈有明顯差別（處理組的用戶價值較高，控制組的用戶價值較低）。"}]},{"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3種方法可以解決這類問題："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"① 隨機實驗"}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"選取部分流量，線上廣告隨機投放，不採用任何策略"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"成本過高，難以落地"}]}]}]},{"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":"② 後採樣"}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"通過重採樣方法，使兩組樣本分佈一致"}]}]}]},{"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":"③ 特徵工程"}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"機器學習常用方法"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"假設X包含所有的confounders，則雖然兩組樣本的P(T,Y)不相同，但P(T,X,Y)相同"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"實際很難做到，且對樣本也有較高要求"}]}]}]},{"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":"在後面的章節中會詳細介紹以上方法在廣告策略實踐中的具體應用。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"5. 變量的統計推斷問題"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"第二個問題是變量的統計推斷問題：在大數據時代常用的複雜模型中，如何精準推斷某一個變量的效應？具體到我們的廣告策略場景上，就是如何推斷廣告是否投放這個干預變量對平臺的KPI產生的影響。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這類問題其實是很多學科裏面臨的一個終極問題，很難得到真正解決。具體到工程實踐上，我們可以通過一個模型結構先驗和loss設計，促使模型對這一單變量學習到正確的效應。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"因果推斷在廣告策略中的實踐"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1. 特徵工程"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/09\/09dc42fdb014214a8b7f999a445022ca.png","alt":"圖片","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":"先從前文提到的兩大分支（特徵工程，樣本重採樣）入手，具體講述工程落地的方法。"}]},{"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":"特徵工程方面，目標是捕捉到所有的confounders（即影響到廣告是否投放，以及影響到我們的平臺效率的所有特徵）。"}]},{"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":"首先參考廣告系統裏其他常見工作，例如廣告的CTR\/CVR預估模型；該類模型對廣告效率的預估，會預先將原始的複雜特徵做彙總。基於此，整體的思想就是基於原生的廣告搜索模型做遷移學習。"}]},{"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，搜索推薦的信息流和廣告策略的數據，兩者的分佈往往存在差異；因此會涉及到模型的泛化性能問題，需要對模型的擬合精度和泛化能力做權衡。具體來講，通過策略採樣的方式（模型結構類似CTR\/CVR多任務模型，但是樣本和模型訓練方式完全不同於CTR\/CVR），將原生信息流樣本和廣告樣本作爲兩個任務分別學習，最後通過loss融合實現多任務學習。"}]},{"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":"對於剛剛提到的模型精度和泛化能力的權衡的問題，我們採用in-batchre-sampling的方式，對於每個訓練batch，通過超參去動態調整原生樣本和廣告樣本的佔比。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/43\/43a7edd917f36d9a18ba56a1c9c961bd.png","alt":"圖片","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":"除了剛剛詳細介紹的CTR\/CVR，特徵工程還包括廣告系統中常用的Search Rank Queue（原生產品隊列）、用戶畫像等，在此不做贅述。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2. 樣本重採樣"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"除了特徵工程，另外一部分就是樣本的重採樣。對於純觀察數據，如何在避免AB實驗的前提下，通過重採樣方法在投放廣告和不投放廣告這兩組的樣本中構造相似的分佈。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/98\/980d42358b0ebea57abd43cf3d73949b.png","alt":"圖片","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":"樣本重採樣常用的方法包括："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"① Propensity Score Weighting\/Matching（傾向評分加權\/配比法）："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基於樣本被分配到處理組的概率進行重採樣"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"採樣結果與隨機分組相似"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Propensity Score本身的準確度難以估計，不僅是用戶行爲用戶偏好的高維分佈，還受到廣告本身的很多策略甚至一些產品業務規則的影響，較難進行機器學習建模，因此實際應用很難"}]}]}]},{"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":"② Original Space Matching（原始空間匹配法）等："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"避免預估Propensity Score，直接在特徵空間進行匹配"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"需要注意分析樣本分佈"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"實際應用可行"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/12\/127e443c070f0344f2f5a6b39225d93e.png","alt":"圖片","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":"以上述搜索結果頁爲例，每個商品都可以得到一個模型預估值（pCTR）。可以將整個頁面pCTR的分佈作爲特徵，用特徵向量Vmatch表示；"}]},{"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":"首先對控制組（即未投放廣告）的樣本進行全量的索引構建；接着對處理組（即有投放廣告），按照控制組的向量索引，使用KNN方法查找最相似的K個沒有廣告投放樣本，進行配對。將控制組樣本和處理組樣本進行兩兩配對以後，很容易通過對比兩組對應樣本的轉化率或GMV來描述廣告投放與否對平臺KPI產生的影響。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/2d\/2dfa423f1ec56e9003e80d0c2ac1d8ed.png","alt":"圖片","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":"3. 模型設計"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從模型設計角度，如何捕捉單變量的影響效應？前面提到過，這個問題很難得到真正的解決；但是我們有幾個啓發式的方法可以得到落地應用："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"借鑑ResNet思想，通過單變量embedding，結合深度網絡短接（shortcut connection）將其接入到網絡的最後一層，從而避免複雜的非線性因素的干擾"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"藉助遷移學習中的領域自適應技術（Domain Adaption），通過多任務學習來實現。具體來說，是將兩組樣本的投放廣告與否，當作Multi-task的方式處理"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"引入正則項（包括Structure Regularization、Task Regularization等），藉助對業務的先驗知識，對任務的loss加入正則項，促使模型學習到廣告投放的效果"}]}]}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"4. Uplift Evaluation簡介"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/3f\/3f625bd89b9c3d6508fb3cdc716f0448.png","alt":"圖片","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":"因果效應模型的評估，和常見其他用戶模型的評估相比，最大的難點在於，每個樣本都沒有相反label。對於這類問題，業界常用分位數分析法進行評估，即對於每一類樣本，按照廣告效應的預測值，從大到小來進行分桶（上圖左側，分了十個桶）；每個桶內聚集廣告效應的預測值相近的樣本（包含不同的label）；將每組樣本內的label的平均值相減，得到該組樣本的uplift。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/cb\/cb6797795ca88480e1161c813ecb572d.png","alt":"圖片","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":"上圖是個理想的情況：左側貢獻比較大，右側貢獻比較小，甚至是負的。基於柱狀圖做累計分佈計算，得到下圖："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/22\/228661cf7f5bcea9037382a397e48ada.png","alt":"圖片","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":"這個圖和常規預估模型中的ROC曲線形似；同樣，曲線下的面積（AUC）越大，模型預估效果越好。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"5. 模型baseline："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"italic"},{"type":"strong"}],"text":"Direct Method"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"假如我們直接做預估模型，選取SRP（結果搜索頁，一部分是廣告，一部分是非廣告）的特徵，和廣告本身的特徵（包括pCTR、pCVR，以及單個treatment的embedding）；模型預測某一條用戶請求是否產生訂單（listing to order）的概率。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/a2\/a2e0f1511f96a0c09b56984f5602b0f0.png","alt":"圖片","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":"這類常用的預估模型，是將“廣告是否投放”這一treatment作爲一般特徵。模型的baseline是基於pCTR\/pCVR（包括最好商品的CTR\/CVR，以及廣告商品的CTR\/CVR和最好商品的CTR\/CVR的差距）進行預估，預估任務的AUC還算理想（0.736），但是評價廣告效應的Uplift Qini指數是負值；增加DNN後AUC有所提升，但是Uplift指數會變得更差。"}]},{"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":"由此可見，預估任務的預測能力和對廣告效應的評價能力並非正相關。"}]},{"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":"模型加入樣本匹配後（表格最後一行），將樣本調整成近似隨機實驗的樣本，可以看出廣告效應的預測能力就變成正向了。以上實驗說明了針對Uplift進行建模優化的必要性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/28\/2840803e85c48eefc09ed9be09c3e3d5.png","alt":"圖片","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":"6. 模型改進思路一："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"italic"},{"type":"strong"}],"text":"Direct Method with shortcuts"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/38\/38a241d7b5bc82f8f3b85db0f05c8db8.png","alt":"圖片","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":"將所有特徵都連到一個DNN裏面，採用了ResNets的思想，如上圖所示：左側網絡對用戶搜索請求預期的轉換效率進行建模，右側網絡對“廣告是否投放”產生的影響進行建模，最後通過線性模型加以合併。對於廣告效應的推斷方面，這種模型相比於DNN會有一定的提升（uplift Qini指數提升至0.6）。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"7. 模型改進思路二："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"italic"},{"type":"strong"}],"text":"Dimain-Adaption: Multi-task Learning"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/0e\/0e58ef3c7323c77b5f80ae801b16209d.png","alt":"圖片","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":"前文提到的建模方法都是將單變量treatment當作一種特徵，建立基於網絡結構的模型，這類基於網絡結構的模型在特徵權重擬合方面會受到模型結構以及數據的限制。"}]},{"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":"這裏嘗試換了一個思路方向，藉助領域自適應中的多任務學習方法，將“是否投放廣告”作爲兩個不同的任務（而不是一個任務中的二級特徵）進行分別預估，通過建立完全不同的網絡來學習“是否投放廣告”產生的效果。在重採樣後的樣本擬合後，uplift指數可提升至0.11。然而這種方法更容易對數據分佈過擬合，從上圖的右表也可以看出，對於未重採樣的原始樣本，uplift指數反而會更差。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"8. 模型改進思路三："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"italic"},{"type":"strong"}],"text":"Effect-Net"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/9c\/9c4bc4b87f56b27f5f4cebdb4774d8c0.png","alt":"圖片","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":"最後一個就是顯式建模單變量treatment效應，即從模型的結構上體現目標效應的作用方式。類似於前文所述的改進思路一，使用兩個網絡分別對預測效果和廣告投放相應；與思路一不同的是，在模型結構上可以顯式地體現出業務理解的先驗，即最後一層的模型融合部分，顯示地將控制組的輸出與uplift相加，得到最終的預測值。使用Effect-Net方法，由於強先驗知識的引入，在原始樣本（即未重採樣的有偏樣本）的表現效果最好。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"未來思考"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1. 因果推理與深度表示學習之間的關係"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基於已有的工作，學界提出了以下幾點思考："}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"怎樣的表徵是有利於因果效應的估計的？"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"怎樣理解表徵與confounder的關係？"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"表徵學習在debiasing、模型泛化等問題中的應用"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基於以上幾點，目前對於表徵學習和因果推理方面的前沿研究方向主要是Balanced Representation Learning。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/e0\/37\/e01e9f5864989f7b0b9e918d0c37d437.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2. Balanced Representation Learning"}]},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/0f\/69\/0f2fa72b9756f5738b8bb6a6f1b0a869.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"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":"Balanced Representation Learning本質上仍然是一種多任務學習，將投放廣告組和未投放廣告組進行分別建模，進而將兩個數字相減，得到因果效應。之所以稱之爲balanced，其實是在目標上加了一個約束，促使兩組樣本上學習到的表徵分佈相似。"}]},{"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":"然而實際情況中，兩組樣本的分佈是存在分佈差異的。主動約束IPM的做法存在方法論上因果倒置的問題：隨機或自然實驗中，正是因爲實驗分組與協變量的獨立性，帶來了各組樣本分佈的相似性，而反過來在模型訓練中通過目標函數的構造來獲取分佈相似的表徵，能夠帶來準確的因果效應估計，是值得懷疑的。"}]},{"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":"從另一個方面看，在學習相似分佈的過程中，必然要求模型忽略掉部分confounder的效應，這又進一步違背了因果推斷中條件獨立的假設。類似的工作，現有的改進方法和改進思路或多或少存在不足，這一方向存在很大的改進空間；團隊也在着手這一方向的深入研究。"}]},{"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":"分享嘉賓："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"觀宙"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"阿里飛豬 | 算法專家"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2018年加入飛豬技術部，先後負責飛豬搜索RANK算法與飛豬廣告算法，建立了飛豬搜索RANK算法架構，包括行業模型、深度預估模型、LTR模型，以及流量調度機制等；目前聚焦於廣告ctr\/cvr預估、受衆定向、出價、預算分配等技術方向。 "}]},{"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":"本文轉載自：DataFunTalk（ID：dataFunTalk）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文鏈接："},{"type":"link","attrs":{"href":"https:\/\/mp.weixin.qq.com\/s\/-yhbFaj9yV1CTNZ5pL6Xqw","title":"xxx","type":null},"content":[{"type":"text","text":"因果推斷在阿里飛豬廣告算法中的實踐"}]}]}]}